287 | GPT 5.5, GPT Images 2.0, GPT Chronicle, GPT Privacy Filter, Claude Design, Kimi 2.6, all in one week 🤯 20% of companies gain 74% of AI financial benefits, Tim Cook is Stepping Down and more important Ai news for the week ending on April 24th, 2026

SECURE YOUR SPOT FOR THE:
MULTI-AGENT ORCHESTRATION AI COURSE: https://multiplai.ai/multi-agent-orchestration-course/
AI BUSINESS TRANSFORMATION COURSE: https://multiplai.ai/ai-course/

Are you falling behind in AI… without even realizing it?

The uncomfortable truth: while most companies think they’re adopting AI, only a small fraction are actually capturing real value—and they’re pulling further ahead every day. 

If you're a business leader, this episode breaks down what’s really happening beneath the headlines from explosive model advancements to the growing divide between AI leaders and everyone else—and what it means for your strategy right now.

The key takeaway? Stop chasing small efficiency gains and start redesigning your business around outcomes. This episode shows you how.

In this session, you’ll discover:

•  Why AI models are already outperforming humans on PhD-level tasks—and what that means for your business 
•  The shocking stat: 20% of companies are capturing 74% of AI-driven value 
•  The real reason most organizations are failing with AI (despite “using it”) 
•  How leading companies are redesigning workflows—not just automating them 
•  The 3 critical questions every business leader must answer about AI strategy 
•  GPT-5.5’s breakthrough capabilities and what “autonomous work” really looks like 
•  Why responsible AI and security risks are falling dangerously behind innovation 
•  The growing US vs. China AI race—and why the gap is nearly gone 
•  What the latest tools from OpenAI, Anthropic, and Google signal about the future 
•  How to identify high-impact AI use cases that drive both growth and efficiency

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, the 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. Many interesting and different aspects of the air world to talk about this week. Four very interesting reports were released in the past week and a half, and we're going to cover two of them. I may cover the other two next week, but I will give you access to them in the newsletter if you're interested. And we had an entire reign of new models from several different companies, so we're gonna cover that as well, including GPT 5.5, including a new interesting tool from Claude, including a very interesting new model from China. So we have a lot to talk about. I have noticed in the past few weeks that the shorter, more focused on big things episode were very successful. So I assume this is a format that most of you prefer. So I'm gonna stick to that this week. If you would like to have me cover more rapid fire things, please reach out to me on LinkedIn and let me know your opinion. By the way, same thing the other way. If you do like the shorter, more focused episodes, let me know as well. And with that, let's get started. the report together with its appendices is over 400 pages long, so I do not recommend you go and read it. It is absolutely great to drop it into notebook LM, and getting a summary or a podcast that will walk you through the different details or just listening to the shorter summary I'm gonna provide to you right now. So the biggest thing that the report is talking about is how good the models are right now. So it is mentioning that on many different aspects. Many of the models that exist today out of the box without any additional training are exceeding human baselines on PhD level tasks. To make this more specific, they're saying that many of today's models and they're talking specifically about the ones that were released in 2025. And that is before the release of Claude 3.7 and GPT 5.4 and 5.5. They're saying all of these models are currently exceeding baselines on PhD level science questions, multimodal reasoning and competition mathematics. As well as code generation benchmarks, different kinds of them, and the performance on some of these benchmarks, specifically on code generation rose from 60% to nearly a hundred percent in just one year. While at the same time, the organizational adoption has risen dramatically and reaches 88% right now. Now I think the 88% number is way overstated. And I'm saying that because I'm working with companies every single week from most of the Western hemisphere and from many different sizes in industry. And when you ask people whether they use ai, most people say yes. wHen you start diving deeper on how to use ai, you understand that there is a very shallow level of understanding and usage of AI compared to what the capabilities that exist today. So if AI adoption means I've asked ChatGPT to find information for me, then I agree with the 88%. If we're talking about what AI can actually do in the organization as far as multi-agent orchestration and the ability to actually perform work, I think that number is potentially more around 10 to 15%. And even that, I think is being very generous because of companies that I'm not aware of. Now the big red flag that this report raises is the fact that responsible ai, the ability to generate rules, regulations, processes that will keep us specific, organizations humanity as a whole, safe is not keeping up with the capabilities of ai and they're seeing aI security and other incidents surge in just one year from 2024 to 2025. A almost 50% jump in such cases, and I'm sure this number is gonna keep on going upward. And I'm sure this number is gonna keep moving upwards. We're going to talk about the mythos leak later on in this episode. To show how much responsible AI is not the core thing that people are focusing on. Think about how much the labs and specific developers and specific companies focus their reports on capabilities and benchmarks and ask yourself, when was the last time you've seen a responsible AI benchmark reported as part of the release of new models? Another topic that the report focuses on is the US China performance gap, and they're basically saying that gap is now practically zero. They're stating that while in early 2025, the Chinese models were still trailing behind. Right now, as of March of 2026, Anthropics top model is leading by just 2.7% over Chinese models. And there are Chinese models who are actually leading over Western hemisphere models on several different categories. In addition, the Chinese are leading in the volume of new publications, citations, patent outputs, and industrial robots installation. So the race is definitely on. There are no clear gaps at this point. If the gaps exist, they're very minor, and again, they're not across the board. There are cases in which the Chinese models are actually better than the Western Hemisphere models. That is despite the fact that AI investment in the US has been 285.9 billion in 2024, which is 23 times the $12.4 billion that were invested in China in developing these models. The other point that they're raising is that while previously, in previous years, US was the biggest place to go to if you wanted to work on AI research, and we were attracting global talent that has been declining since 2017 and has reached its lowest point. Attracting global talent. Is that its lowest point? It's been since 2017 in the us. Another topic that I think is worth mentioning because we are not talking about it a lot on this podcast, is the infrastructure vulnerability of this entire industry. The US currently hosts over 44, over 5,400 data centers. That is more than 10 times any other country, including China, but almost every single chip that is in those data centers comes from TSMC, which is a single company in Taiwan that is dependent on a single company to generate the machines that helps them generate these chips. So if TSMC fails to more. O or the Chinese government decides to take it over by force, then the US data center world is finished, or at least slow down dramatically for the foreseeable future. Despite the fact that TSMC US expansion is starting to finally yield some fruits, and it started in 2025, it's gonna grow dramatically in 2026 and on, but right now the vast majority of chips are still generated in Taiwan. The report does not keep, obviously, the risk to the labor market, and they're stating that in the first quarter of 2026, almost 80,000 US tech workers were laid off with almost 50% of them, 47.9 to be specific, explicitly attributed to AI automation. Their projections are not positive by any means, and there are even worse Two entry level positions ages 22 to 25 kids straight out of college. Were already experiencing a 13% decline compared to the same employment before ChatGPT was released. Now the last point that I wanna touch on related to this report, which will be a perfect segue to the next report I want to talk about, is that they found a 50 point gap existing between AI experts and the public. On whether AI will positively impact how people do their jobs. So people who are AI experts, 73% of them think that it will have a positive impact on how people do their jobs, and people who are not AI experts, the general public, only 23% think that is the case. There are many other important aspects of that report, but we can again, go through all of them. The bottom line is very simple. AI is now extremely powerful and yet not perfect. So they are mentioning the jagged edge of the AI space where again, in some cases it can do PhD work and in other cases it does not know how to read an analog clock, and it fails to do so 50% of the time with the leading models. So they're talking about that as well. But in general, they're talking that the models are getting better. They are at or above PhD level in many different aspects. They can do actual real work across multiple aspects of the economy. The gap in the race with China is very, very real. And the risks are growing with every release of every new model. 'cause that's not the focus of anybody right now. Now, with that, I wanna switch to the second report I want to talk about, which is PWC 2026 AI performance study. That report was released on April 13th, but I was traveling, so I'm covering it right now. And it connects perfectly in piggybacking on the first report. And it reveals a very significant divide between companies and individuals who are focused on AI and successful in AI based on the ones who are not. Now, this is a very significant survey that PWC did. They surveyed over 1200 senior executives across 25 sectors globally. And what the research found is that 20% of organizations who are deploying AI are currently capturing three quarters, almost three quarters, 74% of all AI driven economic value. I'm gonna say that again. 20% of the companies are gaining 74% of the economic value because they are doing it the right way. So a few more statistics. The top performing 20% of companies are generating 7.2 times more AI driven revenue and efficiency gains than the average competitor. Think about how crazy and profound that is. These organizations are not just deploying more AI tools, these are the organizations who are using AI as catalysts of growth, and they are completely reinventing how they are doing their business, including the business models, including the operations, including how they treat customers, including potentially the actual things they deliver. Now what the study also identifies is that the boundaries between current industries are blurring as far as sectors because of ai, meaning companies who learn how to use AI effectively can go and generate revenue streams outside their traditional path. And this exact approach has been the single strongest factor influencing AI driven financial performance far ahead of automation and cost reduction, and that makes perfect sense. Actually. This is one of the things that I'm teaching in my courses and one of the things I'm doing when I'm working with companies as a consultant or in workshops, the idea is that you have to reevaluate the entire business that you have right now and looking at strategic questions that will tell you what is going to be the demand for your current, Let me break this into simple terms. As a business leader, you have to ask yourself three simple questions. Question number one is how AI is gonna impact the demand to your current services or products because your clients potential clients have access to ai. Are they not? Will they be willing to pay the same amount of money to use your products and services when they have access to ai? We've seen the SaaS Armageddon happening in the past few months. Every time one of the leading labs are issuing a feature that allows companies to do things that does not require traditional SaaS as before. The same thing applies to more or less every company in the world. You have to think about how your client's access to AI is gonna impact the demand for your current services and products. That's the negative side. On the positive side, there are two things that you can look at. Number one is what kind of new services or products you can offer to your existing clients in a profitable way that you couldn't do before because you didn't have access to ai. So there are many, many opportunities with existing clients to broaden the amount of services and or products that you're selling because you can now do this more effectively and with a positive ROI. The third component, which connects directly to what this survey covers is what kind of new sectors, new clients with potentially new services and products you can serve effectively profitably, that is aligned with a new strategy that you couldn't do profitably before. And there's a whole blue ocean of that if you think outside the box. this Is basically what this survey is proving that these three questions are very, very important to ask at this point. Now what they've noticed in this survey, and they divided it between AI leaders and the average company. So AI leaders are increasing. The number of decisions they're making in their companies without human intervention are nearly three times the rate of their peers. This type of automation is what PWC calls in this report Trust at Scale, which is driven by the fact that these companies create structured frameworks, including responsible AI governance and so on, that allows them to trust the output of the AI at a much higher rate than the average company, which allows them to then run significantly faster, deploy more automation, and they're creating a flywheel effect that achieves more and more AI capabilities and scales faster and faster compared to the average company that doesn't operate this way. This leads to employees at leading AI companies to be twice as likely to trust AI outputs, creating, again, this flywheel effect that allows them to just keep on scaling and keep on running faster and faster. Now, one of the things that they're mentioning is that these companies that are winning the most are completely redesigning their workflows around AI rather than trying to apply AI to their existing processes. In my courses and workshops, I teach six rules for success in the AI era. One of these rules is called Stop Thinking Efficiency and start thinking Outcome. And to keep it short and not to spend the next 20 minutes explaining it, it is. More or less exactly what the research found. Meaning if you look at your existing business process and you try to apply AI to every one of the steps of the existing process, looking at it through the lens of the existing process. You can create small and sometimes medium efficiencies in every one of, in every one of these steps, which will allow you to increase the throughput of the current process, which is obviously really good. But the reality is with the current AI capabilities, you might be able to circumvent 50% of these steps or 80% of these steps. Meaning if you don't look at the steps you're doing right now and the process you're doing right now, and you just think of the outcome, what is it that I'm trying to achieve and what is the most effective way AI can get me there? You might be able to find a completely different process to do this that will not save you 5% here and 2% there, but will save you 50, 70, 80% of the effort that you're doing right now, putting you on a path to significantly higher efficiency. Now as I mentioned, the study finds that business transformation rather than isolated experimentation, plays a critical role in achieving top of the line AI gains. Now I obviously agree with that 100%. It's something that I do with all the businesses that I'm working with. It's what I teach as well, but, and there's a very big but. One is I don't think these two things are mutually exclusive. I think you can experiment on the grassroots environment and try to build local things that actually help the efficiencies of day-to-day while working on the strategic initiatives, and one doesn't contradict the other. The other thing that I think, and again, I don't think I know I do this. Literally every single day with multiple companies is that it is almost impossible to get to the large scale strategic projects without initially starting with the smaller tests and experimentation, which gives you the idea of what is possible, how your data is structured, what you can and cannot connect, how things work on a larger scale, what limitations you have, how to train the people to work in these environments. Like all these bigger questions are almost impossible to imagine without actually experimenting on a small scale first. And so, yes, I agree with the concept. Looking for the ways to grow the business through new channels will generate significantly higher ROI than small daily incremental gains of efficiencies. However, you need to do both things in parallel, and it will be very hard to get to the big understandings without doing the small steps in the beginning to get the understanding of what is possible. Now PWC warns that the divide between the leaders and the laggers is most likely to grow as the companies are doing it right now. Will continue scaling based on proven AI use cases, which will, they will get to faster than the other companies, which allow them to learn faster, automate more, make more decisions based on this data, and do it safely while the companies who are not shifting to an AI first strategy will lag behind with a growing gap from the leading companies because they will not be able to move at the same speed. Now, this, by the way, aligns with a different PWC survey that is their 29th global CEO survey. That found that only one in eight CEOs, so 12% says that AI has delivered both cost and revenue benefits while 56% report no significant financial benefit to date. Again, small groups are making all the gains. Again, a small percentage of companies are gaining the vast majority of benefits because they're looking at it as a strategic initiative and they're willing to make the right steps in order to get there. And these leaders are focusing not just on cost efficiencies, but also on revenue growth, leveraging ai. And I would like to connect this to the things that I'm seeing in my latest workshops that are focused on creating multi-agent orchestrations that are connected to actual business processes in companies. And the trick is to always pick. Cases, peak use cases that will do both at the same time. Work, on savings and operational efficiencies, but also look for business growth. As an example, in my latest workshop that I did last week in the hackathon, which happens on the second day, so people who did not have a clue on how to build agents or what they actually mean or how the platforms look like on the second day, build tools and build applications and build processes that focus on both aspects. on driving growth while making efficiencies to a few simple examples on the growth side. In this latest hackathon, they built a system that does price research across the board against their competitors. This is something that took their salespeople two hours a day every single day before they had conversations with their clients, and everything was done manually. Going to competitors' websites, checking the pricing, pulling them out, putting them in an Excel file, going, comparing that to the costs and the margins to make decisions on how best to sell their products. And now they've built a system that does it on autopilot. When they come in the morning, they get a dashboard that shows them their exact competitive scenario with suggestions on how exactly to move forward. Now the system is not ready yet. It's not fully deployed yet. It's not a production system yet, but the roots and the first steps were made within 36 hours of knowing nothing about that. This is a system that will drive a more sales and B, higher margins, because every salesperson will have more data about what their margins are and how their competitive pricing looks like, which drives both these components, more sales and higher margins on average. The other thing that they built that drives revenue growth is several different automations around marketing initiatives creation and the capability to create better, more focused, more personalized, more targeted. Highly efficient at scale marketing components, whether it's texts or posts or images or videos that will drive more awareness to their brand and to their offerings, which will lead to more sales. This is something that was very manual, not very efficient, done through several different agencies and without yielding incredibly good results. And now they're building a system that will do it significantly better. On the efficiency side, they are building solutions that will improve client and supplier communication, as an example, which improve relationships with now driving client and supplier communication through a better setup. The outcome of building more efficient client and supplier communication channels is twofold. One, it generates better relationships with your suppliers and clients, which is a very good thing to have. They get answers and responses significantly faster and more accurate responses as well. And B, it reduces the amount of time these employees spend on mundane tasks and allows them to focus on what actually drives business. The bottom line is the transformation is not rocket science. You just need to know what you're doing. You need the right tools. You need the right infrastructure, and you need to train the right people on how to do these things. If this is something that interests you, if you're a business leader and you want your team, your company, your organization, your department to learn how to do this, please reach out to me. I will gladly share with you what kind of services I offer. It doesn't have to be me. It can be anybody else who provides these kind of skills and capabilities, but figuring it out on your own will be a very long, painful and costly process, and bringing in somebody who actually knows how to do this effectively, who have done this with multiple companies, who has business leadership experience is the right way to go. Now, if you are an individual and not a company or not in a leadership position, and you wanna learn these skills in order to grow your career, or maybe you are in a leadership position in a small business, come and join our multi-agent orchestration course. We sold out the first two cohorts very, very quickly. We're now selling cohort three, which starts on June 22nd. That is starting to sell very fast. So if you want to learn how to develop multi-agent orchestrations and learn how AI goes from a chat to teams of AI employees who can do more or less, any digital work in your business, come and join us on June 22nd, because if you don't, the next course is probably gonna be around July, probably the end of July when that course ends, and that is a whole quarter away. So don't wait and sign up right now. By the way, if you sign up before the end of April, there is a $200 off coupon on the website and you can find the link to the page that gives you all of that information and the coupon in the show notes. And I will summarize this entire section with the two reports we covered with a quote from Mohammad Kande PWC Global Chairman, who said 2026 is shaping up as a decisive year for ai. A small group of companies are already turning AI into measurable financial returns, while many others are still struggling to move beyond pilots. Now to the second general topic of this episode, which is all the model releases and starting with GPT 5.5, OpenAI just released GPT 5.5, which is its most capable model yet, and it achieves some pretty remarkable successes. So first of all, as you can expect, this model is designed and focused on autonomous, long, complex, multi-part tasks that involve around coding and any other kind of agentic knowledge work, as well as scientific research. Now the really interesting thing is that this model is beating GPT 5.4, which was re released about a month ago. So it's not like a very long time. It's beating it on more or less every benchmark while keeping the same quickness of GPT 5.4. So very low latency while using significantly less tokens. So let's break down the exact parameters. GPT 5.5 achieves 82.7 accuracy on terminal bench, which is a complex command line workflows it achieves 58.6 on the SWE Bench Pro, which is real GitHub issue resolution. And 73.1 on expert SW, which is Internal Frontier Evaluation with 20 hour median human completion time. All while using less tokens than GPT 5.4 when it comes to real knowledge work. GDP Valve, which is a evaluation that was developed by OpenAI to test actual work across 44 different occupations. GPT 5.5 scored 84.9%. This is higher than average humans on OS world, which is a real computer environment operation. It achieves 78.7%, which is again, higher than average humans. On Tau two Bench Telecom, which is complex customer service workflow benchmark, it achieves 98% without any additional prompt tuning. This gives you an idea how powerful this model is when it comes to doing actual real work. It achieves higher scores than average humans, more or less across the board, while doing it with significantly less tokens than the previous model and keeping the speed of the previous model. Now on scientific research, they also made big strides and GPT 5.5 is now showing significant gains in helping scientists in actual research on both. Multi-stage GE on multi-stage genetic analysis, on bioinformatics and on different mathematics operations and research. So again, these models are now going beyond the ability to help in day-to-day tasks, but they're also going to be helping scientists make new scientific discoveries from a cybersecurity perspective, which is a big topic we talked about in the past few weeks since the release of mythos. And then the parallel cybersecurity model from OpenAI, GPT 5.5 includes a cybersecurity safeguards. That have tighter controls on higher risk activity that's per OpenAI themselves. But they also announced trusted access for cyber programs, which allowed verified defenders, people on the good side and people who work to protect infrastructure, more access to advanced cybersecurity capabilities coming from this model. If you wanna learn more about that and you are on that side of the scale and you wanna be a part of the program, you can go to ChatGPT.com/cyber. GPT 5.5 is already available to plus pro business and enterprise users, and it is also available on Codex and the API and the pricing is $5 per 1 million token input and $30 per 1 million tokens output with a 1 million context window. Very powerful model at comparable pricing to the leading model from Anthropic. A few interesting quotes from relevant people. Pietro Schirano, who's the CEO of Magic Path, said, it genuinely feels like I'm working with a higher intelligence, and there's almost a sense of respect. I'm not sure if that's respect from him to the model or the other way around. I assume from here to the model, but the choice of words of working with a higher intelligence is interesting here, coming from a CEO of a successful company. the second quote comes from a senior engineer at Nvidia who had early access to this model, who said, losing access to GPT. 5.5 feels like I had a limb amputated. So basically working with one arm that's pretty dramatic to say on a change from 5.4 to 5.5, but obviously he feels that this was a significant upgrade to his ability to create and work with code. Michael Trull, the co-founder and CEO of Cursor said GPT 5.5 is notably smarter and more persistent than GPT 5.4 with stronger coding performance and more reliable tool use. It stays on task for significantly longer without stopping early, which matters most for the complex long running work our users delegate to cursor. So what does that mean? It means this model is specifically focused on long horizon complex tasks, including coding and other knowledge work. And it is doing very well at that while using less tokens, which is a very significant achievement for open AI in their race for global denomination against Anthropic and others. But GPT 5.5 is not the only model. It's actually one of a few models they released this week. Before they released GPT 5.5, they actually released GPT Images 2.0, which is their next version of an image generation tool, which achieves incredible results, more or less across the board, and gets really great feedback from anybody who's using it, myself included. It is much better at creating extremely realistic images, including things that look like errors that would happen in actual photography. Which makes it look even more realistic. It is also really good at rendering complex, detailed text user interfaces, elements, condensed compositions, anything that requires significantly more detail. It is very good at. It is also really good at editing existing images and upgrading them and making changes without losing the granularity and the focus and the crispness of images. Any of you who tried to take a nano banana image and manipulate it several times, you end up with something that is not in the same quality and not in the same resolution as the original image, and this model doesn't do that. It allows you to keep the quality of the image while making changes in edits with simple text prompts. It is extremely good, as I mentioned, in creating highly detailed images that include text in them, which opens the door to a huge variety of use cases. Examples are creating a restaurant menu, including the images of the food with whatever background you want, and it knows how to nail that, including all the details in it, which was almost science fiction before that, like it would if you try that with any of the previous models. You know that you get weird artifacts every now and then, and with some models more than every now and then. It also knows how to create multi-panel comic strips straight out of the box. It knows how to create UI elements, actual user interface elements for software. More about that in a minute, and many other capabilities that were very hard and very inconsistent with previous models, including the latest Gemini, nano Banana Pro. This model also is very good in non-Latin text rendering, so Japanese, Korean, Hindi, Bengali, et cetera, which was very inconsistent in previous models, and now it actually does it very well, including very small font and print, again, with a lot of details, with a lot of text on a single page. Now, the other interesting thing about this model. The fact that it's a thinking model, it is connected to the thinking capability of ChatGPT and it can understand your needs and understand what you mean and go and perform research on a specific topic to come back with what it needs in order to generate the really incredible outputs that it generates. Now, back to the user interface use case. This is one that I find very, very interesting. Any of you who try to generate user interface through code using Codex, knows that Codex is really good at coding and really bad at coming up with user interface. If you give the same prompt to cloud code and to Codex from a code quality perspective, they will both do great. Some people prefer that, some people prefer the other, it doesn't really matter. They're very close and they will do an awesome job. But from a user interface, look and feel, Claude will look great and polished, and Codex will look not great and not polished. And what this new functionality enables you to do is to go and experiment and iterate on really advanced complex well done User interfaces for whether it is a mobile app, desktop, website, whatever it is that you're trying to develop, using the new image generation tool and then taking that into Codex and asking Codex to build the components that it sees on the image, and it'll be able to do that. This is probably a use case that we're gonna see a lot of people experimenting with and potentially using regularly moving forward. Now, on that aspect, when it comes to building user interfaces, we're gonna talk about shortly when it comes to the new really cool tool released by Anthropic this week. But for now, let's focus on the image generation capability of ChatGPT. From my perspective, the three biggest things in this model. One is the quality of the output. The second is the realism. If you need to create photorealistic images, it is practically indistinguishable from real life. And the third that for my workflows and probably for many other workflows, is the most critical, is the consistency in maintaining resolution and details across changes, which is something I was struggling with with previous models. If you want to iterate four or five times, you have a significantly degraded. Quality of output, and this is not happening with this model. So I'm very, very excited about it. The moment I heard about it, I asked dispatch, which is the ability to talk to Claude through the phone. I actually saw the dues while I was on the treadmill at the gym, and I asked dispatch to add another path to the tools that generate images for my LinkedIn post and other social media, as well as the path that generates the thumbnails for my YouTube channel. And now it's going to run both of them in parallel. So in a few weeks or maybe a couple of months, when I run it more times, I'll be able to tell you whether it consistently generates better results in real life usage compared to Nano Banana Pro or not. Now I wanna touch on a specific negative feedback that I've seen in one of the articles. It was a specific article from TechRadar who was specifically talking about the use case of creating magazine covers. And what it said is that the model currently can create professionally designed magazine covers that has the images and the graphics and everything you want on a really, really high quality. The problem that they're running into is that is a flat image, means you cannot take it to Adobe tools or any other tools and then edit the different components, resize the text, move things around, control whatever you want to control, which obviously is a limitation when you're just generating the final output. It. However, I have two answers to that. One is maybe you don't need to. If you can generate the final output accurate enough and exactly the way you want it, maybe the extra steps of doing the fine tuning is probably not needed. That being said, I know very little about the design of covers for magazines, so that is probably a necessity, but that technology. Already exists. Those of you who do not know and ever tried this in Canva, and I assume it's the paid version because that's what I have, but maybe it's also available in the regular Canva tools. If you go to edit of an existing image, one of the functions is called Magic Layers. And what Magic layers does is it magically separates a flat image into the different layers. So you got a background, you get the foreground, you get the different components, you get the text, each and every one of them. You can drag, resize, relocate, delete, change the text, edit, change the colors, whatever you want. It becomes its own entity as if it was created separately as a separate layer in Canva. This is one of the coolest, most incredible capabilities that in my workflows and in my team's workflows complement that generation of ai. In many cases, the easiest way to get from a 90% solution to a hundred percent solution is to bring the AI image into Canva. Click on the magic layer thing, wait for two minutes until it does the separation, and then make the small edit that you needed instead of trying to convince AI to do exactly the change that you wanted without changing anything else in the image. I am certain that this same capability, that existing Canva will come to the professional tools as well, and I'll be somewhat surprised if Gemini and the image generation from ChatGPT does not offer that functionality out of the box. But that being said, I assume that is coming. many months ago, and we still didn't get it from the leading labs, but in the professional tools, I'm certain we're gonna get it across the board from everyone. But for now, you can definitely go and use it inside of Canva if you have such a need, I'm sure you will find it very effective. But wait, there is more from OpenAI. This is not the last thing than asked this week. They also announced Chronicles. And what Chronicles is, it is an opt-in tool that you can turn on when you are using Codex. And what it does is it takes screenshots of your screen on whatever cadence. I'm not a hundred percent sure, and then it takes the screenshots and analyzes them and creates summaries in MD file format that then the Codex tool can use as reference to understand what you're doing and have a deeper knowledge and context on what you're developing, how the process works, and so on. So from a conceptual perspective, this is absolutely fantastic. Again, any of you who have developed software, either on your own or vibe coding tools, understand that giving the model context of what you're working on and what other tools and what's your process is extremely powerful and helpful. The problem with that right now is that it saves the images. Unencrypted on your desktop for a short amount of time, but then it saves the MD files that describe what was on the screen unencrypted on your computer forever until you delete them. So while the tool may provide value, it also exposes a very significant risk because you have data on everything that was on your screen available in simple text on your computer that anybody who gets access to your computer can get access to. Now, if this reminds you of things we talked about in the past, you are correct. If you remember, Microsoft came up with a feature that they called recall when they came up with the new computers and that was a huge disaster from a data security perspective and nobody wanted to use it. And I think eventually they canceled that capability altogether because of the backlash on, I don't want the computer to be able to see everything that I'm doing and save that. And that is despite the fact that their data was encrypted and they, it was running completely locally on the computer versus Chronicles, which does not encrypt the data and actually runs it against the API and sends information to ChatGPT to get processed and then saves it not encrypted on your computer as text files. So I don't know how many people are gonna use it in its current state. I'm hoping for open AI because I do see the value in that, that they're gonna make the required changes in order to make this safer to use. And the last thing that I wanna mention, which probably wouldn't have made the cut, but because this past week I did a workshop in Europe and GDPR was a big thing. I want to mention that especially for the people who are listening from Europe or that work in industries that care a lot more about data security such as legal or healthcare and so on. So OpenAI just launched an open source privacy filter. It is currently available on hugging face and GitHub under Apache two license, meaning you can go and train it yourself for your specific needs and specific use cases. And what it is, it basically knows you. It basically allows to use ai, deep knowledge, context, awareness, and understanding of text to find sensitive information, including private address, private email, private phone, private URL, private date, account number, et cetera, et cetera. Secret keys, API, keys, and so on. It knows how to find these in data that you're running through it, which means you can connect it to your pipeline and save you a lot of problems. Now, they do not guarantee the results, but if this is your first pass, it reduces the chances of humans missing. More things and it allows you to then just verify the outputs of the AI rather than going over the entire dataset in a manual way. And the fact that it is open source and allows to train on specific data allows you to make it domain specific and relevant to your thing. In a test they're running, they show that fine tuning improved the accuracy of domain specific tasks from 54% to 96%, and I am willing to bet that 96% is higher than humans doing this manually. As I mentioned, when I did the workshop last week, it was in Europe, so GDPR was a big deal. That was before the announcement by OpenAI. But what I suggested to them, and I'm suggesting now to all of you, is you can build a skill in whichever platform you're using, whether it's Claude or ChatGPT or other, that will be GDPR Aware or whatever other level of data aware. And the way I suggested to do this is send the AI to do the research on the rules of GDPR and then build a skill that will know how to score and evaluate and filter to GDPR related content. And then you can decide in your universe, what do you want the skill to do? Does it block the entire pipeline? Does it highlight specific fields? Does it try to anonymize the relevant fields and just change the client name to. Practically client name or John Doe or whatever you tell it to replace it with and anything like that. So you can decide what's gonna be the output. But the other cool thing that I suggested to my client that did the workshop with me is if you're running on a team's license, at least on the Anthropic universe, you can force that scale onto every computer that uses Claude, meaning everybody in the organization can have that on their computer. And you can force a rule on the Claude MD on every single computer to force that scale to run every time you get data from any source. So again, this is not a firewall that guarantees that nothing will go through. It's not bulletproof, but it dramatically reduces the chances of human errors or just missing specific things. So I highly recommend to anybody who is in a data secure environment to implement something like this in your universe, either by using the OpenAI open source tool or by building your own skills around the specific needs that you have, or a combination of both. The last thing that I will say about OpenAI that wasn't very positive, they had a small size, but relatively long outage. On April 20th, that took down several thousands of users out of tens of millions. So it's not a big amount, but for those thousands of users, that was a very big deal. There's obviously a very strong negative sentiment on social media that shows how dependent we are becoming on these tools. And the only thing that I can say, which I said multiple times, and we had a whole conversation about it on the AI Friday Hangouts, earlier today, so I'm recording this on Friday, April 24th. And we run the Friday AI Hangouts every Friday at 1:00 PM Eastern Time. It's an incredible group of people who care about how to leverage AI effectively in business. And we just talk AI every Friday. It's kinda like an open mic scenario where people share ideas, share thoughts, share things that they build. And if you wanna join, you're more than welcome. There's a link in the show notes and you can come and join us whenever you are available. But in that, in today's conversation, we had a whole conversation about the need to have redundancy and the need to build your processes that in a way that is independent of the model, meaning running. The right structure of files and folders and instructions and MD files and so on that are independent of the model is going to run on. And that allows you to switch the model, either because the model is down or because there's a new model coming out from a different supplier that is now doing better for your use cases and you wanna switch and you don't wanna spend the next four to five weeks in figuring out what is going to break. And after talking a lot about open ai, let's talk about something really cool that Anthropic really this week. So Anthropic just released Claude Design on April 17. What is Claude Design? It is a new tool that is currently in research preview, meaning it's not perfect yet, and I can tell you it's not perfect, but it is really cool. It is currently available to anybody with Pro Max teams and enterprise, licenses. It'll probably flow to everybody else later on, but what this tool does is it helps you design user interface for applications, and it does it in a really effective and really cool way. You can provide it access to whatever you want as reference materials. This could be screenshots, this could be the URL to your website. This could be your actual code base of your existing application on GitHub or other sources. And it uses all of that to analyze what are the brand guidelines. You can obviously upload the brand guidelines as well, but from that, it actually builds the codes. Of all the different components and all the different knowledge it needs to rebuild, reconstruct, or create new user interfaces based on what it has learned. So step one, it does the research on the information. You give it step two, it creates components. Step three, it can create. Any kind of new design based on what it has created, and it gives you multiple choices and you can select how many, give me four variations of this, and then it gives you four variations of whatever you want. So different color swatches, different sizes of fonts, different styles, and you can just click through and move sliders around to see the different variations and pick the one you want and you can iterate from there. The very cool thing about this is, as I mentioned, it is not just generating the view of how it looks like it generates the actual code components. This is a button, here are the two different states of the button mouse over, not mouse over when I'm clicking the button, how does it look like? All of that gets created through code that then you can import into whatever coding platform you want. You can also export that because this can become a design brief. You can export that as an HTML page. You can export that as a PowerPoint presentation. You can export that as a PDF, And you can obviously push this to cloud code. You can also push this straight to Canva, so it is not replacing Canva, at least not yet. Or it'll probably replace Canva on people who are trying to create designs on Canva, which would be weird to me because it's not what the tool is built for, but you can push that into Canva to turn this into a presentation or whatever other thing you want to turn this to. Now you can edit your designs in more or less, every way you can imagine. You can click around on the interface, you can give it commands, you can use the sliders and buttons that it generates for you in order to try the different functions and the different options and versions and so on. It is really fun to use. It is really easy to use. It is really flexible to use. And it makes the design of user interface significantly easier than it was before. If you are following what's happening in the stock market, Figma stock has declined 49% year to date. So they lost 50% of their value because practically you don't need Figma anymore. And tools like this just make it more and more extreme. And as the announcement happened on the release of cloud design, Adobe Weeks and other relevant shares had dropped 2% or more at the moment that Anthropic announced that capability. To tell you how clear it is that this is a competitor to Figma Anthropic, chief Product Officer, mark Krieger was sitting on Fig MA's board and he resigned on April 14th from the board three days before the launch, which tells you that there's a clear conflict of interest and not the best relationship now between Claude and Figma. Now, just a few weeks ago, I shared with you a similar tool called Google Stitch, which does something very similar. It allows you to provide it, access to whatever reference materials you want, and then create user interfaces that turn into the code of the components that can then be imported into whatever tools you want. They're not doing. Exactly the same thing. And they're definitely not doing it the same way, but they're doing very similar things. They're both really powerful. I really enjoyed using Google Stitch. I actually used it with my son to create a website he wanted to create because he wanted to be able to sell his perler beads, creations. So I gained the opportunity to A, test a new tool, and B, teach my son how to use AI to do design. that was a very fun experiment. I also tested Claude design on two different projects. One was more simple and straightforward, and cloud design did a great work, and the other one was a little more complex because I was comparing two different websites and was trying to apply the user interface and the guidelines of one to the other, which is not an easy task to do, and it actually did. Okay. I would give it a seven out of 10 on that task, but that's seven out of 10. Took me about five minutes to do, and if I would've done this in a traditional way, it would've taken me a few days at least, if not a week and a half. So despite the fact that the outcome is not exactly what I wanted, it allowed me to explore how the second website can look in different variations of the guidelines of the first website. And I was able to do that and explore different variations of that in just a few minutes and in a one shot process. So I literally gave it all the information, explained what I wanted to do, and got a decent output. I'm sure if I would invest a little more time and iterations, it will do even better. But Anthropic also had a snafu this week, actually, a much bigger booboo than OpenAI had. And the incident that happened is that an unauthorized group got access to Mythos, which is the latest model that they said is not safe to release to the public, that they're only going to give to a short list of companies in order to make the world a safer place before they release it to the world. that model that is potentially has the ability to break into more or less any digital system out there, any operating system out there, any server system out there has leaked to a Discord group that was able to gain unauthorized access to it by getting a contractor credentials and guessing the URL that they need to put in order to access the model. Now, apparently through this leak, thousands of people across multiple organizations potentially got access to this very dangerous, from a cybersecurity perspective model that anthropic were trying to keep as safe as possible in order to keep the world safe. Now we know for sure that multiple organizations, including partner organizations and beyond, had access to this. We know for sure that people in the Discord group and other people who got access to the people in the Discord group had access, and because of that, we need to assume that nation state actors most likely obtain access to the model. Now, this wasn't a five minute thing. The model was available long enough for a widespread distribution across multiple networks. And it was definitely enough time and spread wide enough that people with malicious intentions would get access to the model. It also took Anthropic quite a while to notify the relevant government agencies of the breach. Now what Anthropic did is obviously it reissued recommendations to all its partners to revoke access credentials immediately and to assume the model was breached and compromised and to take the relevant precautions to do this. But this is obviously all reactive and the damage if is done and we need to assume it is already done. So here are my thoughts on this topic. Just a few weeks ago, Anthropic accidentally released the entire source code of cloud code, which is their best kept secret and the hand that lays golden eggs for them, and that was accidentally released to the public and basically anybody who wanted to, definitely their competitors, and definitely. Nations from the other side of the world got access to it. Now, after that, they made a very big deal out of Mythos saying it is potentially dangerous and that the world shouldn't have access to this because it generates a significantly high level of cybersecurity to more or less every critical system that we know. And they've defined an entire program of partnering companies to increase the safety of the release of that model and future models. And within less than two weeks, we learned that this model that was really dangerous and that shouldn't be available to the public, leaked out to a large number of unauthorized organization. Now this does not look good for anthropic at all. It definitely does not look as a company that has good control over the data that they get access to and or the advanced AI models that they are building and releasing in a safe or non-safe way. Now combine that with the fact that Anthropic has planted their flag on developing AI in a safe way and coming from a company that is trying to get back on the positive side of the government, that designated them as a supply chain risk. So as much as I like Anthropic and their products and I'm building incredible, really mind blowing, insane solutions with their tools and I'm teaching others how to do this, and I'm teaching workshops and courses and really doing stuff that would've been science fiction just a few weeks ago. This kind of level of security is literally unacceptable. I have no other way to say this. And they must find ways to reduce these kind of instances to zero. Otherwise they will not be able to compete at the highest levels because people will not be willing to take the risk if this is what happens with the most secure, scary model that you ever developed. And within a few days, anybody who wants access to it, gets access to it. Why should I trust you with my company's data? Not to mention secure, highly sensitive data from governments or companies, et cetera. And so I really hope for Anthropic that they'll be able to figure this out and turn this around. Now back to the model releases from this week Moonshot ai. The company behind the Kimi models just released an open source Kimi 2.6, which is a 1 trillion parameter mixture of experts model that demonstrates really incredible performance capabilities. Mostly on, as you can expect, long horizon coding tasks and multi-agent orchestration. Now this new model K 2.6 achieves higher scores on many of the leading benchmarks than GPT 5.4, Claude Opus 4.6, and Gemini 3.1 Pro. So maybe not the latest and greatest because we just got GPD 5.5 and Claude Opus 4.7. But it is beating the previous models, which were just recently released from the leading companies in the world, And they were able to achieve some really remarkable real life examples of what this model can do as far as long horizon tasks. As an example, the K point, this new model autonomously overhauled an 8-year-old financial matching engine. So an old piece of software, and it did it autonomously on its own. Over 13 hours, executing over a thousand tool calls across 12 different strategies, and modifying over 4,000 lines of code delivering 185 increase in throughput and 133% performance gain. Again, totally autonomously, 13 hours straight. In addition, K 2.6 scales, 300 sub-agents that can execute up to 4,000 coordinated steps simultaneously. That is three x what the previous model was able to do, and it is the highest number from any model in the world right now. Now, they also introduced some other capabilities that allows to run multiple other agents together with their agents in a unified claw groups. That's what they call it. That allows them to all work together to achieve specific tasks and other capabilities that make it one of the leading models in the world significantly cheaper than the Western hemisphere models and open source, so you can host it and run it on your own. That wraps up the model releases for this week. A lot to test and work with. Everybody's pushing in the same direction. Really advanced, extremely capable models that can run for hours autonomously, spin up other sub-agents to accelerate the work even further, and achieving above human results across multiple different benchmarks. And on the rapid fire. We're going to do just one quick summary of things that were announced on the Google Cloud conference this week. As you can imagine, the focus was all agents, agents, agents, and agents in the enterprise. If you want to be more specific. Google launch or expanded what they call the Gemini Enterprise portfolio, which includes the Gemini Enterprise Agent platform, which is a new platform to develop, run, and evaluate agents. That is an expansion of Vertex ai. This new environment provides a secure, full stack connective tissue needed to build scale, govern, and optimize agents with confidence. That's their words. So if you want, this becomes the mission control or the centralized point of contact to everything agents that you can develop for anything. and even more obviously for the Google ecosystem. This environment also includes their model garden, which currently includes over 200 models. Obviously the leading models from Google themselves, Gemini 3.1, pro, Gemini 3.1, flash, image generation Lyria three, as well as their open source models from, uh, the Gemma family, but also third party models including their top competitors such as Claude Opus, uh, sonnet and Haiku, and lots and lots and lots of open source models coming from any aspect of the world. So you're not locked into the Google Gemini environment while running on this agentic management platform. They're also expanding the capabilities of Workspace Studio, which is a tool that existed before that I actually already use in several of my automations. And what it is, it is a no code automation platforms that lets you build and deploy agents that automate processes across Gmail, docs, sheets, drive, meet, chat, et cetera, all the different Google tools. So think about NA 10 or make.com or Zapier that works just in the Google Workspace universe. So on one hand it does not connect to the rest of the tools, so that's a big disadvantage compared to. Or make or so on. But on the other hand, it is free to use within the Google Universe. So if you just need to build automations in the Google Universe, it is really easy to build. And as I mentioned, I'm already using it and it's working extremely well. They also announced that 75% of all code at Google is now AI generated and engineers just approve most of the code that's up from 50%, just less than a year ago now. Google also addressed their collaboration with Apple when it comes to using Gemini as the foundation models for anything Apple intelligence, and they also announced that the new Siri that has been talked about and delayed for many, many, many, many months, is finally coming later this year based on a custom made Gemini model. There are rumors that OpenAI are going to pay $1 billion annual fee to get a, this custom 1.2 trillion parameter model to run iOS intelligence and Siri. Now, to me, the most interesting aspect of that piece of the news is that whether you're interacting with Siri through iOS or with Gemini as the assistant on Android phones, you are using a Gemini Google model, which means more or less every personal assistant interaction that anybody in the world is going to have is going to go through Google's technology, which will give them a gap that I don't know if anybody will ever be able to close when it comes to understanding the day-to-day needs of people and how to serve them in the best way. Now the expected reveal of this partnership as a product within Apple is probably going to be in the worldwide, in the ww DC, conference of Apple on June 8th with the launch on Property iOS 27 later this fall. From a real world usability and partnership perspective, they are pushing very hard. Their partnership program, they're planning to invest $750 million to deliver new resources to their wide network of partners. And they're working very closely with the leading consulting companies, including Accenture and Deloitte and KPMG two. Deliver, test and deploy agents for their clients. Accenture has already built 450 plus agents on this platform. Deloitte committed its largest investment yet to that environment. KPMG pledged a hundred million pwc, 400 million. You can kind of see where this is going. Really big players with really deep pockets are all committing to investing in this environment. From a hardware perspective, they announced the eighth generation of their TPUs, which is their version of GPUs. That's their models, and they're delivering two versions for the first time. One is called TPU eight T. That is optimized for training new models. And the other one is called TPU eight I for inference, which is going to be optimized for inference. As you can imagine, the world is shifting very, very rapidly to a world in which inference is significantly more important as AI gets embedded into more and more processes, especially in the agentic world where you have 300 agents running in parallel for 13 hours doing everything on their own. You need a lot more inference and a lot less training, and that will allow them to compete with companies like Cerebra and Grok with a queue that was acqui hired by Nvidia that already has these capabilities. That's the future. We're gonna have two types of chips, some of them for training, probably significantly less for training and a lot more chips for inference. Now the last thing that I wanna touch on Google introduced workforce intelligence, which is a unified realtime layer of data and connectivity, which allows to power a Gentech workforce in a secure and dynamic way connected to everything in the Google universe, which again, makes a lot of sense. You want your agents to be able to have access to everything in your environment, in a safe, secure, and effective way. And this is exactly what Google is planning to deliver. So the bottom line of this very quick summary of the major things that Google announced. Vertical integration wins, right? They're offering the models, they're offering the hardware, they're offering the runtime environment, they're offering the productivity suite. They're partnering with the relevant companies to deploy this at scale. They have everything they need. And if you remember, if you go back in this podcast when they were doing miserably in the beginning of the AI era, when OpenAI just launched their tools, and it seemed like Google doesn't know what to do next and they were chasing their tails, I said all the time that they will win because they have everything they need in order to win. And we're now seeing more and more signs of that. The other thing that is very clear is that the Agentic enterprise is inevitable and everything is gonna go through agents, and it's just gonna grow both in means of its ability to access data in an effective way, as well as in its ability to coordinate large, complex, sophisticated processes across every aspect of the organization. And that Google's partner channel is massive and it is going to help them grow and deploy this at a very large scale, and what all of that tells us is the dominant position that Google has right now in the world is probably going to get even more dominant rather than the other way around. The last thing that I will say, because I have to mention that, is that Tim Cook, the CEO of Apple is stepping down. His last day at the job is going to be September 1st of this year, and he has been at the company for 15 years. In those 15 years, he achieved incredible financial growth for Apple. He took them from a market cap of 350 million to over 4 trillion, making them the third most valuable company in the world. He was able to take their service business from 3 billion quarterly in 2011 to 30 billion quarterly in the at the end of last year, and really made Apple what it is today, one of the most loved and successful companies on the planet. But he failed miserably in leading the company into the AI era. They had one failure after the other. There were many, many voices from large investors calling from him to step down, to bring in somebody who's gonna be focused on product development and new ingenuity versus just efficiency and operations. And the person who's stepping in is John Tennis. Who has been with the company for 25 years was involved in building, developing, and deploying some of the most iconic products ever built, including different variations of IMAX and MacBooks and iPhones and AirPods and other different things that Apple has built through the years. And he's going to take the position of CEO. How will that evolve and how will that impact the success of Apple in the AI era together, combined with what we report previously as far as their tightening partnership with Google and Gemini? Only time will tell, but these are. The last news that I'm gonna report today, a lot of other really interesting things happened this week. There are two other really interesting reports, one from McKinsey, which released their AI transformation manifesto with 12 themes that separate AI winners from the rest. Another report from Forbes that released their 2026 AI 50 list. Additional announcements on interesting partnerships such as SpaceX and Cursor, which is an interesting combination which might turn into a $60 billion m and a outcome. New plugins from Anthropic for Word and other interesting news that you should probably check out, and you can find all of these in the newsletter, and you can sign up through the link in the show notes. That's it for today. If you find this podcast helpful, please rank us and rate us on your favorite platform, whether it's Apple Podcast or Spotify, and click the share button. You can do this right now. It will take you a few seconds unless you're driving. Don't do this if you're driving, but anything else that you're doing, you can stop for a second. Click the share button and share this with other people who can benefit from this podcast. They will appreciate the fact that you're giving them such a resource. I would really appreciate it and you'll feel good about helping people learn about AI and hopefully helping the world end up on a better state because of AI rather than the other way around. We'll be back on Tuesday with an interesting episode in which I'm going to share with you my Claude Universe, how I'm actually working and building things through my different companies right now, and I'm gonna open the kimono and show you everything that I'm doing so you can do the same. That's it for today. Have an awesome rest of your weekend, and we'll be back on Tuesday.