211 | America’s high-stakes AI action plan, why an AI skillset could add $18K to your pay-check, leaked GPT-5 details spark frenzy, and more AI news for the week ending on August 1, 2025

23:22It is less than a week away from the time this podcast is aired, so you still have a few days to sign up. If your career is worth to you more than$1 ,000 if getting a potential raise of$18 ,000 a year. Again, I'm not promising any of that, but that's what the research is showing that investing$1 ,000 in order to get the course is a no-brainer. It's the highest ROI of any investment you can make right now, and you're making it in your own future. If you remember a couple of weeks ago, I shared another research with you that said that people with AI skills are getting 56 % more money as salaries and compensation beyond what people in the same positions get without AI skills.

24:00It is very obvious that this is becoming a critical aspect of career success across almost any position, across almost any industry. So again, invest in yourself, invest in your future, invest in the future of your company. Come join us on August 11th for the AI Business Transformation course. And if making$18 ,000 a year more money or 28 % more money on average than you're making right now is not a big enough incentive, there's another small incentive for being a listener of the podcast. you get$100 off the price of the course by using promo code LEVERAGINGAI100, one word, all uppercase. So click on the link on the show notes.

24:35It will take you to the course page. You can learn everything about it, about the curriculum, etc. And come and join us on August 11th. And now back to the impact of AI on the job market. AI referrals to websites has skyrocketed by 357 % from the end of last year, hitting 1.13 billion referrals in June of 2025. This is based on an article by TechCrunch quoting information from similar web. The highest jump in referrals were to news and media outlets that saw 770 % increase in AI referrals year over year, but the increase in referrals were more or less across the board. The largest contributor to that is not surprisingly ChatGPT that accounts for 80 % of AI referrals to top domains, which basically means they're driving more traffic than Gemini, Claude, Grok, and Perplexity all combined.

25:24But it's not surprising. Again, they have, I don't know, 700, 800 million weekly active users. That is way more than all the others. So that just makes sense. Another thing that became very obvious from this report is that AI overviews that show up in search results are dramatically reducing the amount of clicks that are coming to websites. A different company called Pew Research find that users clicked links 8 % of the time when there are AI summaries versus 15 % of the time when there aren't any AI summaries on the page, meaning almost half of the number of clicks when AI summaries are there.

25:57The top companies that are getting these referrals are not surprising, or maybe they are surprising, but on e-commerce, Amazon tops the list with 4.5 million, followed by Etsy and then eBay. On tech and social media, Google is seeing the highest number of referrals with 53.1 million referrals, then Reddit, than Facebook. And why am I saying maybe it's surprising, maybe it's not? These are the biggest websites on the planet. They get a lot of traffic anyway. And what we're seeing so far, which again is the surprising part, is that most of the referral traffic is still going to the same websites, meaning the underlying concepts that drive traffic to websites traditionally is also working in order to drive traffic from AI.

26:34But that is not completely accurate because there's a very big shift in the amount of traffic that websites are getting. And if you are a company that drives most of its business or a significant amount of your business from traffic to the website, you need to start taking this very, very seriously. Now, while Google search still completely dominates the field with 191 billion referrals in June of 2025, again, compared to 1.13. So almost 200x more traffic comes from Google than comes from AI platforms. But the growth in AI platforms is very, very significant. And it is going to chip in the traffic that's coming from your traditional ways.

27:13And if you don't take action right now, you may start losing critical traffic to your website, which may hurt your income, which may put your company at risk if you don't take action. Again, especially if you're some kind of an e-commerce website and your entire livelihood depends on people getting to your website. But it's not just website. There's a huge growth in AI applications as well. So a report by Sensor Tower reveals explosive growth in generative AI apps in the first half of 2025, with total downloads of 1.7 billion downloads, and in-app revenue doubling to 1.87 billion. And not only that users have downloaded the apps more and spent a lot more money in these apps, users spend 15.6 billion hours on AI apps across 426 billion sessions, nearly doubling the amount of time they spend in these apps in the second half of 2024.

28:01So almost 100 % growth in the past six months. Again, not surprisingly, ChatGPT tops the charts, including in revenue, except in China, where DeepSeek is at the top of the download table. That may change right now that they're moving outside of China and are detaching themselves from the Chinese ecosystem. But to put things in perspective, ChatGPT users currently average 12.1 days per month on the app. This is similar to X and Reddit, which are two of the websites that get the most amount of time spent on them, and they're only behind Google that is showing monthly usage frequency. Now, again, if you break this down to daily usage, Chachupiti reached 16 minutes of daily average use with its users on the app, closing very, very close to search engines such as Google with 18.2 minutes.

28:46Again, that tells you very clearly where the wind is blowing. Now, in addition to the main large language models and the leading AI labs, apps with AI in the description were downloaded 7.5 billion times, 10 % of global downloads. That I think will go away because very soon, maybe in the next year or two, more or less all apps will have AI built into them, and that's not going to be a factor anymore. But right now, people are looking for everything with AI, and if you have AI in your description, whether it's real or not, you're going to get probably more downloads. And now we'll dive into our rapid fire items, starting with an interesting breakthrough.

29:20A Singapore-based company called Sapient Intelligent has unveiled what they called an HRM or hierarchical reasoning model, which is a new architecture to drive artificial intelligence different than large language models. They're claiming that they can get answers up to 100 times faster with just 1 ,000 training examples. So what results did they achieve? They scored 40.3 % on the ARC AGI test. That's surpassing large language models like O3 Mini High with 34.5%, Claude 3.7 Sonnet with 21.2 % and that's only with 1 ,000 training examples, which is significantly less than the amount of effort that was required in order to train these models.

30:02The way it works is that HRM uses two models, a high-level module called H4 High that does slow abstract planning and then a low-level L module that does fast detailed computation. They're claiming that it is significantly more efficient than traditional large language models chain of thought and that it's achieving better results with significantly faster computed time as well as significantly smaller training runs. To put things in perspective, one of the tests they run is how much GPU time they need to invest in order to solve different things. So they used two GPU hours for professional level Sudoku solving that does very well on that particular benchmark or 50 to 200 GPU hours to solve the Arc AGI test.

30:44That is compared to, I don't know, hundreds of thousands of hours that large language models are being trained. What they are claiming is that HRM excels in anything that has a lot of accurate data, like healthcare diagnostics, climate forecasting, and robotics, like healthcare diagnostics and climate forecasting, and it also excels where low latency reasoning is critical, such as in robotics. Now, to make this even more interesting, Sapient Intelligence open source their code on GitHub. So anybody who wants to continue research and evolve this new concept can do that right now and just get the code from and continue from there.

31:19And they believe this could be an evolution beyond the traditional large language models. Now critics question the scalability of this thing beyond just puzzle and benchmark-based capabilities, and they're saying that they haven't seen any real-world enterprise problems being solved with this new technology or when there's ambiguous data that is not accurate, which is in most cases when you're in a business environment. So, but that being said, it will be very interesting to watch and I will keep you posted as I hear more about that. We are continuing to get more and more signals that GPT-5 is going to come out in the next maybe before you listen to this podcast.

31:52But GPT-5 traces appeared in GitHub Security Benchmark on Microsoft Copilot Smart Mode in a Cursor Employee Screenshot and now on Mac OS ChatGPT application in its config files as well as several different trusted testing partners. Now there's been several different weird variants potentially of ChatGPT on LM Chatbot Arena and other benchmarking arenas as well. They showed up, they performed very well, and then they disappeared, which also increases the tensions that this was different variants of GPT-5. So when exactly are we getting GPT-5? It's unclear. It is getting clearer. That is coming probably in the next few days.

32:29And some leaked YouTube stream schedules suggest that it may come out either just before or at the same time as Google's August 5th Tokyo Next event. And so it's not going to be the first time that OpenAI jumps the gun and jumps ahead of Google in order to release something and steal their thunder. So that might be when we get this, which means this coming week. Now what's in GPT-5? There are a few things that we already know. One is it consolidates their models and is supposed to choose for you what kind of model to run in the background, whether to reason or not reason, to run fast, run slow, how many steps to take, and so on.

33:03Similar to the way Cloud4 and Gemini 2.5 Pro are working. In addition, it's supposed to be very good at coding and put them back in real competition with Claude, which is currently the leading platform for most developers around the world. More about that in a minute. But going back to interviews in which we learn a lot, there was a very candid and open interview of Sam Altman this past weekend with Theo Vaughn. And in this interview, Altman mentioned a few things that I must admit are very scary to me and should be scary to everyone. He's talking about the fact that this whole AI thing has been fast, but he's saying that the past few months feels very fast, basically hinting that the performance that they're getting from GPT-5 are significantly better than they expected.

33:42He's also talking about his personal experience. And I'm quoting now, I felt useless relative to the AI. He was mentioning that he was testing GPT-5 and he just got an email that he didn't know how to answer. And GPT-5 seamlessly and effectively solved the problem that was presented in that email. And again, he personally felt useless. He also said, while testing GPT-5, I got scared. He's also saying that in many cases, AI feels like previous historical developments in which the development people said, and I'm quoting, what have we done? And he's connecting it back to the Manhattan Project. Sam Altman himself is not sure whether JATCPT5 is safe or not if he's comparing it to the development of the atomic weapon.

34:22If he got scared and surprised and felt useless, and he's the one that knows this technology better than most people on the planet, then I think we have more and more reason to believe that we need to be really worried about where this technology is going or where it can go without additional oversight. And something that really troubles me every time I hear him speak, and I said that multiple times before, either him or everybody else, it's not just him. The only one that maybe are taking actions in the right direction is Anthropic. But he said when he was referring to the lack of higher level oversight, he said, and I'm quoting, there are no adults in the room, basically suggesting that there needs to be more regulation and frameworks around what they're doing.

35:01And the reason it's driving me crazy is you are the one doing it. If you feel that this is wrong, if you feel that there needs to be oversight, create the oversight. Get together with the other leaders from this industry, from the US, from Europe, from Asia, from China, I don't care where they're from, bring together some governments and create a body that will look into all the things that we need to worry about. Right now saying, oh, I'm really worried somebody else should monitor what we're doing just doesn't make any sense to me. And it's the same thing with job losses. It's the same thing with figure out what's true and what's not.

35:31It's the same thing across multiple scary aspects of this technology with the people who are driving it are saying, yes, this is really dangerous. Somebody should monitor this, but they're not taking any action. Definitely not at the level of investment that they're doing everything else. And this is a very big concern to me and should be a very big concern to everybody else as well. In addition, OpenAI announced this week that they are starting Stargate Norway, which is going to be their first large data center in Europe. It is a part of their AI for countries program that they've announced a while back.

36:00The good news about this particular data center is that it's going to be at Narvik, Norway, and this facility is going to get 100 % hydro power and cooling capabilities. And it is supposed to have 100 ,000 NVIDIA GPUs in the beginning. It's supposed to grow afterwards and it's supposed to go live by the end of 2026. So in summary, big data center in Europe, all powered by renewable energy, which is a great idea from my perspective. In this next section, we're going to talk about interesting changes and shakeups in the competition in the AI world. In this article on the information, which is one of my favorite publications when it comes to learning stuff about AI, they're talking about the fact that AI makes it a lot easier for companies to shift their underlying AI infrastructure.

36:42And the specific statistics that they're saying is that AI-driven software transition can cut costs by up to 50 % for signed companies when it comes to transitioning from one big software to the other. And by reducing the switching cost, it enables smoother transition between software providers. The company with the highest risk right now is Microsoft, with a significant dominance around the world when it comes to enterprise software. And currently, you have companies like Amazon, Salesforce, and Palantir that are aggressively targeting their existing clientele and offering using different AI tools, reducing the risk and shortening the time to get the benefits from this transition.

37:16What does that tell you? It tells you that you yourself in your business, in your company, need to look into your existing tech stack and see what you can do more effectively with different AI tools right now and not just think in the traditional ways as you're looking at right now. The most famous example is obviously Klarna that said they're going to stop using Salesforce and stop using Workday and just develop AI solutions that provide them and just develop in-house AI solutions that will provide them the benefits that these really expensive and cumbersome software are providing them right now.

37:45Still on the topic of competition between giants, in a report this week from Menlo Ventures, they reveal that Anthropic AI models have overtaken OpenAI to become the top choice for enterprises. So per this report, Anthropic currently holds 32 % of enterprise LLM market share usage, surpassing OpenAI's 25%. Now that's a very big shift because in 2023, just a year and a half ago, OpenAI ended with 50 % of the enterprise market. Anthropic only had 12. Now, the biggest driver for that is obviously the coding aspect of it with Anthropic dominating with 42 % of enterprise market share compared to only 21 % of OpenAI.

38:23This started with the release of Cloud 3.5 Sonnet, followed by 3.7, and then Cloud 4 that are all performing better than the parallel OpenAI tools. Again, this is one of the reasons that GPT-5 has a significant focus, per the rumors, on coding capabilities in order to take some of that market share back. How critical is coding capabilities to Anthropic's dominance in the business world? Anthropic just announced that they're revoking OpenAI's Cloud API access, and the reason they took it away is that they're citing their terms of service have been violated. So Anthropic's terms of service are prohibiting using Cloud to, and I'm quoting, build a competing product or service or reverse engineering its models.

39:04OpenAI apparently is using the API to test coding, creative writing, and safety prompts, comparing it as a benchmark to its new models that are being developed. And so Anthropic decided to block OpenAI's access to its Cloud model API so they cannot test their own models against them and benchmark them against it. OpenAI chief communication officer Hannah Wong replied and said, it's industry standard to evaluate other AI systems. It's disappointing considering our API remains available to them. What does that tell you? It tells you that right now, Claude is ahead and OpenAI wants to make sure that GPT-5 will be ahead of Claude, at least on the code development capabilities, because this has been driving explosive growth for Anthropic.

39:43How explosive? Anthropic annualized revenue just hit 5 billion by the end of July, a jump from 4 billion in the beginning of July. This is showing you how rapid the growth of these tools are overall, but definitely how crazy the growth is for Anthropic from a coding perspective and why they will do everything they can in order to maintain that lead over OpenAI. Anthropic is close to securing$3 to$5 billion new funding, which is supposed to value them at$170 billion, almost 3x its$61.5 billion valuation this March. So we are four months away and they are tripled their valuation to these crazy numbers.

40:22But again, been driven by a very clear and strong growth. About 80 % of Anthropik's revenue comes from enterprise subscribers, including some very significant companies like Zoom and Snowflakes and Pfizer that are using Claude across everything that they're doing. Now, if their explosive growth wasn't enough, Anthropik is expecting their annualized revenue could grow to$9 billion annualized by the end of this year. That is doubling their old forecast of$4 billion by this year. So it's not just that they're growing twice as fast as what they thought they can grow at. Now that obviously doesn't mean they're profitable.

40:56They're still losing a lot of money. They lost$5.6 billion last year and they're expecting to lose$3 billion this year. But this is a smaller number while they're growing very fast and it definitely looks very positive for Anthropic. On a personal level, I can tell you that I love some of the things that Claude is doing that other models cannot do. Mostly it's in-app coding capabilities and its ability to display and run the code and troubleshoot the code within Claude itself. I create very cool dashboards for myself or for my clients using Claude, and I don't know how to write code, and it's the most effective at visualizations of data, at least based on my personal experience.

41:31Staying on the topic of Claude and programming, starting on August 28th, Anthropic introduces new weekly usage caps for Claude Code Pro, which is their$20 a month and for their max plan. The biggest focus is on their most advanced model, Claude Opus 4, which had at least seven partial or major outages in this past month. Now, it is most likely due to power users that are running it nonstop, which is straining Anthropics computational resources. Now, these new caps are most likely going to impact, based on Anthropics assessment, only 5 % of their subscribers, which they're saying are most likely sharing the usage of their passwords and so on with the world, which they think is unfair.

42:10Now, they're not the first one they're doing. Enesphere, the company behind Cursor, Replit, and other companies have adjusted their pricing for top users. The good news for people who are not sharing their password and still want to use Claude a lot is that if you're a Max user, you'll be able to continue using the app just paying based on their standard API rates and not based on the one-time monthly fee. I think that makes perfect sense and it gives people an out if they want to keep on using the model, just pay as you go versus just paying the one-time monthly fee. And from competition to Meta, Meta is continuing to hiring top talent from many different companies and to try hiring even more.

42:47But what we've learned this week is what is Mark Zuckerberg's vision for this new superintelligence and the direction Meta is going to take in the AI space. And what they're planning to create is what they call personal superintelligence that knows us deeply, understands our goals, and will help us achieve them because he believes that's going to be the most useful way to use AI. So they're not planning, at least at this point, or at least based on what they're saying, to focus on work automation, but rather making it a personal assistant that actually knows you and can empower you with everything that you do, not just at work.

43:20Zuckerberg notes that 200 years ago, 90 % of us were farmers and that technological advancements allowed us to go and develop science, health, culture, personal fulfillment, and so on that wouldn't have been possible without technology. And he believes that personal super intelligence that knows everything about you and complements you can take us to the next level in those topics. Another thing that we've learned is that Meta is aggressively pursuing partnership with an AI video startup. They had conversation with Pika, Higgsfield, and Runway and trying to get to either a partnership or potentially an acquisition of one of those companies with the goal of integrating it into everything Meta.

43:55That makes perfect sense to me, especially after the huge success of VO3 and its direct integration into YouTube Shorts, So they currently don't have anything that comes close to that. And acquiring a company that can bring them advanced video capabilities is perfectly aligned with everything meta because then they can integrate it into Instagram and into WhatsApp and into all the other tools that they have. Stay competitive in one of the cases which is directly tied up to how they generate WhatsApp revenue, which is consumption of video content online. Now, an interesting shift in their approach to AI has to do with the fact that Zuckerberg also said that they may not open source all their super intelligence models.

44:32So while they're still committed to doing everything open source, which has been the focus of the company since the beginning, he's saying that the company will need to evaluate very closely and very carefully what they're going to open source and that they're saying it may not be all their latest and greatest models that they're going to develop. This makes perfect sense to me because I think the risk of developing super intelligence and then putting it out there as an open source model significantly outweighs the potential opportunity that this will. The next topic has to do with the World Artificial Intelligence Conference week that just happened in Shanghai.

45:03There's been big announcement in there. So Chinese tech powerhouses like Tencent and ByteDance both released big models and had big announcement in the show. Tencent Cloud Unit showcased dozens of AI agents designed to automate multiple tasks. Tencent demoed their online marketing agent that does autonomous workflows across the entire marketing process. And both companies together with Baidu showcase their coding platforms as well. So more and more competition is coming from China. Alibaba share that they're entering the smart glasses field in China with a new set called Quark AI that is going to be powered by a special version of their proprietary Quinn model.

45:40And they're positioning themselves directly against one of the hottest field in the world today with companies like Meta with their glasses and Xreal and Rocket. And another product that did a big splash in that show is PND Robotics from China that revealed their new model called Adam Yu. It's not their first, it is their third robot that is AI based, but this one is very, very different. It actually doesn't have legs. And what they were showcasing is the advanced human-like motion that it can do with 31 flexible joints just in its upper body, which is all it has is its upper body. If you watch the videos, it's pretty remarkable to see how similar it is to human movements.

46:15And it actually combines what they call a PN link motion capture switch, which is basically a human that is wearing these gloves and helmet and so on, and it can mimic everything that the Shiman can do. It is pretty incredible to watch how accurate and versatile its movements are with its arms. And they're currently open to pre-orders with stock units costing$45 ,000. Is that something you're going to buy yourself? Absolutely not. Is this something that is showcasing where the future of robotics is going? Absolutely yes. They will be able to do everything that we can do in the real world and a lot more because they're going to be more accurate, faster, stronger, and so on.

46:50And every one of those small steps in that direction is showing us the direction that the robotics world is going to take. And the price ranges are also going down with every one of these presentations. And maybe the coolest product release that have happened this week is a company called Brilliant Labs have shared their Halo glasses, which is their second generation of glasses. And it has lots and lots of cool features. First of all, the glasses are going to cost$300, which is roughly the price range of most of the other glasses. But there are a few very cool capabilities to these glasses. First of all, they introduced Noah AI, which is their AI agent development platform that runs within the glasses.

47:28Basically, you can use your natural voice in order to ask the glasses to develop specific applications that you need, and it will develop the application for you. So that means you don't need an app store. You can actually tell the glasses what you want it to create. It creates an app specifically for you, and then it runs on your glasses. Another thing is they're developing what they're calling private and personalized knowledge base, which is the idea here. It can remember everything about you, about people it sees, about things that it see and so on, and be able to use that information in order to provide you more personalized and relevant information, whether it is navigation, shopping recommendations based on your own and previous taste, introduction to people that it knows knowledge about people that it has seen before and so on.

48:08So very unique and interesting approach that again is just showing us where this is going. We will all wear these AI wearables that will be aware of the world around us, will be aware of our taste, of our personal experience, about other people, connected to multiple different systems, and will become these cyborgs that will have significantly more information about the world around us at any given point. Now, is that a concern? It's a huge concern because it means that everybody will record everything that we're doing everywhere we go, that there are other people. And that obviously generates a lot of issues that I can think about and probably even more that I cannot think about.

48:39But as a society, it's something we have to start paying attention to, something we have to address, and something we need to be aware of because this is going to be ubiquitous in the very near future. That's it for this week. Quick reminder, the next AI Business Transformation course starts on August 11th. That's just around the corner, so come and join us. If you're enjoying this podcast, please share it with other people. Click now on the share button on your phone and share it with four, five, six people that you know that can benefit from this. I'm sure you know a few of these people. and this is your way to help increase AI literacy around the world.

49:10And I will personally really appreciate it. We'll be back on Tuesday with another fascinating how-to episode. And until then, enjoy your weekend, play with AI, share what you're learning with the world, and I will see you on Tuesday.