Hello and welcome to a weekend news episode of the Leveraging AI podcast, a podcast that shares practical, ethical ways to leverage AI to improve efficiency, grow your business, and advance your career. This is Isar Matis, your host, and we have a lot of interesting topics to talk about this week. First of all, we're going to deep dive into what is the current status of open source and the exponential growth it has seen in this past year and even more in the recent months, and how that impacts the overall global AI race. We are going to talk about very interesting developments in the local regulation situation inside of the US. We're going to talk about the latest annual report about the AI state by Boston Consulting Group, which was released a few weeks ago, and I've been delaying talking about this because there were bigger things, so we're gonna touch about that. We're going to talk about the current state of OpenAI's hardware ambitions and where it stands. They actually released a piece of hardware this week, but there's a much bigger story behind that. And then we have a lot of rapid fires items to talk about. Some of them are very critical and interesting. So lots to cover. Let's get started So we talked in the last couple of weeks about the big drive now towards more efficient models as costs are starting to hit hard on companies who are starting to implement AI at a larger scale, and those who are using more agentic solutions even more. So we've been talking a lot about that in the past few weeks. But a subset of that is the drive towards open source models and more specifically Chinese models. but now there's a new participant in that space from the US. So a few interesting of data points that I collected from multiple sources to paint the picture for you. The first one comes from Hugging Face latest report from this last month that is showing that open weight Chinese models now account for 41% of total downloads on Hugging Face, surpassing US models. Now, if you wanna dive specifically, earlier this year, Alibaba Qwen, which is a whole family of products from Alibaba, not just one model, has reached one billion downloads in January of this year, when late last year they surpassed Meta as the number one company in the world when it comes to the number of open source downloads of its products. Now, if you want another crazy piece of information, is that currently new repositories of open source models on Hugging Face are opened every seven seconds Another piece of information that is crazy is that half of Fortune 500 companies currently deploy models via Hugging Face as well. So it's not just small experiments or little tests or smaller businesses. Half of the Fortune 500 companies have open source repos running on Hugging Face. Another really interesting piece of information comes from OpenRouter. So we talked about OpenRouter many times on this podcast. They're a platform that allows you to integrate and connect one API and then get access to any other API you want. so all the different models are there, and they track everything that's going on the platform. Obviously, the top six most popular models on Hugging Face right now are all open weights, all Chinese, coming from Tencent, Xiaomi, DeepSeek, MiniMax, Z. And then Anthropic Opus is just the seventh model as far as usage on that platform. That tells you again that people who are looking at using multiple models are by far preferring using open source cheaper and yet very capable Chinese models over the closed source top leading models right now. Another angle of the same thing, Vercel, those of you who don't know Vercel, it's one of the most popular platforms in the world right now to deploy web applications. They are saying that open models currently handle about one third of all requests on the platform as of June of this year. So this is very recent information. That share, meaning the share of open source models early last year was negligible per them. So in one year, usage of open source models went from nothing to about 30% of everything that's happening on Vercel right now, and that number keeps on growing. Another great example comes from Chatbot Arena. So we talked about Chatbot Arena many times in the past. There are platforms that allows you to go in and do a white label test. You give it a prompt in whatever topic, whether video generation, image generation, agentic usage, etc. It gives you two different answers, and you choose the one you think is a better answer, and based on that, they're ranking the model. So these are actual use cases by actual people. And the open versus closed source gap has shrank as of March of 2026 to 3.3% on preference, meaning only 3% of people prefer the closed source models, open source models. Basically, it's in full parity On tasks such as coding, instruction following, and general knowledge, the remaining gaps are confined to reasoning and agentic tasks, and even that is closing fast because this information is from March, and in the last couple of weeks, we got two incredibly capable models, which we talked about one and we're gonna talk about the other, that is practically closing that gap as well. Another point of, data source is from Linux Foundation. Linux Foundation estimates that open models run at 6% lower cost and at 90% quality parity. So you give up on 10% that in many cases are negligible, meaning they're not gonna make a big difference on your particular task, and you're paying 80% less money or 85% less money for the same results. So the bottom line is that the Chinese open source models are dominating platforms from a download share, and they are live in production and being used more and more across multiple companies around the world, including Fortune 500 companies Clem Delangue, the CEO of Hugging Face, summarized it perfectly and he said, and I'm quoting, "Maybe in a few years, the frontier models will be for experimenting and for some really high-value tasks, and most of the production workloads will actually be powered by either private models within the companies or by open source models." And I agree 100%, unless the closed source models find a way to make their pricing and quality competitive, and then they stand a chance on staying in this game. Now, the second thing I'm going to talk about is what is currently happening at the frontier of the Chinese labs. we talked about GLM 5.2, when it came out just two weeks ago, and we said how powerful it is and how close it is to the frontier. But now we've got two new models. We got DeepSeek V4, which has just been released, and we've got Moonshot AI Qimi K3, which was just released on July 16. It's a 2.8 trillion parameter model. It is the largest open-weight model in the world right now. Its full weights are going to be available on July 27th, so just 10 days from the day this podcast is being released It has a context window of one million tokens. It is multimodal from a native perspective, so it was built as a multimodal model, it is scoring extremely high on basically all the major benchmarks. So it is currently ranking first on four out of the top eight benchmarks, including Automation Bench, including Spreadsheet Bench, BrowseComp, SWE Marathon, and Program Bench. It outperforms Opus 4.8 and OpenAI GPT 5.6 Sol, the latest and greatest model by OpenAI, and the best model other than Fable by Anthropic on multiple benchmarks, and it is consistently in the top three across more or less every piece of benchmark across the entire evaluation set. So this is a frontier model coming from a Chinese lab that is released as a full open source model with weights and everything. Now, the pricing is not as dramatically cheaper than the US models compared to other ones. Right now it is $15 for output million tokens, and on the input tokens it is $3 for non-cached input tokens and 30 cents for cached input tokens, meaning if it's something-- it's a piece of information you've used recently, then you're paying significantly less, one-tenth actually. If you compare that with Fable 5, the top model from OpenAI, the output tokens on that model are $50 instead of 15 compared to the Chinese model. If you go to Opus 4.8 is at $25 on output tokens, still much more expensive than Kimi K3. Now, if you wanna go extreme, as I mentioned, we have DeepSeek V4, which was just recently released, and if you go even to DeepSeek V4 Pro, which is not as advanced as Kimi K3, but it is not that far behind, it is 87 cents for output tokens and about a tenth of that for input tokens. So it is significantly cheaper. We're talking about more than an order of magnitude cheaper than even the cheaper models of the Western closed source models So it is very clear that the Chinese models are closing the gap, and what was discussed previously about the US models having a six to nine-month lead, that lead is shrinking very quickly and potentially completely eroded in several different cases with Chinese labs are now at the frontier at multiple aspects of using the models while still keeping them cheaper, some of them extremely cheaper, some of them are very close. What I think, it's very obvious from all the results and everything that I'm hearing, is that the fact that the US leading labs have aspects in which they are still in the lead is true, but it is probably not relevant because you can do a lot of the knowledge work that is required to be done in multiple companies with models that are probably two generations behind the current frontier, meaning capabilities that are far within the capabilities of the open source models and while paying, again, potentially 1% or 2% or 5% of the cost of the US models if you go to DeepSeek as an example. The third component that has been thrown into this mix this week is that Thinking Machine Labs, which is the company that was founded by several people who left OpenAI, including Mira Murati, which was the CTO, and John Schulman, GPT architects, and Lilian Weng, which is OpenAI's VP of safety and robotics. So a lot of people who came out of the top leadership in OpenAI founded this company in February of this year, and they finally released their first model. This model is called InkLing, and it has a nine hundred and seventy-five billion total parameters. It is a mixture of experts architecture. It has forty-one billion active parameters. It has a one million tokens context window, which is now the standard in the industry. And it also has a smaller Inkling Small model with only 276 billion models or 12 billion active parameters. Now, the benchmarks of this model are not at the top in the frontier. They are scoring well, but they're not close to the top most advanced models. But this was the plan. From the beginning, what Thinking Machines Lab said is that they wanna build models that will be effective and will be able to be used in multiple use cases and not necessarily to compete with the frontier. That being said, we've seen things like that from Meta before as well. When Meta started and they released their first model, it was, eh, not impressive, and now their second one is extremely good. So I think the fact this is their first model is actually a good point for them. They have built this to be efficient, and efficient it is. The other interesting thing about this model is that they've used self-improvement capabilities during the development of the model. So the model was using its own capabilities to fine-tune and make itself better during the development process, and it actually worked well. What they're saying that is very interesting is that the chain of thought reasoning of the model became more and more concise over time while not being a part of the plan. It just happened because of the mechanism that they have created of self-improvement. We talked about recursive self-improvement many times on this podcast, so this is not a full implementation of that, but it's first signs that it's actually running and working inside the labs themselves when developing new models. Now, this model has now been deployed across, anywhere you can basically imagine, everywhere you can find open source models in addition to Databricks and Fireworks and Together AI and a lot of other platforms The other interesting and really related aspect of this is you're saying, "Okay, why do we need another model? The Chinese models are already there. We have open source models from them. how they're even going to compete?" the way they're going to compete is the companies who invested a crazy amount of money, largest seed money any company in history ever raised, just over $2 billion. they got investments from Cognition and WQ Foundry and SkyGrid and SparkCognition and AMD and Cisco and Andreessen Horowitz and Cursor and Coinbase and Airbnb and OpenRouter. All these companies will most likely become users of the thing they've invested in as long as it's good enough. So they can drive distribution significantly faster than just a Joe Schmo who started a model just because of the rank and the clout and the connections and the impact of the companies that invested in them. Now, another thing that we talked about in the past few weeks that is driving towards open source models is what happened with the US government that has put a halt to the release of Anthropic's latest models that since then came back in a more refined or a safer version, and also slowed down or stopped for a short little while the release of the latest versions of OpenAI, that again now are also released and available to the public, at least some of them. So the concept that the models you depend on, the closed source models, can be pulled back by a wave of a wand by the US government puts you as a user of these models at risk. So having a control on everything, meaning the model runs on a server you own with the weights that you own and nobody can take it away from you, becomes a very big point. Now, we discussed last week that there were rumors that Xi Jinping, the, leader of China, has had conversations with companies like Alibaba and other open source models to potentially limit their distribution in the US and the Western world in general. He hinted that may not be the case in a address this week. So he indicated this week that open source is the future and that openness in the AI space is the key to a better global future, and that China is going to play a major role in that. So he didn't specifically say that he's not going to pull or limit the distribution, but he very clearly said that openness in the AI space is the future and that China is going to play a big role. I think that hints that potentially they're not going to pull away their models, but I think that's regardless of what he says now, whether he believes it or not, this may change tomorrow depending on what's happening from a technological perspective, whatever breakthrough that makes, the Chinese models significantly more powerful that he wouldn't want to release it at that point or whatever the geopolitical situation is going to be at that point. So while that's what he might be hinting right now, this may or may not be relevant in the reality in the near future Now, two things the CEO of Hugging Face, Clem Delangue said with regards to this, and I'm going to quote both of them. The first one is, "The biggest risk in AI is concentration of power. The way you make the world safer, in my opinion, is by leveling up the playing field and creating transparency on these models." And I agree 100%. If you allowed closed source companies to control most of the data and the processes in the world, you're just making them more and more powerful and grow a larger dependency on them, which is never a good thing. The second thing he said, and I agree with that as well, and I'm quoting: "If you're an AI company or a technology company, you don't want to outsource your core capabilities to another company, to, to a black box API that you don't control, don't have any visibility on, and don't really have any sort of ownership," which is what's happening right now if you're using OpenAI, Anthropic, Groq, et cetera. Now, to add gasoline to this fire, Satya Nadella just wrote a essay that touches a lot of these points. He calls it Token Capital, and he just published it, recently, and he talks about a few very interesting concepts in this essay And the two main concepts that he's discussing, one is called token capital, and the other one is called reverse information paradox. And I'll explain both because I think they both matter a lot. In token capital, he describes the proprietary value enterprises accumulate by generating their own AI processed data and fine-tune model weights rather than simply buying inference from third-party APIs. So think about how a company builds capital in different ways, how a company builds value for itself. What Nadella is claiming is in the future, your ability to control your own models, have them train for your benefit, for your value based on your information, is gonna be a part of the capital of a company. In other words, instead of feeding the beast, instead of letting OpenAI and Anthropic know your biggest secrets, and even if they don't use your data for training, they can use your prompts for training. They can use your methodologies for training. They can learn from how you engage in your industry in order to make their models better for that industry. Which leads to the second thing that he talks about, the reverse information paradox What he's arguing is exactly what I just mentioned. If you are allowing the leading models to learn how you work, you are giving them your information daily, constantly, in every action, in every interaction that you do. So they learn from you, so they can improve on the topics that are your proprietary methodology, while you know nothing about what they do with the data. The winners in this particular case are those who accumulate the data, which are these proprietary companies that has proprietary model weights, and you just depend on their inference, and nobody can replicate it because they're gonna consume more and more knowledge and methodologies and so on. He argued that basically, when you use intelligence, you are training, creating more intelligence that you do not control, which he is saying is a really bad thing, A, from a concentration of power, and B, from the position of your token capital that you're not building, but actually allowing somebody else to build. So he is claiming very strongly that every company has to build its own AI capabilities and train its own models and use what it's doing in order to develop this token capital internally rather than give it away. A very interesting viewpoint, especially from somebody who is running the company, who's invested more than $12 billion in OpenAI, and that owns about 30% of OpenAI, which their share is worth now hundreds of billions of dollars. and if the And if and when they go public, this may become even a larger number. So for somebody like that to come and say out in the open that's the wrong future and the right future is something else is very interesting, and we need to pay attention. We need to pay attention anyway. Satya is a very smart person, and he obviously knows one or two things about how the tech world actually works. So interesting viewpoints on how it works. So what is the future if more and more things are going to go open source? What is going to be the moat? And I'm not sure there is going to be a moat, but I have a feeling if there is going to be a moat, is inference infrastructure, right? Whoever will allow you to run your models, your trained models, your weights, and so on, the open source things that you're going to use. Whoever will allow you to run it in the most efficient way is going to win from a financial perspective. So when everything else is commoditized and multiple companies can create open source models that are maybe not the best, but are good enough for most knowledge work, you're going to go with the one that will allow you to do it the fastest, the most efficient, the safest, and the cheapest way possible. So if you can achieve safety and you can achieve speed and you can achieve consistency, the result is going to be give me the one that is going to be the cheapest option of them all. Again, with the trade-offs of speed, potentially you will be willing to pay for speed a little more, which still goes back to infrastructure Now, two other big trends that are happening right now that are going to play a role in this. One is custom silicon, right? Who controls the actual chips that will allow to do that? And this is why you've seen recently everybody jumping into this game. NVIDIA is still the 800-pound gorilla, but Meta has their own chips, and OpenAI now have their own chips, and Anthropic are talking about doing the same thing. And Google, we already know, has their own. So this is going to continue going. The other thing that we'll start seeing more and more that is still in its very early stages of infancy is on device, right? So not just custom silicon, but custom silicon that run on your local computer or on your watch or inside your headphones or glasses or whatever the case may be. So at the device you're engaging with, which will dramatically reduce the load and the need for huge data centers, because a lot of the AI work will be able to done locally, which means the data doesn't go anywhere from a data security perspective. From a speeds perspective, it's going to provide a lot of value. So we're not there yet as far as running top-of-the-line models. But again, with improvements in hardware and improvements in the algorithms and how it runs, we may get to models that are good enough to do most of the things we need that can run locally, and this can change the equation a lot as well So quick summary of where that puts us. The closed model era is not what it was just a few months ago. The gap is shrinking much, much, much faster than anybody anticipated, and in many cases, there is no gap anymore. Now, the six to nine months lead that existed is probably not real, but even if it is just a small delay, and it doesn't really matter because most of the knowledge work can use the previous version of models and still deliver high consistency, solid work, which the open source Chinese models have passed that point a long time ago. The risk from the government pulling the closed source models if you're not a US company or not a short list of US companies that the government will allow to use the most advanced models, puts you at a risk unless you are using open source. The talent exodus from some of these companies. So again, if you think about Mira Murati, if you think about Anthropic, about other labs, they all are people who left the big labs and have started companies that are now potentially pushing open source models. And now you have people like Satya Nadella, who is one of the most influential people in the tech world, and somebody who has invested billions, and that his company owns hundreds of billions of dollars of two of the, of Anthropic and OpenAI, is saying that this is the wrong way to go, and that going open source and having your own models with your own weights is the right way moving forward from a capital and the value of your company, and to reduce the risk of concentration of power So what is the bottom line? And I said that time and time again, the bottom line is we're going to see more and more companies shift to open source models, including Chinese models, but potentially US models as well, if more of those are going to be coming out, because the math just works out. If you can get 90% of the capability, which covers most of your knowledge work, and you can get it at $15 to a million tokens compared to 50, or if you go more extreme, 87 cents compared to $50, it's a no-brainer. Now, the only thing in which the closed source models still win is from a safety and reliability and regulatory compliance. Now, Anthropy came out with a benchmark that they call Fortress, and it stands for something like it's an acronym, it's not the actual word. but it's-- it allows to check the adversarial robustness score and the enterprise safety level of specific models. Even on that, the open source models are starting to close the gap, specifically new model Inkling has already posts 78% on the Fortress adversarial robustness and 98.6 on strong reject refusal rate, which is another benchmark that looks at these things. So even on those aspects, the closed source models don't have a significant lead anymore What does that mean? It means that with-- If you haven't experimented with open source models, you should start. And the potentially easiest way to do this is to connect to Hugging Face, or if you want an easier route, connect to OpenRouter. It will take you, with the assistance of Claude or ChatGPT, about five minutes to connect to their API. And then you can run your use case across four, five, six different models, some of them are open weights, and see if they are handling your use case as good as your Anthropic or OpenAI model. If they are, just look at how much it's going to cost you and then decide what the risk level are, and then you can decide which models to run which use cases. There are more and more platforms out there, including open source setups that you can install on your own servers that allows you to do this in real time, meaning it is going to send your prompt to a large advanced model to check which other models can handle it effectively, and then you can lock it in after you test it several times. And then you just work against a regular chat, and it will route it to cheaper models as needed for you to get consistent, safe results at a much, much, much lower price point Now from the topic of technology to the topic of regulation, we mentioned a little bit what has happened with the US federal government and how it has stopped the release or pulled back the release of models. But there's a much bigger story that is brewing and is accelerating, right now in the legislative side of the US, part of the world. So the first of all, that the US government has failed three times to pass actual bills, So not an executive order, but an actual bill that goes through the legislative process. So in three different attempts, one through the Great American Artificial Intelligence Act of 2026, and two other attempts, the legislation did not go through, meaning right now the federal government has not put in place an overarching, clear legislated approach to how to govern AI and all its different aspects. In two different attempts, the government decided or wanted to decide to block state-level rules from moving forward for at least three years in order to prevent a patchwork of legislation that then the US labs will have to handle. Now, this is obviously being aggressively pushed by the big labs themselves and the tech companies that are saying that this kind of patchwork will be a disaster for the labs to be able to deal with, which will slow down US innovation, which will allow China to close the gap, which they're closing anyway, and to potentially take the lead. government still hasn't actually done anything effective. This has left the door open for state level to take action, and we have seen two new bills from two different states in the US this past few days that are aiming to figure out how to handle the growing and accelerating needs and capabilities of the AI models. So the first one is in New York Hochul, and I hope I'm pronouncing her name correctly, signed an executive order on July 14th of 2026 that is imposing the nation's first statewide moratorium on new hyperscale data centers. Basically, they're pausing all environmental permits for any facility that is requiring 50 gigawatts or more of electrical capacity for a whole year The other story comes from Illinois, where Governor JB Pritzker has signed a Senate bill, that is Senate Bill 315, into law, making it the third state to regulate frontier AI model developers, requiring companies with $500 million in revenue to publish transparency frameworks and submit third-party audits for everything that they're doing in order to be allowed to deploy it in the state. But this is just a big, large example because the numbers are much larger. In Q2 of 2026 alone, states has enacted 35 AI-related bills into law Now, in addition, there's a growing community pressure to stop data center development. So let's dive into some of these topics. So the first one I'm gonna add some information about is the New York, uh, new legislation about stopping giving permits to data centers. What they're requesting is to put together a framework. So those 12 months, they're supposed to work on a framework that will define what kind of additional things a new data center needs to do in order to get the permit. that includes grid funding, so how they're going to fund building a better, stronger grid that will prevent increasing the cost of electricity to people of the state. They're also talking about providing a community investment framework with 60 days that will supposed to allow to drive additional benefits to the local communities beyond just the grid funding itself Or as Governor Kathy Hochul stated, and I'm quoting, "New York has always been at the forefront of innovation and change, but we've also always guaranteed that New Yorkers benefit. As data centers development threatens to hike up utility bills, deplete our natural resources, and create uncertainty to New Yorkers, it is my responsibility to take action and lead." Now, to be fair, this has not been a totally straight path forward, and there's a lot of backlash from that. The biggest one is the amount of time that they've put this in place. So Julie Samuels, who is the president and CEO of Tech:NYC, said, "12 months is far too long and will have the negative impact of encouraging companies to move their investments elsewhere instead of working with communities and state agencies on plans to keep energy affordable and reliable." So what she's claiming, and other are claiming as well, is while the plan is the right plan, the measures are not the right measures, meaning the goal of making sure that the New Yorkers, or I assume other similar things will happen in other states, so the citizens of a specific area do not suffer because of these new developments make absolute sense. Saying nobody's gonna get a permit for the next 12 months is basically saying, "Okay, thank you. We're gonna go to a different state and build it over there." And government laws in place, this will depend on the local states, and each state may decide a different thing. And so this will most likely drive investments into other places because the race is on to build this infrastructure, and it will happen in the next 12 months, whether in New York or in other places. So now let's switch to Illinois and the law that they have put in place. Senate Bill 315 is targeting companies with over $500 million in annual revenue or anybody that has a massive computing measurements. So basically OpenAI, Anthropic, Google, xAI, and Meta, these are the people who fall under this, thing, with an effective date of January 2028. And what they're basically saying is that all these companies will have to publish the following. One, how they measure model capabilities. Two, how they assess catastrophic risk probability. Number three, how they identify and responds to safety incidents, and that they need to have third-party auditors that will verify their compliance with whatever that is being defined And the auditor's qualifications are going to be defined in the statute of the state versus defined by the companies themselves Now, a few interesting things about this legislation. The first one is OpenAI and Anthropic both supported it. So they basically said both that this is a good step in the right direction. The other one is this was a very wide bipartisan support. So it has passed the House of Illinois by 110 to zero vote and the Senate by a 52 to five vote before the governor signed it. So this is very clear, straight past you, everybody agrees this is the right thing to do. Now, New York and California already has similar laws in place since 2025, and so this is just another big state that is moving in this direction that is requiring the labs to describe more details about what they're doing and what the risks are and so on. and like I said, the surprising thing, or maybe the not so surprising thing, is that OpenAI and Anthropic both supported it. One of the reasons they might have supported this is because they truly believe in that. The other reason they might have supported this is because it's very unclear how that is going to actually gonna be enforced. And even if it will be enforced, it's a million dollar fine for the first time you get caught not following this and a three million dollar fine, if you do this the second time or more, which is a lot of money unless you're in the hundreds of billions of dollars revenue and fundraising and so on, and then a million dollars is a drop in the ocean Now, on the community level, there has been 75 data center projects that are facing active community resistance and that actually has actions taken against them from a legal perspective, stopping them or blocking them one way or another in just Q1 of this year. This is in states like Wisconsin and Virginia and Texas and other states as well, and has targeted companies including Google, Microsoft, Amazon, Meta, and Vantage Data Centers So we have local communities that are taking legal action that is actually blocking or slowing down the development of models across multiple t- states against multiple developers. That is, by definition, slowing down the development of more data centers, which is the driver of everything as far as AI in the future. And this is a very interesting situation right now, where the lower you go down the totem pole, the more resistance there is or the more request for additional actions about how AI will grow. So at the community level, it is very aggressive, and it is happening across the entire United States. There's a few states who is trying or starting to take action, and then on the federal level, it's mostly crickets at this point when it comes to actual real legislation and definition. That will most likely change as we get closer and closer to the 2028 elections. and we're going to talk more about this later on in the first item of the rapid fire. But there is a very interesting paper, I'm throwing you a teaser, that is called AI 2040 and how that may evolve, and they talk a lot about how the elections in 2028 are gonna play a huge role in the decision of the future of AI in the United States. But more about that in a minute So what I think will happen, first of all, we've seen the US federal government getting more and more active about this. So both the executive orders that have been coming out stopping the models from being released and more and more discussions. We haven't seen any bipartisan agreement yet on legislation should look like, which leads to more and more lower level legislation and/or legal battles that I'm sure will continue until something broader will evolve, if it will evolve The next topic I want to share with you is information that was released by BCG Boston Consulting Group in a paper they called AI at Work: Strategy Matter More Than Tools that they released in June of this year, and I was looking for a place to mention that, and I kept on pushing it week after week. even though there's not a lot of time in this week either. But they have some really important findings that I want to share with you, so I'll try to make it brief. there's a link obviously there's a link to that in the newsletter if you wanna dive into that and read more. But they found a lot of really interesting things So first of all, let's talk about how they got this information. They got the information by surveying close to 12,000 frontline employees, managers, and leaders in more than a dozen global markets. So this is a very significant survey. This is not like a, somebody doing in their, garage, and they have done a very thorough work in analyzing information and providing interesting results. So I'll give you the top headlines, and again, I suggest you go and deep dive into this because it has a lot of very useful information. So first of all, 74% of frontline employees now regularly use AI, a 23% point increase from the 2025 survey, with 42% of these users are reporting saving eight hours per week or more. If you think about that, this is a full day of work that is being saved by these employees. 42% of the people who say they're using AI regularly are saving a day of work a week. This is very significant. I'm actually surprised and blown away because I work with many different companies across different industries and different sizes, and I don't see anything even close to that. So 42% out of all these people that are reporting that is a very high number, and even if that's what they feel and it's not the real numbers, it's still very significant. Now, the flip side of that, that despite the amount of time that's being saved, 66% of frontline employees receive limited or the time that gets freed up, and over half are not reinvesting it into more strategic work. So the whole promise AI is to take away mundane, simple tasks in order to allow employees to focus on better, more strategic, more important, more fulfilling tasks. And the reality is this is not happening because of lack of strategy. Now, the survey also found that employees who receive clear strategic direction on AI usage significantly outperform those who only have access to tools, even if they have access to more tools. The reality is very simple, and this is something I'm working with all the leaders that I'm working with. There has to be a plan. There has to be proper training. There has to be tracking mechanisms in place, actual real business deployment. Giving everybody the licenses to use whichever tool you want, it doesn't really matter, is not an AI strategy, and it's not going to lead to the results you want. It will add a lot of cost on that side of the, math of the organization without necessarily generating the right value. And I'm really glad that the survey found this, but on this particular case, I'm completely not surprised because this is what I see time and time again. Many companies that I come into to help them with AI training and implementation already have the licenses in place, but nothing else in place, which again, just becomes a cost with no real value because there's no strategy and process that is defined mind-blowing is that the number of organizations that have advanced to using AI to redesign end-to-end workflows or invest in new business models has nearly doubled, reaching 42% in 2026, up from 22% in 2025. This is another big deal that I talk about a lot, especially in my lectures, in my workshops. Building efficiencies around existing processes is awesome. It's gonna gain you whatever percent efficiency, 10%, 50%, 6%, 20%. If you are able to address new markets or new products or new services that you couldn't profitably do before, and now you have AI helping you do this, you can generate significantly more money, which drives significantly themselves. And again, the survey proves that as well Two more important points. one is that over six out of 10 people, so 60% or 61% if you want the exact number from the survey, believes that AI agents could perform at least half of their jobs within the next three years Now, the interesting thing about this, while you think this drives crazy fear, regular AI users are also reporting higher job satisfaction and increased cognitive load, meaning people who have learned and have been trained on how to use AI effectively actually enjoy their work more, and they feel that they're using their brain more because AI is now doing a lot of the mundane work that they had to do before Now, maybe the most important part of this, talk about a lot in this podcast, is that 72% of CEOs, double the number from 2025, now consider themselves the primary decision-makers on AI within the organizations, with 50% of them believing that their job stability depends on successful AI implementation. Now, when I work with companies, I tell them that the two most important, most critical aspects of successful AI implementation and transformation are leadership buy-in, which is exactly this, like the CEO personally and the top leadership personally involved in the process and taking ownership of the process versus telling IT or somebody else to take control over this and manage the process. This is number one. And number two is figuring out ongoing education on how to deploy the strategy that these people have defined. So top leadership defines a strategy, and then how do we trickle this down in an effective way through continuous learning and training and education to the employees are the two most critical factors, more than how good the models are, more than how many licenses you have, even more than how solid your data is. And so if you need help with that, please reach out to me because this is what I do on daily basis with multiple companies. Now, I want to finish with two quotes from different people in BCG. The first one is David Martin, the global leader of people and organization work at BCG, and he said, "Companies have moved quickly to give people tools, but many have not yet redesigned the work around those tools. Saved time does not automatically becomes value. If a frontline employee saves a few hours a week but has no direction on whether to use that time for customer service, quality improvement, innovation, or faster execution, that value can simply leak out of the organization." I strongly agree with that. The second quote come from, Sullivan Duranton, who is the global leader of BCG Tech Build and Design unit at BCG X, and he said, Employees don't push back on AI intensity. They thrive when the strategy is clear, the direction is real, and the message reaches them. Business value and employee engagement aren't trade-off. The organizations capturing the greatest business value are the same ones where employees enjoy the work the most." Both of them I will clap and cheer for if I heard them speaking on stage, because I'm a complete believer in both these two things. And again, if you're searching what to do as a leader in your organization, keep these two things in mind The last thing that I wanna dive into, and it's gonna be a quick dive, but I think it is very important, is what's currently going on with hardware inside of OpenAI. So in literally just two weeks, less than two weeks, three very important things happened altogether. On July 10th, Apple sued OpenAI for allegedly stealing hardware trade secrets. We're going to dive into more details about this. Five days later, OpenAI product, But it is nothing anybody expected. It is not the device everybody was expecting from Jony Ive. It is a $230 keyboard extension that allows you to control AI coding agents in a more effective way. So those of you who have been in the, let's say, video editing space know you can buy these little control panels to allow you to edit while also using the regular keyboard. It's a very similar thing, only to control AI agents, and it has different functions and a little joystick and a little meter the model is going to work hard and stuff like that. And there are also now more concrete rumors on exactly what is the device that OpenAI is developing with Jony Ive, and it is going to be a screen-less, movable, voice-activated home computer that is designed to have a voice interaction with a very sophisticated AI platform. But what I wanna go back to is the timeline of this whole thing very quickly and then dive into some of the components that I just mentioned. So in May of 2025, OpenAI acquires IO Products, which is Jony Ive's hardware design studio, for roughly $6.5 billion. Jony Ive was the guy inside of Apple that has designed most or all of their iconic products and then decided to venture out and build his own design studio that then was bought out by OpenAI. On January 22, 2026, Cheng Liu, who is an eight-year Apple electrical engineer, leaves Apple for OpenAI, and Apple later alleges he kept exploiting Apple network access after departing. They actually have him bragging about this in writing. This is not gonna go well in court. On April 16 of 2026, Yu Ting Pang departs Apple for OpenAI. Apple says that he coached other colleagues in Apple to get him files out of the system after he left or while he was leaving. On July 10th, 2026, as I mentioned, Apple files a 41-page federal trade secret lawsuit against these particular individuals and against OpenAI. On July 14th, OpenAI issued a fuller public rebuttal saying, "We're not aware of any evidence of that complaint has merit." And a day later, on July 15th, OpenAI ships Codex Micro, their keyboard, and then on mid-July 2026, there are new reports on Bloomberg and others about the new thing that OpenAI is actually developing that is going to be their device, that is going to be priced presumably between $200 and $300 with a 2027 commercial release. So I wanna dive first of all into the lawsuit because I think it is profound. I don't have a better word to do that, and I'm quoting a few things straight from the lawsuit itself. "This case is about Apple's former employees stealing Apple's trade secret for the benefit of OpenAI." This is from the complaint itself. They're also saying what they're sharing is the tip of the iceberg. This is a quote as well. And then the summary, if you want, sentence for all of this is, "At every level, from members of its technical staff to its chief hardware officer and in coordination with business partners, OpenAI has been stealing Apple's trade secret and confidential information. hardware business now rests on the shakiest of foundations, rotten to its core by its illegal reliance on misappropriated trade secrets." So this is from the lawsuit itself They have many clear examples of how these people, including Chang Liu and Tang Yutan and others, have had clear pattern of stealing specific designs, of downloading specific files, of encouraging existing employees to send them information, and interviewing people from Apple for hardware-related positions and requiring that they bring specific designs and information to the interviews in order to be included as part of the people that they're interviewing So does that relate anything to the smaller Codex micro keyboard release? Probably not. I'm actually really surprised with that particular release. I don't think anybody expected this. But this is basically a small little keyboard extension that has specific keys that can help you control frequent Codex functions. There's a little joystick that allows you to launch and control specific common agent workflows and a dial for adjust how much reasoning you want each agent to use. And again, those of you who have done anything such as, video editing or recording advanced studios, there's these little mini keyboards that do very specific things that are easy to use and program for tasks that you're doing. It is the same thing for controlling, agents and particularly in the first state, coding agents, but we already know that doesn't matter because these coding agents can do more or less any other knowledge work. So this is a cool, interesting device that I think will be a very geeky kind of status symbol, but nothing more than that. But what they're actually building, the home computer, which we've learned this week more information about As I mentioned, it will have no display. It will be one hundred percent voice-driven. It will be portable within the house. It will have a rechargeable battery, and the idea is that you can move it with you from one room to the other. So it's not like an Alexa that is supposed to be plugged in all the time. You can actually take it with you. So that is significantly smarter than everything we know. It has cameras and sensors to read the surroundings and actually know what's going on to enable to have more of a personalized, proactive approach to what's currently happening around it, and it can control smart home devices, play media, answer questions, and notably access personal information such as your calendar and your email and stuff like that. The price point will be two hundred to three hundred dollars, and it is probably going to be available in twenty twenty-seven. This device depends a lot on everything from the lawsuit, which means if the lawsuit is successful, it may never see the light of day. Now, I assume that this device plus the future devices that OpenAI is planning is a is going to sell during its IPO. If people think this is not moving forward, this may reduce the ability of OpenAI to receive the valuation it wants for the IPO. Combine that with everything we're talking about in the past few weeks, which is the fierce competition from other sources, reversal from token maxing to let's save as much money as we can on tokens, the improvement of open source models, and that OpenAI is in a tough spot right now when it comes to its IPO. and also its entire hardware business might be at risk. The lawsuit sounds very specific in some of the things that it is stating, and if there is truth in half of those things, OpenAI will have a very, very serious problem launching this product or at least launching it without paying Apple huge amounts of money, either as fees from every device that it's selling or to settle this lawsuit. Either way, again, it doesn't look very good for OpenAI. doesn't look very good doesn't cut it. It looks really bad for OpenAI. Also, the product itself that they're developing, the product, sounds like a really smart version of Alexa. So I can see the value in that, but I thought that what they're going to be delivering will be a lot more portable, meaning I'll be able to use it everywhere, anytime, all the time, versus just when I'm at home and I have to move it with me from one room to the other. That doesn't sound very user-friendly or very helpful. I want something that I can use always, again, like smart glasses or a pendant or a pen or something that I can take with me wherever I'm going and include in everything that I'm doing. If I'm going to expose my universe to the AI to benefit from it, I might as well do it all the way. I'm very curious to see how this thing works out in the end, exactly what value it provides, and why Jony Ive, who is one of the most legendary product designers in the world, and Sam Altman, one of the most successful entrepreneurs in history, think this is the biggest game changer we have seen from a compute perspective And now to a few important rapid-fire items. The first one is AI safety experts propose what they call a Plan A, which is an international deal to delay superintelligence until 2040 and prevent a global concentration of power. So this group has written a previous paper last year where they talked about short-term impact of AI, and they now just released a new paper that is trying to speculate how we can, as humanity, the planet, everybody, get to a situation where we benefit from AI while reducing the risks. And they're stating a very detailed plan on how this may work, but the general premise is what I said, which is let's have an international collaboration, slow everything down, allow as many people to catch up to the frontier and have a real discussion on the risks, handle them while enjoying the benefits in a slower pace. they also raise strong concern about the concentration of power through a few small list of companies that will control the intelligence for everybody else, which is something we addressed several times on this podcast, including today And basically the general idea is having a global collaboration, including mostly US-China agreement to avoid reckless competition, scaling AI within human range through 2030 to 2035. Basically allowing AI to be as good as top human expert level and not more than that, which on the current trajectory is happening significantly sooner. so we can maintain control and then responsibly advancing to super intelligence in those next five years between 2035 and 2040, implementing across everything that we're doing, R&D, and allowing international verification and enforcement of guardrails that we've defined jointly Now they're stating that there are four alternative pathways. Plan A, which we discussed. Then we have plan D, which is a race to super intelligence. Plan C, which is burn the competitive lead. Plan B, race against China. And plan S, which is shutdown of all development Now they're stating that the decision point for the US is going to be 2029, and the reason for that is it will be following the 2028 elections, which will happen at the end of 2028. They're claiming that the trajectory of the current situation of things, and they specify a lot of it, and because this is a rapid fire, I'm not gonna dive into all of those. But they have very solid, concrete reasoning why that is the case. And they're claiming because of these things that the 2028 elections, the main topic is going to be AI and its implications on society, and I tend to agree, and I also agree with the reasonings that they're using. Again, you can go and read this or drop it into NotebookLM and get it to give you a summary or drop it into any other AI and ask it to give you a summary. But they're claiming that in 2028, that will become the most critical aspect that will decide the election, and because each side will have to take its own position, there's going to be one of these paths that they mentioned that is going to be the formal approach of the US government that we'll start seeing the implementation 2029 Now, I've been saying the same things for a very long time, that I think the only thing that will allow humanity and this planet to benefit from AI to its full capabilities or most of its capabilities without introducing risks that are unacceptable is by international collaboration of as many nations as possible, multiple companies, leading labs from all around the world, including industry, including government, including researchers and universities and so on. Basically, everybody that can contribute into figuring this out needs to be a part of this. The problem is I don't see that happening. I think the race right now is too fierce, and while the plan is great, and if I'm quoting Eisenhower, plans are worthless, but planning is everything. So I think the fact that somebody put this plan on the table is important. This is a 90-page document, so it's not like a short little snippet of an idea. I really hope this can somehow happen, and I really hope it will happen before a big catastrophe happens first that then will force us to do this. if you think about what happened with nuclear weapons, well, we figured out how to control nu- nuclear weapons after they were used twice, and a huge amount of people died, and an even longer huge number of people have suffered for decades because of the outcome. I truly hope we won't have to get to that point with AI before we figure out that we need to collaborate on this together as a humanity and not as specific individuals or specific labs or specific nations. A few other important things that happened this week. A urgent warning that was signed by over 200 economists and AI researchers, including 16 Nobel Prize winners, signed an open letter that is organized by Stanford University Digital Economy Lab that is demanding immediate action to manage AI economic impact They're predicting or warning, if you want, that AI could become radically more powerful over the next 10 years, driving an economic transformation larger than the Industrial Revolution, but unfolding over a vastly shorter timeframe. I've stated exactly this multiple times on this podcast. Every previous revolution took either decades or in some cases hundreds of years, and now we're dealing about single-digit years where there's gonna be dramatic shifts to the economy. And even if eventually you will figure it out, the short to medium term extremely shaky, risky, and potentially with devastating results Now, again, they're not ignoring the positive side. They're saying that this rapid transformation significant risks, such as large-scale job displacement, and major opportunities, including substantial gains in living standards. So they do see the benefit side, the benefits of this, but they're saying that there needs to be much more proactive governance and ways that will include incentives and guardrails and institutions that need to steer AI in the right direction. Again, very much aligned with what we just mentioned before, coming up from different locations On the flip side of news, something very interesting that I read this week is that IBM is now using a lot of AI to help it with its HR decision-making process. And IBM's AI-driven HR is now hiring more entry-level employees than they did before they applied AI to do it. So IBM plans to triple its US entry-level hiring in 2026 And this is while their askHR AI system is now automating 94% of HR routine inquiries. So there's more and more AI in the system, including in the HR space, but they're hiring more and more entry-level positions, which is exactly the opposite of everything that we've seen, in, in the industry overall But if you remember just two weeks ago, we talked about several reports that many companies are rehiring the people they let go initially because of the AI transformation. Nearly 2,000 US hiring managers reveals that 32% had eliminated roles due to AI, only to rehire the same or similar number of positions shortly after. That figure jumps to 44% in finance, and 55% of leaders believe now that the AI-driven redundancies that they have thought are redundancies were actually a mistake. So what does that leave us? I'm not 100% sure. It is clear right now that companies that are right now implementing AI effectively at scale is hiring more people, and we talked about this as well. And this is very obvious to me why I truly believe that there's a short to mid-term huge competitive advantage for companies who can implement AI effectively. You can outgrow any company in your industry and run circles around them. That requires more people. The thing is, once everybody figures out how to implement AI effectively, the playing ground is leveled, and you don't need all these people because you cannot drive that kind of growth anymore. So what does the total math add up to? I'm not 100% sure. What I am sure is that very interesting times are ahead when it comes to HR and workforce in general From a model updates perspective, as I mentioned, DeepSeek have released its awaited V4 model as a full open source and optimized for agentic tasks. And as I mentioned, it runs on significantly lower cost than the US models, and it is using mostly Huawei domestic chips and not US, Nvidia chips to create this model. It is not at the tip of the spear, but it is not far from that, and it allows very advanced capabilities on numbers and rates that are, again, 1% to 5% of the cost of running the US models the model space to partnerships that again are popping up left and right. Anthropic has officially named its new AI deployment venture Ode or Ode. I'm not sure how to spell that, how to pronounce that. It's spelled O-D-E. It's a $1.5 billion joint venture with Blackstone, Hellman Friedman, Goldman Sachs and others, which is supposed to deploy frontier AI labs now putting actual employees and providing actual services to implement the AI capabilities into organization. Going back to what we discussed before, it is not about the licenses, it's about how to actually put a strategy in place to do the deployment and train the people, and so on. We've seen other labs and other companies do exactly the same thing, including Google and including OpenAI From this topic to infrastructure and how the current infrastructure may not be the right infrastructure for the future. Meta's VP of engineering, Barak Yagur, has warned that the enterprise infrastructure built over two decades since the beginning of the internet boom for humans is going to be breaking under the weight of autonomous AI agents. So he's saying that agentic queries at Meta are growing 30X in the single half year since they started deploying such, and now represents 51% of its total internet traffic. And what he's saying is that organizations faces a 20-month window to rebuild systems designed for humans to be designed for humans and agents co-creation and co-usage of the current infrastructure Now he's giving several different examples on how current infrastructure will not work and why. So I'll give you two examples. One of it, he's saying that now every software engineer can spin up 10 agents and sub-agents that will basically enabling a thousand-person organization to generate the load equivalent of 100,000 users, and they can do this basically overnight Now, the other thing that he's talking about is identity systems inside the organization, and these collapse because the agents fit no traditional category. They're not actual human users, so there's no infrastructure right now to control it. Another examples that it gives is the fact that all these vibe coding capabilities, whether it's, Copilot or Claude Code or whatever, can write most of the code in seconds, but the CI/CD, like the actual continuous deployment capabilities, including testing and so on, remain the way it was before. So we're generating significantly more code without any solutions on how to deploy this code in an effective and safe way, and many other reasons. So definitely a problem and a growing problem across multiple industries. Like everything else, it is starting with the coding because this is how these labs design it to help their own work. But the same kind of problems are going to be exposed very quickly in other industries as well And then the last thing I wanna mention this week that I find really interesting and potentially scary, I'm not sure where I land on that yet, is that 1Password, which is one of the most commonly used password, solutions where you can and then you can use them in every website you go to, or you can share specific passwords with people without giving them access to the actual password, but allowing them through a token to use, the tools that you're connected to. So they have integrated with Anthropic Claude to enable Claude chatbot to access the login credentials through a secure framework that prevents Claude from actually viewing the password, but allows it to access different things in your universe. I'm a heavy Claude user, and I can definitely see the benefit in doing so, because right now what happens is every time Claude needs to log into something, it stops and it waits for me to log in for it, so it can continue doing the work that it needs to do. Using this integration, Claude will be able to do it itself. The thing is, I'm not 100% sure I want Claude to always log into things it wants to log into, especially as I'm giving it more and more agency to do more and more things on its own. And without asking me, I will never know what it is planning to access and what it wants to, do in those platforms that it is accessing. So on a practical time-saving perspective, I'm potentially excited about this. From a control, security, and governance perspective, I'm somewhat terrified in this, solution. I must admit, I've had myself a partial internal solution for this. So I'm a Mac user, and I have multiple specific things saved on Keychain specifically for Claude to be able to use them. So I have done similar things to this, but not across the board for everything that I have access to. maybe that's not exactly how it works. I didn't get a chance to deep dive into is releasing. Maybe you can control what you allow it access to and what you're not. But I think on the broader sense, we won't have a choice in the future because most of what we will be doing across the internet will be done by agents and not by us, which means they will need access to these things, which means solution like this will become the norm sometime in the not too far future. There's a lot more news in the newsletter that we cannot cover because there's a limited time, I can record and a limited time you'll a lot of other interesting things that happened this week and a lot of new releases and announcements and things that you can find in the newsletter. You can sign up for the newsletter in the show notes. There's a link, and you can quickly get that newsletter and get access to more stuff. I will mention one more thing which is our Friday Hangouts. If you haven't heard about it yet, we meet a community of people who want to learn about AI and teach each other about AI. We meet every Friday at 1:00 PM Eastern and just talk AI, and some of the discussions are very theoretical and conceptual of where the world is going, and most of them are highly tactical on things people are actually implementing or struggling with, Everybody else helps them resolve, and people are very open in sharing solutions that they find and different things, that people can benefit from. So if you want to join us, again, there's a link for that in the show notes as well. You can click on it, you can sign up, and then you can join any Friday that you would like. That's it for today. we'll be back on Tuesday with another how-to episode that is going to show you how to implement a specific AI use case. And until then, enjoy the rest of your weekend.