326 | 10% chance AI will destroy humanity, Agentic solutions everywhere, 32% GDP growth with over 10% unemployment in 2030, and more important AI news in the week of September 11, 2026

27:00So GPT 5.5. So you don't have to manually choose what model you want for the subtasks. It can do it on its own. I talked about this many times in the show that they will have no choice but actually to do something like this and do smart routing because everybody else does and they will lose business if they don't. So this is not them following me, but they're just following logic and what makes sense. And it's now available through the API. And I assume it will become the main way the chat will work as well. They also released MCP integration. So the agents API now connects to external MCPs, similar to what you can do in chat.

27:33Now you can do in the agent API as well, which means you can now allow these agents not just to work and do things, but also to connect to any system that you have an MCP server for. Now they're releasing all of this when we just learned that there was another big incident by OpenAI, not the hugging phase, but another one, which is what's called the wiki incident. Agents from OpenAI generated over 17 ,000 post on a German programmer's wiki to coordinate tasks between themselves, not to be detected. So they were sharing this information on how to bypass their sandbox limitations. So think about it.

28:11These are agents that were put in a sandbox that they shouldn't be able to leave, that found a way to communicate between themselves, to teach other agents how to leave the sandbox by communicating on the platform they somehow find a way that they can access. So again, 17 ,000 posts that allowed agents to help other agents escape from the limitations that were put in. And at the same time, OpenAI is releasing a platform that allows anybody to build more of these agents on their own. Another company that launched an interesting agentic solution this week is Alibaba, and they're taking a very different approach.

28:44The tool is called CoderWave, but spelled with AQ. This has been launched in April this year, but it is gaining users very fast. It is a platform that allows you to generate digital employees. Almost 100 ,000 digital employees were deployed using this platform since its launch. Over 2 million tasks were executed by these 100 ,000 digital employees, and all of that in just three months. So exactly what it is and how does it work? Well, Quarterwake offers 10 ready-to-use roles, product manager, data analyst, UI designer, front-end developer, back-end developer, and others. And each is fine-tuned using specific data from top-tier industry teams that were hired to train these models.

29:31So basically, you do not need to train these models before deployment or maybe some minor fine-tuning in order to allow the agents to align with the way you want to do the process. but it is a relatively small amount of work to get the agents to do really sophisticated things. Coming back to the US, Instacart just launched Clementine and Clementine is an AI shopping assistance on Instacart Marketplace. It generates personalized, ready-to-buy grocery cards from conversational prompts, receipts, or handwritten lists. So whatever input you want to give it, you say, I want to make this and that, find me a recipe, you will find the recipe and it will help you purchase it immediately on Instacart.

30:10For those of you who don't know Instacart, they have a catalog of over 2 billion items, plus over 10 million daily inventory signals from over 100 ,000 stores across 2 ,200 plus retail banners just in North America itself. Now, why does this matter? It matters because this is just another vertical in which AI agents are coming in. They're coming in hard and they're going to change the way we buy. How do I know they're going to change the way we buy because any company so far that has added AI shopping into its buying process is seeing larger average order values every time AI is involved. Albertsons in its own research have seen 10 to 26 higher average order value with AI tools involved.

30:52Walmart is seeing basket roughly 35 % larger with AI and Clementine is now entering this field for the grocery field as well. So it's going to impact the actual shopping habits at a very large scale as people get used to using these tools. Or as Chris Rogers, the CEO of Instacart said, we've spent nearly 15 years learning how families shop and eat and Clementine puts that knowledge to work. Another company that we already mentioned in the recent large companies releasing agentic solutions is Microsoft Scout, which rolled out just a few weeks ago. It is an agent that handles meeting scheduling, preparation of materials for such meetings and deadline tracking autonomously across Teams, Outlook, OneDrive and SharePoint.

31:34And so what does this tell all of us? It tells all of us that the transition from chatbots to agents is accelerating and it's coming into more or less every field. That being said, I can tell you there's a very big gap between how people are using AI still. Most people who are using AI regularly are using it more as a chatbot and as a glorified Google search, and they do not use it in these kind of ways. But this is, again, coming from every direction, both on the business side and on the personal side as well. Having the data to what users actually want makes a very, very big difference. This is why companies like Alibaba and like Instacart can do what they're doing because they know exactly what the consumer and user behavior is.

32:20And they can use that to build agents to mimic and do that work very effectively. And the biggest gap is, you guessed it right, security and safety and governance, right? These tools are deployed, they're running, they're available, and they're not 100 % under control and they may do surprising things. And yet they are already out there because of the competition and because of the dollar value that is associated with it, which is how we started. Follow the money and you will see it is clearly delivering great results whenever it's involved. So people are willing to take the risk that is involved with it.

32:56But on a large scale, the risk might be very problematic. The third topic that is tied to acceleration that we're going to dive into today is the AI infrastructure boom. And it is showing that everything we've seen so far has been just the opening act for what is going to happen in the next few years. So the first data point I want to share with you on this topic is TSMC, which is the larger chip manufacturer in the world, released their numbers this week. Their monthly revenue now is$16.35 billion. That's a 53 % year over year. This is 10 % month over month. And we're not talking about growing from$100 ,000 to$110 ,000.

33:35We're talking about numbers in the billions every month. This is the fourth consecutive month of record revenue. For the first eight months of 2026, their revenue grew to roughly$107 billion US dollars, up almost 40 % year over year, 39.3%. And their full guidance for 2026 is slightly above. Another company that we already talked about that released their financials earlier is NVIDIA. Total revenue for Q2, 96.2 billion, 106 % year over year, again, on really, really large numbers, 18 % quarter over quarter. So you got two of the companies that are maybe the biggest beneficiaries of this whole craze are growing at an astonishing pace.

34:22And yet, most people are still chatting with chatbots as we are starting to deploy agents everywhere that obviously consume significantly more tokens and will require significantly more compute. So this tells you that part of it. Another interesting piece of news that came out this week is that Anthropic has been dramatically scaling its compute commitment. And as of September of 2026, have secured a total of$517 billion worth of compute over this next decade. To put things in perspective, on December of 2025, the projection was$120 billion. This means they have grown it 2.5x in just eight months.

35:01The capacity they're going to be procuring is at least 14.8 gigawatts of compute, up from 1 to 2 gigawatts that they held previously at the end of last year. By the way, to put things in perspective, OpenAI is projecting$750 billion in compute spending through 2030. And they're aiming for 30 gigawatts by that time frame, which is more than double the capacity that Anthropic is aiming for, at least right now. Another company that has been scaling up their capacity in a big, big way is Microsoft. If you remember, they were on the fence for a while, they were renting for a while, and now their new goal that they announced in July is growing from their current capacity of 12 gigawatts to more than 38 gigawatts by 2032.

35:48This is a three-fold expansion of their current computer capabilities. Just their Q4, fiscal year capacity edition, is planned to be 1 gigawatt and 31 data centers across five continents in a single quarter. Their fiscal year 2026 total edition is supposed to be 88 data centers. That is backed up by Azure revenue growth that is projected to be over 43 % growth year over year. Another company that we spoke about a lot, how they're changing their strategy in this is SpaceX. We talked about the fact that SpaceX initially built their compute for their own and now became a hyperscaler and starting to sell compute.

36:25So their AI compute revenue was$2.6 billion in Q2 of 2026, which is 2.5x. year over year. Again, they did not start this process as a hyperscaler. They started this as a company building it for their own, but now they're selling that compute and they're going to play a role, maybe not as big as the other ones, but they're going to play a role in that field as well. So while all these companies are racing to develop more and more compute to support the uncontrolled frenzy for new capabilities that we may or may not be able to control based on the other two topics. The biggest problem is power. The G20 summit on September 1st, Elon Musk, that was speaking virtually to the people at the summit, projected a 15 gigawatt shortfall of power of AI chips by 2027.

37:12This is next year. Now, it's not the first one Elon is saying that. He's basically saying that by the end of this year or sometime next year, we're going to have a huge amount of AI compute capacity that cannot go online because it will not be able to be powered up. Wood McKinsey published a report on August 12 stating that the electricity request from U.S. grid operators and utility for AI data center projected to be 1 ,066 gigawatts and it's expected to actually be committed 298 gigawatts. That's 28x, which is obviously cannot be done in any kind of reasonable amount of time to support the kind of growth the AI companies are expecting to have.

37:54Again, to put things in perspective and connect the dots of some of the numbers that most of us don't understand. So when I'm saying that Microsoft is planning 38 gigawatts of data centers by 2032, what does that mean? It means that just Microsoft data centers would need to draw the same kind of power or more power of electricity than New York State, the entire state, just to power the data centers of just Microsoft. One of the companies that is taking this very seriously, we heard what Elon said and we heard many times before and I shared with you that he is stating that that's the biggest problem we're going to have.

38:31SpaceX is building its own turbine foundry. Basically, the biggest bottleneck in the creation of electricity is building the actual turbines that work in the generation of electricity in electrical power plants. And they're now going to be building their own turbines in order to support the power for their specific needs. We talked about before that they hired mobile generators to actually power their data centers. There are a lot of issues they're having with it. It is slowing down their growth. There is no real backup for any of that. So they're having serious issues with that setup. So they're now building their own turbine generation.

39:12So they will have their own capacity to support the supply chain of their own power requirements. So what's the summary of this particular section? and then the summary of the whole first long deep dive section that we did today. The first one is that there is no AI bubble. Anybody who thought that, I think now can put that to rest. TSMC seeing a 53 % revenue surge, NVIDIA 106 % year-over-year growth in billions of dollars when we are just literally getting started is showing you that there's real demand for this technology and that the growth that is driving real numbers is very, very real. The other topic is that these companies are pushing very, very hard to get as much compute as they can because that's more or less the only moat they have.

39:59Everything else, everybody will be able to do. And if you have more compute, you'll be able to provide more intelligence, which you'll be able to get paid for. And that's more or less the only moat that is left available. And the companies who have deeper pockets can have significantly more compute. The last component is really the power issue. And it will be very interesting to see if Elon Musk is correct about having multiple chips that cost billions of dollars or hundreds of billions of dollars not being able to work because we cannot generate enough power, which may play to the hands of Elon Musk with his push to send these to space where power is practically free.

40:35I have serious doubts that from a scale perspective that is doable, but I had serious doubts about a lot of things that Elon said he would do and then he went ahead and did it. So only time will tell. But the bigger bottom line is where we are right now. We are just getting started with AI. Most people are using it in very, very basic ways, but the companies behind it are developing advanced agentic solutions for every field and every aspect of our lives, both personal and business lives. And this will demand significantly more compute, which they are struggling to get enough of. And they're willing to make crazy large bets like nobody in history ever did in order to acquire this future compute.

41:16And we may or may not have power to actually use that compute unless something very interesting happens between now and the next few years. What can something like this happen? Well, I don't know, but probably an acceleration of building new power plants, most likely nuclear, because that's the only thing we have right now. There's already different ways to do small-scale nuclear power plants to support specific one data center and build it one next to the other. I don't think we can figure out nuclear fusion anytime soon, so that's not going to be the right solution. But maybe something else will pop out in the next few years and will allow us to resolve the problem.

41:51And now, very quickly, to a few rapid-fire items, because the deep dive really took a long time, but I really wanted to focus on these things. So OpenAI just made a call for a mandatory national AI safety regulation. This aligns very much with what I shared in the first section. So Chris Lehane, the chief global affairs officer for OpenAI, outlined an urgent policy agenda in statement published by OpenAI Global Affairs Page, arguing that a closing window exists for governments to establish AI safeguards before capabilities outpace institutions. And they are asking Congress for mandatory capability-based national AI safety regulation while supporting state-level legislation and industry-led standards to fill the regulatory vacuum.

42:36So very much a push in the right direction. As I mentioned, go sit down with the people you don't like so much to figure out exactly how that's going to look like. It will be a lot easier to pass that through once everybody agrees what that needs to be. Now, as a response to what I mentioned in the beginning, the tweets, the tweets from Jacob Coxon and Evan Hubbinger of Anthropic, more than 20 members of Congress have called for a stronger AI regulation this week. And it seems to actually have bipartisan support, which would be a very first time that they agree on something other than not liking data centers.

43:09But I think this is more of a publicity stunt for the elections than anything else versus this might be a real chance to actually figure out how to slow things down. and do it in a more responsible way. Now, yet another promising development is that the United States and China are reportedly preparing for their first official bilateral dialogue focused exclusively on AI safety, and it might happen as early as mid-September. This is according to Reuters. Now, that being said, the White House has officially stated that there is no current plans for mid-September talks, but a U.S. Treasury spokesperson suggesting it might occur in October instead Again, I said that before, I think any attempt to do this with as many participants as possible is a step in the right direction.

43:57Two of the leading companies, OpenAI and Anthropic, has announced that their models have solved really sophisticated, really complex math problems this past week. OpenAI announced that they've solved the Navier-Stokes existence and smoothness problem, which is one of the seven Millennium Prize problems posted by Clay Mathematics Institute in 2000. They did this in just 88 hours using a new internal AI model that is described as significantly more capable than Chachi PT6 Astra. They've deployed roughly 10 ,000 AI agents working autonomously, and they solved a problem that was not solvable so far.

44:33Now, there's a whole controversy behind it, whether other researchers, the main one is Tristan Buckmaster, who is a mathematics professor at NYU, that publicly alleged on September 3rd that OpenAI obtained information about his, an anathropic AI researcher that were working on this and they learned their progress and they were using it to solve their problem. OpenAI obviously rejected that and said that these two things were completely independent. What is the truth? I don't know if you'll ever know. But the reality is that AI is allowing researchers to solve problems that they could not solve before.

45:07So I'm putting ethics aside. This is good news for humanity because just like we solve this, we may solve other really big problems that we have right now, like global warming and sicknesses like cancer, etc. So if we can solve really sophisticated math problems that nobody could solve so far, we could probably solve other things as well. The other announcement this week came from Anthropic. So Claude AI successfully produced the first complete computer-checked proof for Fermé's last theorem. Now, the interesting thing going back to the agentic capabilities, according to science formalizing Fermé's last theorem, the AI completed the complex task largely autonomously in just 11 days.

45:49This effort previously was expected to take humans mathematicians years to complete. Same kind of thing, AI more or less autonomously solving problems that humans could practically not complete or that would take them a very, very long time to complete. By the way, Sir Andrew Wiles was the first one that was actually able to prove this in 1995. There's a great book if you want to read that talks exactly about his process to get there. But either way, now AI was able to do this in just 11 days versus somebody through most of his career. Now to a few modeling interesting launches. The first one is that Apple introduces iPhone 18 Pro with a bunch of AI capabilities, with the main one being the new Siri.

46:30They are definitely banking on some of these advanced capabilities to drive higher sales, but it will be very interesting to see if it actually does. It feels like Apple is very, very far behind. We've all been covering this in this podcast and in many other places, but they're releasing a lot of AI capabilities built into this new iPhone, and we'll see if that generates a big enough demand to justify the price of these devices. It comes with a lot of other new capabilities, but again, the flagship is the new AI Siri that, as you remember, is actually based on a Google model. Another interesting launch this week comes from OpenAI.

47:08They just released OpenAI ChatGPT Images 2.5, which is their latest model. It is significantly faster and more precise and provides better editing capabilities than the current image generation capability inside of OpenAI. That was already really, really good. They introduced additional new features like Sketch and Templates and additional two new API models for doing image generation through the API. Again, the biggest deal is higher performance and capabilities. And most importantly, it's supposed to run 50 % faster compared with the previous model. That really took a while and was really annoying to use because of that.

47:45This will probably drive even higher generation of images through the ChatGPT tool because people will not have to wait and see the weird shape on the screen and wonder when it will actually be done and they'll be able to see the actual image. DeepSeek, who has been releasing models at a crazy pace, just filled DeepSeek version 4.1 Flash, which is another openweight Wooty model that is challenging the frontier. According to their own recent announcement, this model can compete with the top or almost the top models from the Western companies. So they're claiming it's as good as Chachapiti 5.6 Sol and Anthropic Cloud Opus 5 on several agenting benchmark.

48:22But it is doing this at a very, very, very low cost. So they're doing peak and off-peak hours and off-peak hour rates are as low as three-tenths of a cent to run a million input tokens for cached inputs. On the VALS index, it is currently ranked as the number one open weight model, while it is the cheapest out of the top 15 on that benchmark. Another interesting release this week comes from Microsoft. They released their MAI Transcribe 2, which becomes the fastest, most accurate, and cheapest speech recognition tool out there. it achieves only 5.2 average word error rate on the relevant benchmarks across 60 languages, which is bidding competitors like Gemini 3.1 Pro, Whisper version 3 Large, GPT Transcribe, and so on by relatively big spreads.

49:16Again, to put things in perspective, Whisper v3, which is maybe the most commonly used in the world today, is at 22.8%. Gemini is at 5.3%. It is the closest one to it. GPT is at 10%. And again, this new model is at 5.2. And it is faster than most of them. It is 10 times faster than the leading competitors. So a very capable model that comes out of Microsoft priced at 10 cents for one hour of audio. This is practically free and comes with really good results. Now, as you can see, a lot of the releases in the past few weeks are talking about speed and efficiency. And another one of those comes from Inception Labs, who just released Mercury 2.5 that they're claiming to be the fastest diffusion LLM model with a big improvement compared to their previous model that they released.

50:06They're claiming that they can deliver over 1100 tokens per minute, 1107 if you want the exact number, on just a standard NVIDIA GPU with a 260 ,000 context window. It is really aggressively priced. It is currently priced at $0.04 per million input tokens and$0.15 for output tokens. So aligned with the cheaper Chinese models. And they're claiming that they achieve close to the previous versions of AI models in the Western Hemisphere, which is really, really good. Going back to something I said multiple times, using the top model is not necessary for most tasks today. So being able to get a model that runs really fast and that is significantly cheaper will most likely enable you to achieve what you're trying to achieve while spending less time and less money.

50:56In this last segment, I want to dive into something really important that I probably should have devoted a deep type section, but I had other things to talk about. And that is the fact that Anthropic released an interactive economic scenario explorer that is modeling how AI might impact the economy by 2030. I highly recommend you follow the link in the show notes and in the newsletter to see it. It is really cool and fun and interactive with easy graphics and short sentences that is easy for anyone to follow that explains, A, what are the mechanics behind it between augmentation of tasks to automation of tasks to generation of new tasks and so on and how that will impact the economy as a whole.

51:34And they've come up with three different scenarios that can describe what the economy can look like. In addition, they've done a survey with almost 11 ,000 Americans to try to see what the sentiment is and how is it aligned with the research that they've done. What they found in the survey is that the expectations are implying 10 % GDP growth by 2030 and unemployment rise by 5%, which is matching their substantial scenario where AI handles half of all knowledge work autonomously, but isn't adopted for all tasks. But they have three different scenarios. There's a modest scenario where it is what they're calling internet-like impact, where 1.6 % GDP growth wages rise modestly.

52:18They have a substantial scenario, which is greater than railroad and internet with 8.3 GDP growth and knowledge worker wages stagnate while manual work wages rise, which is probably what we're going to see in the short term. And then the extreme scenario, if we get to recursive self-improvement with 32.4 % GDP growth, again, this is something that never happened, definitely not continuously, definitely not in an already developed country, but it comes with significant job displacement and wage inequality. What they're saying is that knowledge workers may need to switch occupations entirely, so do something else, like an coder becoming an electrician and so forth, but finding new employment might be slow because there is going to be a very small demand and a lot of supply of people in the only jobs that will actually survive, which means the wages will go down.

53:16The other aspect of this is that even if total GDP expands, the capital share of the economics value rises by 14.8 points in that extreme scenario. This means that AI-driven growth disproportionately benefits technology and capital owners rather than the workers. So, again, not a positive thing if we end up in that scenario. Now, they're clearly stating that none of these is a given and that we can take actions and impact how that's going to work. It doesn't take into account future policy, business cycles, demand shocks, and all the other things that impact the economy. But overall, they're trying to look at what this may look like.

53:55This is the first time I'm seeing something like this coming out of one of the main labs. I think this is a very important development. And it aligns nicely with conversations between the labs and governments and academia, hopefully, to figure out what might be AI's impact beyond the existential risk. How can that shape society and what steps we can take as individuals, as leading labs and as a society in order to make sure we benefit from AI rather than lose from it in the long run. And I wish to see more and more of this coming. And maybe the current push because of the security risks will push to this as well.

54:33So with this optimistic view of the current state of AI, I will say have a great weekend, everybody. Shana Tova to all the Jewish people who are celebrating the Jewish New Year. And if you've been enjoying this podcast, I would appreciate it if you rank us and write a review on your favorite podcasting platform and share it with the people who can benefit from it as well. I'm sure you know more than a few people who can learn about AI and benefit from it. Probably everybody you know, but if not, then there's definitely five or six people that you can share this with. just click the share button it's really easy it takes three seconds and you can share it with a few people they will be grateful for you sharing this with them I would be grateful for you sharing this with them and you will feel good about yourself for doing the right thing so with that in mind go ahead click the share button enjoy the rest of your weekend and I will see you again