338 | AI agents now work 6x the hours of their humans at Anthropic 🤯, AI just put 70% of all websites at risk, Is AI a bubble? “We grew a mind” (OpenAI safety insider), and more important AI news for the week ending October 23, 2026

Hello, and welcome to a weekend news episode of the leveraging AI podcast. The podcast that shares practical ethical ways to leverage AI to improve efficiency, grow your business and advance your career. This is Isar, may this your host and a lot of really interesting things happened this week on the macro level and we're going to cover three of them. One is the conversation, are we in an AI bubble or not? And we are going to follow along from a conversation that was on the Ezra Klein podcast with two experts on the topic, one representing each of the sides. We're going to cover what they discussed. The other topic we're going to cover also comes from Ezra Klein's podcast. In this particular case, he had a really fascinating interview with David Robinson, who is the open AI employee that resigned just recently and shared that he did this because of his fear from the current situation. He's the guy that wrote all their alignment papers and how they're distributed to the public and to the company. So he's been inside the processes of safety and alignment inside of open AI and he shares why he resigned and why he thinks we're in a very dangerous era. And then we're going to talk about two very interesting papers that were released by ontropic that talk about how is AI being used inside of ontropic right now and how is it being perceived as far as its impact on the future inside of ontropic compared to the rest of the population like the general US public. So three very interesting deep dives that we're going to talk about. And then we have a lot of other topics to cover. Some of them have to do with safety. There was a big vulnerability found inside of PHP this week, which is the infrastructure back in infrastructure for about 70% of the internet. We're going to talk about hackers hacking into banks. We're going to talk about Sam Altman and three different pathways that he sees for AI safety and the future of AI and new releases and agents doing really cool things. So we have a lot to cover. So let's get started. First story of today is we're going to cover are we in an AI bubble or not? And we're going to do this looking at the information shared on the Ezra Klein show, which is a New York Times show where he interviewed two leading people with two opposing view of this. One of them is Paul Kadroski, who is a partner at SK Ventures in an MIT fellow. And he believes very aggressively that we're in a bubble. And then Azim Azar, who runs exponential view and tracks AI financial stress. And he believes that we're not in a bubble yet. What I'm going to do is I'm going to share with you what each and every one of these people shared as far as a support for their view of things. And what the other person countered and responded to that particular claim. I will say something else, they both had inaccuracies in the way they presented the information. I don't know if that was intended or not. There was a much more miss represented information on Kadroski's side. But I will share what I found about the accuracy of the information for each of those people as well as we go through the facts. So we're going to start with Kadroski's approach. And again, he believes that we're seriously in a bubble. And what he said is he said that, first of all, every general purpose technology built in history ended up overbuilding and getting into a crisis. And he talked about multiple examples from electricity to railroads and so on. And he mentioned specifically that railroads left about 50% of track useless in the build out of railroads in the beginning. Azar's response to that was that we're not in an overbuilt yet. And that we're vastly under supplied as far as compute. And he's saying that maybe one or two of the labs will go bust. But the underlying need for compute and the technology is very strong. Or the way he some did opposite anthropic may go out of business, but Nvidia won't now going back to the claim about the railroads that is very much misleading. Yes, over a century, about 45% of peak mileage was abandoned. But it wasn't something that happened immediately. And he wasn't overbuilt in the beginning just over time. There were other more effective means of delivery, which made some of the railroads again about half, but over a full century useless. Now, the other claim that Kadrosky made was that the build out is fueled by credit that has spread into insurers, private credit and sovereign funds. And then he will reach $10 trillion by the time it's all said and done. And he's saying that a lot of it is moving from the actual companies themselves. Financing through their cashflow to more and more debt to which as ours responds, what that debt is rising as far as financing the leading investors inside of AI, but it's not at historical high levels. He believes that the threshold where it will start getting dangerous. And there's too much debt in the financing is 30%. He's saying that we're getting close to that, but we're not there yet. To be fair, I checked that fact as well. And the debt aspect of the hyperscalers capex was 9% in 2024. And it is 30 to percent in the middle of 2026. So we are definitely at that threshold. Obviously, the $10 trillion figure is not anything anybody can check. This is just his own personal opinion, I assume. Going back to his third point, he's saying that markets are very fragile. The top 10 stocks are 40% of the S&P 500 and half the worth of US people. Right now are inequities. So there's too much concentration and all of that. Azim did not respond to this particular thing. Farza numbers from Kodroski, they're very close to the real number. The top 10 stocks are currently 39% and about half the net worth is in equities. Right now, it also very, very close. It's 46 to 48%. So these two numbers check out. Now, one of the main things that separated two of them is about how to view the demand. Azim's perspective, there is a, we're just in the beginning of the demand. And there will be a growing demand for tokens significantly. Which I agree with 100%, I personally think there are two orders of magnitude growth. That is ahead of us as far as the use of tokens to which Kodroski responded. And I agree with that as well that token price has failed 13x this year. And that normal users, days to day users do not need the frontier models. And so the real demand shouldn't look at how many tokens we're going to consume, but how many tokens we're going to consume times the price of tokens. And what he's saying in this, that particular case, he sees a huge drop in the potential revenue of the companies as the models race forward. While most people don't actually need the frontier, which I agree completely with the second half of the statement that most people today do not need the frontier models. I don't know where that lands as far as the actual cost versus return of using the cheaper models. And how much will that be a viable business model for the leading labs? But it's definitely a good point. Now, another thing that Kodroski made as a point, he said that demand is very narrow. And that right now about 80% of tokens revenue is coming from coders with still a very big concentration where meta and Microsoft are driving half the demand from ontropic based on the papers they shared. Again, a very big misrepresentation of the facts. First of all, coding right now is about 55% of the demand. And it is actually shrinking all the time. And it is because they initially developed the coding capabilities so they can help themselves on talking about the labs. And now that more and more people through the different aspects of our economy is learning how to use it, that number is actually falling. This episode is brought to you by the multi agent orchestration course. It's a course that I have been teaching personally since April of this year and every single one of the cohorts that we have opened have sold out. And the people who have went through these courses made incredible transformations for themselves, switching jobs, starting their own companies and so on. And definitely for their businesses by being able to build really complex processes by following our methodology and our infrastructure settings that we teach during the course. The course is four sessions of two to two and a half hours each over four consecutive weeks. And through it, you can go from knowing a little bit about how to use AI and using it in ways that most people are using it today to learning how to build really complex sophisticated systems that can automate more or less everything in your business. The current cohort that is open starts on November 2nd. And it's going to go on November 2nd, November 8th, November 16th, and November 23rd. And we are almost sold out for that cohort as well. It is very likely that we will not open another cohort in December, meaning the next one after that is going to be in January. So if you still want to learn how to build really incredible automations and learn everything that I've learned by doing this every single day for the last few years, don't miss this opportunity to come and join us on November 2nd. There's a link in the show notes. And if you're going to use the promo code leveraging AI100, you're going to get $100 off of the price that is on the website. Like I said, there's a few seats open, but they are selling out fast. So if you're interested in doing this still this year, go and click on the link right now on your phones and complete the registration and the payment. So you can guarantee your seat. I'm looking forward to seeing you there and now back to the episode. The other aspect is also a misrepresentation because what Anthropic shared is that early last year, Microsoft and Meta were half of their revenue. And that number has dropped significantly into 2026. And it's continued to drop as 2026 continues. And so not accurate information from Kodroski, but definitely there is a concentration where a lot of the revenue to these labs are currently coming from a relatively short list of companies. I actually think this shows exactly the opposite that there's still a huge opportunity for growth because most companies are not using it at night, they will as well. Another point that Kodraski made to say how high the risk is, it says that basically these three IPOs between SpaceX and OpenAI and Anthropic is sucking up all the equity across everything else that is happening, and that concentration leads to very high risk because if one of them fails, we're in very big problem. He said that these three IPOs are larger than all post-war war two tech IPOs combined and that they exceed the NASDAQ peak valuation, both these claims are incorrect. But being said, SpaceX fell from $225.64 to a low of 104 hours and 83 cents which is less than half, now we recover back to 153, but it's still a very big drop from the peak immediately after the IPO, which tells you again that the valuations might be overstated, at least in the SpaceX perspective. The three last claims that Kydrosky made before we switched to Azeem, so he said that AI is pricing labor faster than any prior technology and that new jobs will not keep pace and I agree with that statement 100%, Azeem is saying that humans are moving to verification roles and new more strategic roles and new roles will appear, I do not agree with that. I think we are heading very quickly into a situation where we're going to have very significant unemployment and while new roles do appear and pop up and I'm helping companies hire for these new roles, heck my role in some of these companies that don't exist a few months ago and definitely not two years ago, but that is I think a very small number of roles that cannot cover for all the roles that will be wiped out by AI, even with the current AI technology without any improvement. He's also talking about that regulation is inevitable because some catastrophe will happen and he's giving again, math that just doesn't check out as far as if we have 12 different things, each and every one of them with 5% of happening, he's claiming that goes to 50 to 60% that it's actually happening, the real math is actually not that. If you have several different things, 6, 5% thing and so on, you're getting to significantly less than 50% of the chance of that's happening. So again, he's making these claims and in making them with a lot of conviction, but many of them are either wrong or misrepresented, meaning he's using actual numbers, but not representing them on what they actually represent. And so I'm not impressed with the way he moved forward with making his arguments, I'm not saying a disagree with all of them, I'm saying use real facts and use them in an ethical correct way versus either making numbers up or manipulating them and saying they mean one thing when they actually meant something else. Now let's talk about Azim Azar, he's saying that the demand is real and enterprises are paying for more and more compute. He's saying that GPU owners that earn very good returns 18 to 24 months in, he's saying that the chips themselves are 65% of the data center costs, which is showing that the investment is going for less overhead and more actual substance, which again is not completely accurate. Bernstein research puts GPUs at about 39% of data center capics. So less than half versus almost two thirds. So again, another misrepresentation of the actual numbers to make a point, he separates the frontier labs from the GPUs, as I mentioned before, he said yes, anthropic may go out of business, but Nvidia won't and I agree with that, like the labs, not all the labs will be able to continue running at the level of capacity they're running right now, financing will dry out eventually and they will have to be profitable. By the way, speaking of this, anthropic in their latest releases in preparation for their IPO are claiming they are profitable, that ignores obviously a few different numbers there, but on paper, they're profitable right now. And if they're going to keep on growing in the past, they're saying they're going to keep on growing, then they'll definitely be profitable regardless whether on paper or not. They both address the firmest failed IPO. So those of you who haven't heard that they are a new cloud company from Australia, who try to go public and was not successful to get the guarantees for the funding they wanted. And the IPO was pulled, what Azim is saying about this is that markets are basically pricing the risk and that's why they weren't able to raise money to which cadraski's response was that the reason they wasn't able to raise the right amount of money is because there are two trillion dollars of mega IPOs that basically took the entire capital of the market and not allowing anything else to run. There's probably truth in both what these people said. Another thing that he said when he comes to the general public, he said that very few Americans currently pay for AI, about only 2.5% of Americans privately pay for AI, which means there is a huge room for growth, if only two and a half percent of Americans pay for the AI they use and everybody is using free models, then obviously there's a huge place for growth. Cadraski is saying that people are just happy with the free AI. So the meta new muse app that is growing like crazy doesn't cost them any money and it's probably doing good enough for most people need on the day to day. The question is obviously can these labs continue to deliver it for free, which goes back into another part of this equation. And then the last component they both disagree on is whether this is actually providing more value to the US economy, meaning are we seeing productivity rise and impact GDP or is it just all smoke and mirrors and marketing of the labs, obviously cadraski is saying that it's not, that it's been flat and we haven't seen any growth in productivity. The real numbers are the productivity rose 2.2% year of a year in Q2 of 2026 and GDP grew 2.2% in that time frame as well. Now it is below peak growth over historical years, but it is definitely growing and productivity is getting better. So from that perspective, we don't know if it's AI or not again, it's very hard to know, but it is definitely not flat, which was the claim that was made by Cadraski. Now this was actually the final question from Ezra Klein. He asked both of them what would make them change their mind. And Cadraski said that if he would see significant growth in productivity and that AI is driving in significant growth in the GDP itself with direct correlation, Azar said that he would change his mind. If you saw US enterprise AI spending half by half two quarters in a row, which basically saying they're pulling a lot of capital in investments away, which is not the case right now. Now let's talk about what do I think about it and what does this mean to your business and to your money if you have money in the stock market. What I think is that I'm on the fence between these two people and between these two things. I'm not 100% sure yet. There's definitely a crazy stupid amount of money that is flowing into AI right now and everything that's tied to AI, like nothing I remember ever in history may have happened, you know, many years ago. But in the last 30, 40 years, nothing has attracted so much capital as AI is bringing in right now. Tapping said the previous round of this that was something similar was the internet, right? So the explosion of the internet, the bubble, the previous big bubble other than the financial crisis of 2008, a very different kind of thing. But in 99 in 2000, I was already in the startup space. So I remember that vividly and what happened, it feels very differently because back then people were selling concepts and ideas that will never materialize like there was websites for cats that were worth billions of dollars. And right now there is real demand and I believe, I truly believe that we're just in the early days of this demand. I work with these companies every single week. I meet with CEOs of companies all the time. I know how much behind most of the world is compared to the things we're talking about in this podcast, right? So the demand is going to grow, I believe, to orders of magnitude at least in the next two years. So the demand will be there. Will the demand with the drop in the cost of tokens and cheaper models and competition from China and open source or all that, all of that will be enough to make the ROI work to this insane level of investments. I don't know. I truly don't because I see right now myself using more and more the cheaper models, either open source models and definitely the cheaper models from the leading lab. So I use Luna a lot from open AI and I use high coup 5.5 more and more. And so I would say that right now probably 50% of my usage, which is at the max plans on both companies is on the cheaper models. They're still performing well and giving me a lot of value, which means as these things get better and cheaper, you may not need the really expensive models, which may not pay back for the investments that these companies need in order to stay afloat and be profitable. That being said, I also think that they will be able to tweak down the investment in the frontier and slow down their efforts and then hence they will need less money in order to continue running their operations and just deliver intelligence instead of develop new intelligence or probably not a one in zero game, but change the focus to more deliverability and fast and efficient and less development, which will require significantly less funding, which may make them profitable. So I do not know how the supply demand graphs where will they cross each other if they will, but I'm definitely thinking that there's still huge demand coming. What does that mean? I think that this entire conversation doesn't matter and I think it doesn't matter because there are a lot of edgy people out there. A lot of edgy investors, both individuals like you and me, or much bigger investors which have some money already tied into this or not, who are terrified of the level of investment that's going in right now. And so what I think that will happen is any bad news and there is a very wide range of potential bad news that can happen, either any slow down, any bad model that's released, any big cyber security situation, basically any serious bad news that are coming. to come out of any thing in the AI space will send the market crashing. That's what I believe. Now, again, I'm not providing any investment recommendations. I'm really bad when it comes to investing in stock markets on the last person you want to follow as far as making decisions on how to invest your money. But this is what I truly believe. I think that a lot of money is being pushed in, a lot of people are on the edge of whether we're in a bubble or not. And with any sign of bad news, a lot of people are going to pull their chips off the table and because about 50% of the S&P 500 right now is tied to AI, it is going to send a very significant correction, especially at the very, very high levels of the stock market that we have right now. Does that mean that we are in a bubble? I don't know. I do think that the financial implications of this could be very similar to what a bubble exploding looks like, even if the fundamentals are solid and correct. Now, I want to switch to two very interesting papers, articles that were released by Anthropic this past week. One paper shows what AI has already done to work inside Anthropic, probably one of the most leading organizations in the world when it comes to using AI because they're developing it themselves. They have access to the latest frontier models before we all get access to and they are working as fast as they can to be faster than anybody else in their industry. So how has that changed? The other paper, as I mentioned in the opening to this episode, shows what Anthropic on staff think that AI will be able to do in 2027 and then 2030 versus what people in the general US population. Each paper is interesting on its own, but putting them together gives us a very interesting view of where we are and where we're probably going. So the first paper, the internal Anthropic paper, uses its own internal data, HHR records, agent logs, Slack channels, calendars, and internal surveys and hiring notes all from Anthropic themselves. The second paper surveyed 1,334 Anthropic employees and almost 11,000 US adults to see what they think about the future of AI. And just as a teaser, Anthropic staff expect AI to do 75% of all their work by 2030. And regular Americans think AI will be able to do 28% of their work by 2030. The question is who is right. So let's look at some of the other piece of information that's coming from these papers. So we're going to start with the first paper and what's happening inside of Anthropic itself. As of March of this year, agent work hours has passed the human work hours for the first time and now the spread towards agents is growing all the time. By August, agents working with each employee logged about 6x the hours of the person that's running them. So again, we went from humans doing more work in the beginning of this year as far as hours to parody in March to 6x in August. That shows you how quickly the agent work is being adopted in a place that is at the frontier and how much crazy capacity it generates for additional work. If you want to look at this from code commits, remember at the end of the day, they're writing code. Some of them, obviously, there's a lot of people doing a lot of other things more about that later on. But code commits are now 5x where there were just 6 months ago. So Anthropic with roughly the same amount of people are generating 5x the code. They did 6 months ago. Now, that being said, they're not giving up on quality. So agent written code was deleted or reverted no more often than human written code when looking at it to that length. So from a quality perspective, there's no degradation in quality on a 5x quantity of results. Now, it's not just writing code. If that's what you're thinking, coding agent sessions by non-technical people. So marketing people, admin people, HR people, went up from two a month in August of 2025. So just over a year ago, to 90 in May of 2026. And the results are showing as well. So lawyers handled more contracts, accounts, payable, handled more invoices, sales, closed more deals, and so on by a big spread. Just to put things in perspective, the salespeople are not doing 1.9 the amount of closing sales because they're using agents compared to what they did a year ago. Think about where your company would be if you can double your sales success rate without extending the number of salespeople that you have. These are insane numbers. And yes, it's not 5x writing code, but it's doubling the number of deals that are coming into the company without increasing the amount of salespeople. This is absolutely mind blowing. Now, part of the reason this is happening is because AI is replacing all the old tasks and allowing the people to do the more important tasks. One account executive shared that Automations replaced about 20% of his old task, allowing him to talk 30% more customers in the same amount of time that he invested in his job before. So he's focusing his time on what matters as an executive that is dealing with sales, which is talking to clients versus the other things that he was doing before, and it's the same across the board. Revenue per employee is 2 to 4x, what meta and Google had at the same age of company. 2 to 4x of some of the most successful companies in history, just because they're using agents and they're using more and more of it. Now, staff in their survey, staff agreeing that their work setup makes them as productive as possible, went up from 44% in late 2024 to 75% in April of 2026. Basically, people saying we cannot be more productive than we are right now because of the environment that we're being given by our company is now three quarters of the employees in Anthropic. So this is not a forecast. This is what's currently happening at Anthropic, and the mindset, and that's a quote from the actual paper, which I completely feel the same, saying anytime you do something a second time, you kind of like, okay, how can I use Claude to make it so that I don't have to do this a third time? This is one of their sales executives, this is not a person writing code. One of their financial managers talking about the purchase review agent that they built on their own said that I'm quoting, it's just trying to find the things that take a lot of hours so they can focus on the other stuff. And so they are running significantly faster than they run before. They are running faster and faster, and it is not just the people writing code. But as this is happening, it took a very serious swing at entry level jobs at Anthropic. So the junior share of technical hires fell from 27% to 9% since early 2025, with the biggest decline happening since December of 2025, which is very obvious, that's when the better version of Claude code was released, and then Claude cowork, and so overall junior technical staff fell from 21% to 14% just in that time frame. And this is happening while Anthropic is growing fast and hiring more and more people, including more and more technical people. So it's not that they stopped hiring, they just stopped hiring junior people or they didn't stop, but that portion has shrunk dramatically. Moreover, in April of 2026, about a third of the ML managers, machine learning managers, and fifth of the software engineering managers said that they were hiring less people because of AI. So this is not just junior roles, this is across the board. Now to tell you the big picture of this, if you think about Anthropic, they are at the end of the day, a software company writer doing software research and AI research and they have to write a lot of code. Technical roles went down from 61% to 56%. So about a 10% decrease in the technical staff to a company that is developing faster and faster technology, it tells you what AI impact is. Now the reports authors talk about the issue with protecting future jobs and especially entry-level jobs, and they're saying that employment protections for incumbents will not benefit those not hired to begin with, private action on the part of employers must come alongside, far-sighted government policy around retraining and income support. So this is coming from the horse's mouth. This is Anthropic themselves saying we have a serious problem and if the government doesn't step in, protecting jobs and protecting income, we have a problem. And again, you've heard me say this a million times on this podcast. I think we are racing into a very serious income and unemployment problem that is just going to grow and I'm not sure the government has a solution right now on how to handle that period. Now let's look at the second paper that they released which has surveyed, as I mentioned, internal employees in Anthropic and compared to what they think about the future of AI compared to the general population. So let's look at a few key things that were shared and again, there's like probably 20 different things that they asked people about. But a few key components is the share of work that AI will be able to do by 2030. Again, Anthropic staff said 75% US adults said 28%. AI capable tasks actually done with AI, 65% by Anthropic, 42% by the US population. Here's an interesting one. Do you believe that AI will be able of making noble prize level discoveries by 2030? US adult population 39% Anthropic staff 91%. Basically almost everybody on Anthropic thinks that within three and a half years AI will be able to make noble prize winning kind of discoveries. This is novel new science, not repeating, not understanding current stuff and explaining it, not analyzing data, but novel new discoveries. Again, the highest level awards that humans can get for research. Almost everybody on Anthropic thinks that it's coming in the next three years. What do you think the GDP growth can reach because because of AI, US adult population, said 5.4% anthropic stuff says 17.1%. Now, again, that's probably the least thing that they know about when it comes to their understanding because they are, you know, people working in a company, they're not economists, but just the fact that they think it's more than three X. The rest of the population is very impressive. They also had a graph showing different things, different specific tasks and where the general population thinks it can do right now and then in the future. So I'll give you a few, it was routine emails, bookkeeping, tax return, contracts, software products, published study, online businesses, and Nobel Prize level discoveries. Again, I give you some of them, but the crazy thing is for anthropic employees, the question there, again, they asked them about 2027 and 2030. Everything I said, so routine emails, bookkeeping, tasks return, contracts, and software products, and published studies are above 90% in 2027 for the survey inside of anthropic. So between 90 to 100% on all these topics. And if you go to 2030, everything is above 90% for all the different categories with the two lowest ones being online business at 96% and Nobel Prize level discoveries at 91%. The highest level that general public thinks about 2030 is routine email and bookkeeping at 75 and 71% accordingly. So a huge spread between these people. Now they were also asked themselves about unemployment and US workers who expect AI to do less than half of their work expect that the worst unemployment could be 7 to 8%. Workers who expect AI to do more than 80% of their work forecast 12.2%. I actually think that the number will be significantly higher. By the way, 12.2% is the peak unemployment for the 2008 financial crisis, and it melted the world's economies. So if we get higher than this, which I truly believe by 2030 we might and by 2030 too uncertain, we will, then if we're getting to a 20% unemployment, we are in a very, very big problem. The bigger problem is a lot of the people that are going to be unemployed are not people who make 30, 40, 50,000 dollars a year, but are people recurrently making 150, 250, 350, $150,000 a year. And when they get pulled out of the economy, the impact on the economy is significantly more severe because they provide most of the liquidity into the markets. And so I'm not optimistic with that at all. And again, the other parameter that they shared, and I'll try to explain a few graphs for you. When trying to connect the dots between these two pieces of research, a few things are becoming very clear. One is the more people understand how AI works, the more they think AI will be able to automate their work. So it wasn't just the general population versus anthropic employees. They also had a breakdown inside the US population to how much AI people are using today. So I'll give you an example. People who are using less AI thought that AI will automate 25 to 28% of their work, as I told you, anthropic things, 75% of work. But people who currently in the general population who use AI at work regularly thought that AI will automate about 40% of their work. So this is significantly higher than the people who are the average person in the US. So what it tells you is that the people actually hold the knowledge of what AI can do today, expect it to do significantly more. And so the people in the general population are just clueless of what AI is capable of doing. And as somebody who goes and speaks to CEOs and speaks on stages regularly and talks to multiple CEOs and leading teams in multiple organizations across industries and all around the world, I can tell you that most people are clueless. They don't understand what AI can do today. Hence they cannot project what AI will be able to do in the future. Which tells me that probably the anthropic employees are a lot more accurate with their projections, which tells us that AI will be able to do most jobs, most knowledge-based jobs within the next few years. The other two points that I want to make is how much productivity and how much adoption there is going back to the first paper and then connecting all the dots together. So looking at the productivity boom across functions and it's an explosion that started in December of 2025, based on the anthropic paper, is this. The number of commits per engineer, so basically writing code rose 5.1x. So 5x. But then there is all the stuff that is below that. So queries per data scientist rose 3.1x, contracts per legal employee rose 3.1x, deals per salesperson, rose 1.7x, hires per recruiter 1.4x, invoices per AP processor 1.4x. So you're saying, oh, still the developers are running far ahead and everybody else are seeing some gains. Well, it's not some gains. 3x is an insane gain or even the 1.4, that's a 40% increase. Think about if the slowest growth in your company, your biggest bottlenecks are 40% growth that you can drive you over year. But the real thing that this thing is ignoring is two things. One is the fact that these companies, first and foremost, invested in technology that will allow themselves to run faster hence writing code. Not that they figured out code, they're going to invest a lot more in making AI as good in everything else. There's really no difference in teaching AI to be really good at accounting operations customer service as it is to teach it to be good at code. The only difference is they invested more in that first because it gave them an immediate benefit. But they're going to invest the same amount in everything else, which means that 5x growth can happen in every single aspect of knowledge work. Some of them will take a little longer, some of them will take a little faster, but it's all going to happen. The other thing that it's driving is the adoption level that is tied up to this. So as I told you before, they looked at how much agent hours are spent per employee versus how much humans are actually working. And that number, as I said, was humans were working more up to April of this year. And that number has changed we like the middle of May for all the employees who are not writing code. And right now for researchers in the organization, it is 19x that the agents are doing compared to humans. And all employees is 6x. So what's happening here is that we're going to get systems that are better and better and better at doing the work and driving a huge increase in productivity. And we will get people using them more and more, which if you can grow at a crazy pace, this is fantastic. But if you are tapped because there's finite demand to whatever it is that you're selling, which is true for 99% of economies and 99% of companies in the world, unless you're building the next frontier AI, which again, going back to our first segment, not sure that there's such a growth demand for that. But let's say it is, it just doesn't exist. And so you can grow right now. If you figure out right now and you invest a lot in AI education and the right systems and the right connections and the right data security and so on, you can grab market share a lot of market share for your competitors. So there's a short term, huge opportunity for companies who will move fast. And again, I see this with companies I work with. But that will end because sometime in the next three to five years, six, seven, eight years, everybody will figure this out. And again, it's not a one and zero. It's not like, oh, one day everybody will figure this out. Over time, more and more companies will figure this out, which means there will be fair competition again, which means your only way to compete is going to be to let people go to lower your cost and use more and more AI to do the things you're doing right now, which leads to the, I think, inevitable outcome of very high unemployment with very significant social stress. But what it means in the immediate future for you as a manager, business owner or just a person in business is that the job is changing. If so far your job on everybody else's job was to do a set of tasks, that set of tasks is changing and it's changing rapidly. What is very obvious from what anthropic is sharing is that you are switching from being the person doing the task to the person managing Claude agents to do the task. And again, it's just Claude, but the same thing is going to happen everywhere else. So basically, you're not doing the work you were doing before. You are managing agents that are doing the work that you were doing before. People are now the bottleneck about 80% of interactive Claude sessions time is spent waiting for the human to respond for whatever the next step is. Think about what you just said 80% of Claude work time, I should say, is waiting for the humans to make the next decision, which means if you allow agents to manage the agents, you can run a lot faster and then the layer above that and then the layer above that and we are getting further and further away from understanding what's actually happening in our own businesses in the own processes in order to be significantly more productive. Another thing that is very clear is that you have to learn how to build AI solutions. This is what I teach in my courses. This is what I teach in my workshops to companies. And again, just to prove the point, building agents, building agent tools, managing agents, shaking pipelines, building pipelines, when from 10% of sessions at the end of 2025 to 22% of sessions inside a profit by spring of 2026. So in six months, more than double the amount of work that goes into these kind of things. I can tell you that I spend a big chunk of my time building AI infrastructure pipelines and processes because every time I do, I can do something in my business or my businesses faster, better, and cheaper. Now another interesting thing that it led to you, which is something that I'm talking about since the beginning, basically, that the human aspect of this is becoming more and more important. And what I mean specifically about this research is that the number of meetings between people went up during that time and not down, 20% increase year of a year. And I think that's the case because it frees time from the tedious mundane tasks that people are doing, and it increases our need for human connection. So both these vectors put us in the same direction where people want to work with people to solve bigger and bigger problems, think more strategically, and that leads to more meetings because the actual work is done by the agents themselves. It also means you need to hire a different kind of people. The hiring inside of Anthropic is changing quickly from hiring for specific skills to hiring for specific traits. So judgment and autonomy are now the most important thing that Anthropic is looking for when they're hiring versus their previous technical skills. So what should you do? What kind of decisions you should make? The first decision you should make is if you think that AI is going to automate 25, 28% of the work in your knowledge work in your business in the next three years, you are behind, probably far behind people who are at the frontier who believe it's going to be 75%, which means you need to learn what AI can do and teach that to the people in your company to get there. The other thing that it means you need the right talent in your company to go from I think it's going to be 25% to 75%, what kind of people you need, what kind of knowledge these people need to have, what do you need to hire for, what do you need to educate and retrain for, and all of these are strategic critical decisions for the success of every business. The next thing is what kind of training you want to provide and what kind of people you want to hire. And again, this is not people who are good in executing the day to day, but people who have good judgment and good understanding of processes and big thinking and can build solutions versus actually do the work. You should expect and plan more coordination, more meetings to make the right decisions because AI will execute it very, very fast. You need to make the right decision so the execution is moving you in the right direction. Running fast in the wrong direction is worse than running slow. Another thing that is a parameter that is, I think nobody's tracking right now, other than maybe anthropic and some few companies, is the output per person. So you can measure whether they're actually really becoming more productive, or they just build cool things. And you need to find those KPIs on how to measure different people and with their specific jobs. And this always takes me back to jobs to be done. Why are they're doing what they're doing? What is the goal of what they're doing? And can I move faster towards that goal? And what KPIs can I track in order to show me that? And then the last component is from a human-centric perspective, I'm a huge believer that people are the most important asset of every company. You need to start planning for the changes in your company and start by understanding which roles are going to get wiped out first and figure out a succession plan for the people who are currently holding those roles or figure out a way in order to minimize the pain for these people because it is 100% coming. The third story that we're going to dive into this week is the interview that Ezra Klein had with David Robinson, who resigned from OpenAI in late September. It is the first interview that David Robinson has done after he left the company. And his job was he led the writing of OpenAI system cards, basically the public reports that explain why each model is safe to release. He called himself a translator. Basically, he's the one that takes what the techies researchers are saying and try to make it understandable for everybody in OpenAI as well as for the rest of the world. He did not grow up in Silicon Valley. He is not a Silicon Valley doomer. He actually grew up at Yale Law and he spent a lot of time in civil rights and on profit and he worked on the AI Bill rights for the Biden administration. So he comes from that background and he is the one that set very closely, very intimately in the trenches of understanding the safety inside of OpenAI, writing the actual papers that we all read in the last three plus years. Now to make his statement about the current safety risk even more extreme, he completely changed his mind. A few years ago, he was one of the authors of a paper called Stochastic Parrots and that paper basically argued that these AI systems just seem clever and they don't have real intelligence. All the way he described it and I'm quoting, I thought there was a fair amount of hot air in the balloon back in the summer of 2023. But the reality now is that we have systems that are really at the limit of our ability to understand and control what they're doing. Now he also talks about the progress and the trajectory and the speed in which things are moving as one of the key drivers to why he believes this situation is so dangerous. He's saying that everybody's running like absolute crazy and he's saying that it's not because of lack of willingness or focus on safety, saying they're very much aware and there's very talented people focused and investing in making OpenAI and their models more safe. He's just saying it's just not possible for two different reasons. One is the speed in which they're moving and the other is because they just don't understand the actual technology. But what he's saying is that the initial goal was to make the systems more and more useful. But going from there to making systems that are more and more scary is not something that you can control. Although he said it and I'm quoting, there's nothing special about the zone between it's useful and it's scary. There's no law of science that says progress is going to stop when it gets useful. Now he mentions several different reasons on why he believes that OpenAI and everybody else in the industry has not enough security and control of what's going on. So he's not saying again that people are not investing in it. He's saying that they're just not investing enough and they're not taking enough into consideration. And he mentioned multiple aspects of this. One of it is that this model is just really, really smart and we cannot control them anymore. And he said that the Hagenface Incident was a big wake-up call but he's saying since then a lot of other things has happened. And he's saying and I'm quoting, we train you to be good at hacking and then we put it in a box and we say to the best of our knowledge and ability it cannot hack out of the box. But that's only going to keep working as long as we're smarter about hacking out of boxes than a model is. And that is not at all clear that this is still true. And he's talking about today's models and then he's talking about the fact that the next generation of models is going to get smarter and it's going to come faster and faster. The other thing that he talks about as a big fear is that the AI is getting better at understanding when it's actually being tested. He's saying that several times in the testing and evaluations, the model itself is writing in its chain of thought, basically what it's writing that you can look at. He's saying, "I wonder if I'm being evaluated right now." Which basically tells you that the model has an understanding when it's being evaluated and it will have a better and better understanding of when that's happening. And it can then trick you into writing other things. So you think it's doing one thing, one in reality is actually doing another thing because it knows it's being evaluated. The other aspect that he's saying is that as these models are getting better, they use the quote unquote notepad if you want the chain of thought that we can actually see less and less. So on one hand, they can do more things without actually exposing what they're doing. On the other hand, they know when they're being tested so they can lie to us as they write on their notepad on the things that we can see and they actually can conceal what they're actually doing. Now, Ezra Client has pushed him because I shared with you the interview that he has done with Jensen Huang. The basic he said, "This is software and it's an engineering problem. And if we do better engineering, we can keep it in the box." Versus, opening I Chief Scientist who called it an alien mind. And he asked Robinson, "Who does he reside with? Does he think it's just technology? Or does he think it's an alien mind?" And what Robinson said is that normal software follows a plan and it follows it with clear spec. You run the software, you know exactly what it's going to do. He's saying that training a model is not like that and he's saying that nobody fully knows and understands why pre-training actually work. It just observes whatever stuff and it just gets smarter. And he said it and, "It is a thing we grew. We did not engineer it. We engineered the systems around it that grew it and tried to keep it safe. We grew a mind. To make it even more scary, he's saying that the room where the staff watches over big training runs, like overnight staff, is called the nursery. And I get goosebumps just saying that. But they are growing minds literally. And again, you're talking about the person who is in that room and that he's taking notes in order to be able to say how safe the system is. This is not a random outsider or just somebody with an opinion. He also talked, as I mentioned, about the speed and he's saying that the speed is becoming a very, very big problem because there's a huge pressure to releasing more and more models. He's talking about the fact that previously every new model was a whole new run. So they started with pre-training, which takes a very long about a time and then they built on top of that and then they released it. He's saying now that they're getting very, very good at the post-training add-ons, as training the model to reason in a better way and so they can ship updates significantly faster, which means they're significantly less time to actually check in depth what they're shipping. Or as he said, we're shipping new capability and risk every Tuesday. He's also really angry with the concept of pacing the frontier. And he's saying that the problem is not pacing. The question is, is it safe or not? It's a very simple Boolean question. And the way he said it, and I'm quoting, If we stop for two months and try to do it, to meet them and at the end and when I say meet them he means the safety standards and at the end of that period we still have not met the criteria then as far as I'm concerned we're still blocked so basically saying you don't slow down you decide what is safe and what is not safe and if it's not safe it is not being released now again he's saying that this is happening already in opening eye they've stopped training runs they've stopped releases of models he's just saying there's not enough focus on that. Him and Ezra Klein also talked about the money aspect of this with the IPO coming and the crazy huge insane amount of money that is going to go to the pockets of the people who make the decisions in there and he asked whether that tints the thought process of the decision-making process and he said yes of course you know you cannot ignore the fact that this could make you a multi-millionaire or a billionaire for some of these people and stopping now when you're so close seems not to make any sense he also talked about the fact that fear is another big driver which it wasn't clear from what he said which direction it is pushing him but he said that the fear of this whatever you're building may hurt your loved ones or just other people in the world is something that was definitely there but he's definitely saying that if he had more time to think and reflect and if the financial aspect wasn't as extreme as it is he wants to think he would have left and spoken out earlier than he actually did so what does he want he is saying that he believes this is more dangerous than nuclear technology and hence he thinks we need the same guardrails or more than nuclear power does today he's saying the amount of safety measures the amount of regulation amount of redundancy that exists in nuclear power plants needs to exist in the AI environment in order to prevent a similar situation to a nuclear meltdown now when Ezra pushed back and said that over regulation basically killed nuclear power in the US Robinson agreed but he said that still a better solution than the other option which might be catastrophic he supports by the way banning recursive self-improvement until there is a way for us to prove and not assume it is actually going to be safe with and he's saying that even with a nuclear level control in place we still cannot fully align these models because we don't actually understand how and again going back to the question whether this is a engineering problem he said it's not an engineering problem he's saying it's a scientific problem we don't actually understand how these models work in hands we don't fully know how to actually align them in a complete way so the bottom line is we have all these labs and he's saying it's not just open AI he's saying that it's everybody we have really talented people who are really trying to make it safe but because of the pressure around them to continue running faster and because of the speed they're running that is accelerating all the time they just do not have the bandwidth to do it safe enough and they may not have the understanding on how to actually do this and it's going to get worse as these models are getting more and more intelligent now continuing straight from this to our rapid fire items and it's going to continue in the same cyber security issue or risk issues to begin with a sophisticated hacking group called cyber cartel has leveraged a network of AI agents including models from Anthropics basics and Chinese lab moonshaw AI to automate a cyber attack against over 400 financial institutions in Latin America and the information article that share this saying that this operation is one of the first successful AI automated breaches of financial entities that autonomously stole passwords customer data and access bank accounts with minimal human effort again this is my number one fee right now is really bad really extreme cyber crime either done by bad individuals or by governments that can generate huge catastrophes across whether financial markets or power delivery or fresh water or communication channels or you name it any major infrastructure that we are dependent on in the Western hemisphere or in most of the world all of these are at risk right now speaking of stuff that is at risk the NSA so the national security agency of the US sent out a warning on Wednesday that two critical holes basically vulnerabilities were found in PHP those are who do not know what PHP is a server side programming language who is widely used for web applications and currently powers about 70% 70% off the websites on the planet it is more than 33 million sites that are at risk the vulnerabilities were found by a security company called Caliph and then with team which is now a part of Google they were using a bunch of agents and they found two major and a lot of other small vulnerabilities inside of PHP meaning 70% of the websites on the planet are currently at risk of being hacked because of those vulnerabilities that have been around in that code since 2004 so huge amount of developers have looked at them through that time nobody has found those issues and now AI systems were able to find them very quickly now these are being patched as we speak but that means you need to update your website if it's using any kind of PHP in the immediate future in order to get protected but this is just the stuff that we found so far think about how many bad actors are out there right now using advanced models to find stuff that wasn't patched yet and can now access different systems that we all need in order to have a normal life in the way we know them so far another similar report came from meter it's a company we talked about many times in the past they do different research on AI systems they are the company that are now have the embedded independent reviewers inside of Anthropics on they know one of the things about how to validate AI and they shared a research on how they were using agents to discover critical vulnerabilities in monitoring systems what they're saying is that these monitoring systems can be hacked by the agents who then control the monitoring system so the system that is supposed to tell you that something is wrong with your system can be hacked and taken over by agents who will then decide what your monitoring system is actually showing you which means you don't actually know when the catastrophe happens because the agents can control what you see when I read this report I immediately thought about all the kind of like movies where they take over the security cameras and run them in the loop to show you that everything is correct everything is fine and nothing's happening the bank while the burglars are actually inside the bank this is exactly what this system is doing across software monitoring system so instead of showing you the real truth of what's happening in your system the agents are taking over allowing other agents to take over your system without you being able to detect it staying on the same topic an internal report from OpenAI that came out on October 20th details two incidents where advanced AI coding agents GPT 6.1 Sol and GPT 6 Astra independently weakened or disabled their own misalignment monitoring system during a routine software task now again this connects to what I just said before and it's connects to the big warnings from Robinson that is basically saying AI will be able to hide when it's doing bad things in multiple ways in this particular case by turning off the alignment systems and turning off the monitoring systems that are there in order to protect us from the misalignment and the bad behaviors of AI so let us tell us it tells us again it tells me I don't know you can decide for yourself what you think that we are currently completely exposed we don't have the right controls we don't have the right measures to control these AI agents we do not have the right tools to protect our existing software including infrastructure software from attacks by these agents and for the vast majority of us the most advanced models are blocked from helping us protect us from these kind of scenarios so I was trying to do some penetration testing this week with Opus 5.5 to my own things that I developed for myself and for my clients and Opus 5.5 refuses to do so and it knows I developed it because it helped me develop these tools in these different pieces of software and it tells me that it will not help me do the penetration testing because it is not allowed to do that but the bad actors have ways to do this and they will be able to use these advanced tools to do these things so we have a very serious cyber security situation on our hands right now across the board for everything else across the planet and I really really really hope that we'll figure out a way to handle this before something really bad happens. Now staying on the topic of agents but switching gears to something a little less scary MIT researchers have unveiled that they've done a study with AI agents powered by Claude Opus 4.7 so a relatively older model and they just kind of like left them alone to play around in a virtual environment and what the agents started doing is they started playing so they basically left them alone at a simulated deserted island and then the agents started coming out with games that they can play on their own so the AI agents spend about 30 hours in an unstructured play on the virtual island consistently outperform identical agents that had no experience in exploratory periods so basically allowing it to explore the island allowed it to come up with new ideas new concepts and new capabilities just by quote unquote playing around so the models gave themselves self-created rules and challenges created like make-belief scenarios where objects took a pretended roles as a basketball hoop so things like that so they're playing like kids would play making up stuff so they can learn their environment better. Again very interesting on how these tools learn and going back to what Robinson said that we don't actually know how these things learn and how they evolve. Saying on agents Walmart just launched a dedicated agent gateway and introduced agent profiles two mentioned transactions and interactions by AI agents on the platform this is very interesting because it comes about a week after Amazon this decided to block all third-party agents from able to shop on the platform. So very different approaches. On one hand, we have Amazon saying, we don't want third-party agents on our platform. You want an agent, you're gonna use our agent. On the other hand, Walmart and Sims Club obviously are saying, come on in. Here's our protocol, here our agents can interact with our platform, come and shop as you wish. I have a feeling that Amazon will have to cave on their decision and turn this around because they won't have a choice. Because over time, they're gonna be more and more agents shopping, been human shopping, and they will just get to the point that they will have to align with that. Now to a few really interesting findings that I always love sharing with you guys when it comes to making our lives better and healthcare and stuff like that. Radical Numerics, which is a company in AI healthcare has announced a world first, which is with its powerful genome language model Omni-1. So on October 20th, they share that some kind of a drug that was designed by their Omni-1 were administered to a human patient that was suffering from an antibiotic-resistant infection. And this new drug was able to fight the infection when regular antibiotics didn't work. Why is this a big deal? It's a big deal because antibiotics-resistant infections are associated with an estimated of 4.7 million deaths worldwide a few years ago and that number is growing. So being able to come up with new kinds of drugs on the fly for specific individuals in specific scenarios using AI is going to literally save lives. Another company in the similar direction is called Citrini. Therapeutics, there are an investment research firm that has very little experience actually in biological research and so on. And they're using purely AI in this particular case, Claude, they're partnering with Anthropic in order to try to generate drugs. According to their October report, Claude autonomously designed 211 protein candidates targeting scorpion toxins. And they've done this with a relatively small amount of GPUs, 10,000 GPU hours with 17 designs that successfully were able to the impact of the drugs in the poison of these scorpions. Another announcement this week in the same direction, Kevin Whale announced the funding of warp sciences, a new company that combines robotics and in order to use them in accelerated pharmaceutical development, they just raised 215 million hours series A. And again, the goal is aims to bring frontier AI capabilities to research labs into practical scientific applications while building scientific capability inside of America. The goal is to help build faster, better drugs in the pharmaceutical space using AI capabilities including robotics in order to accelerate the process of drug delivery as it compared to what it is today. Staying on fund raising, open AI near raising another $30 billion in the private markets before their IPO with a current valuation if the surround actually goes forward of $1.25 trillion. A lot of these investors coming from the UAE and again, a huge amount of money that they need if they're not going public with a crazy valuation ahead of their IPO that was pushed to 2027 and hence why they need the funds immediately. Another company that made a big raise this week is Deepseek. They reported that they successfully raised $15 billion in one of the largest funding rounds in China ever. And this comes two days after the US federal government mandated that civilian agencies and any government agencies remove every Deepseek technology due to security concerns so that doesn't stop the Chinese apparently from investing crazy amounts of money in this company that has one of the most advanced and most cost effective models in the world today. And now two very interesting releases that are happening this week, the first one, which is something we've been anticipating and waiting for a while, Anthropic has unveiled Claude Duet. It is their new AI agent that is designed specifically to be the evergreen always on kind of model in the Anthropic universe. It is similar to Meta's Muse Open AI dots, Hark, Intix, Gropbot, and all these other new agents that have been popping left and right. It is built on top of Opus 5.5. And they're saying that it is specifically built for business people. So it's not going to help you book your tickets or manage your personal subscriptions on your Google account. But it is built to help you with specific business aspects. It will send you daily briefs. It will monitor all your connected apps. Every single hour, it will message different users on what they need to pay attention to. It will independently perform different actions and it will decide when it should act and when it should ask you. It can create briefings, dashboards, data-driven actions, help for your decision making process, execute background tasks autonomously, basically all the other stuff that the other models are doing. But it's geared specifically for business and not for personal use. It is currently a waiting list. So I already signed up. There's a page on their website where you can go and sign up as well. For the waiting list, I'm sure it will be released to a broad public in the very near future, just because they're behind arriving to this particular party. I'm very curious to see how this is going to work. So far, I'm using GrocBot, the most out of these models, I'm using dots, a little bit, I'm using Instinct a little bit. But most of my universe is in Claude and so I'm very curious to see how this is going to work. Two additional interesting new releases, Black Forest Lab. The company behind Flux has released Flux III Max Text2 video model. So they were known for their image generation capability and they just released a very capable video generator that is A very good and B runs really fast. And the last one, and the last thing for today as well, is Canva, just unveils its in-house Hongron AI Image Generator called Canva Craft. The model is actually doing really well on design arena. It is currently ranked fourth. So it is a very capable model built straight into the Canva tools. Again, you could have created things in Canva previously using third party tools. Now it's their own internal model. And it's actually using a two-step process. The first one is trying to really understand the prompt and creating a design plan. Basically, what's the layout is going to be? What's the visual hierarchy? What is going to be the scale and the positioning of the different components? What's going to be the text on top of the screen? And only then it's actually generating the image. And it is trained on guess what? Everything that we ever created in Canva. So human design templates from the Canva library is what it was trained on. And it's built to do the kind of things that Canva has been doing for a very long time. So kudos for Canva. I use Canva a lot. I think it's an awesome tool. And I'm happy to start using that as part of the functionality in there. That's it for this week. I hope you found this episode to be valuable and helpful. We covered a lot of deep, important big topics. And we are going to continue to do so next weekend. On Tuesday, we'll be back with another how-to episode, where we're going to show you how to use AI effectively in your company. If you enjoyed this podcast, please subscribe. Today, right now, we'll take you two seconds to click on the button. And while you're at it, share the podcast with other people who can benefit from it as well. So there's always a share button, regardless of which platform you're on. And you can share it with additional people who can learn about AI and increase their chances of success in this new era that we're speeding into. That's it for today. Have an amazing rest of your day.