245 | AI Tsunami warning! 11 % of jobs replaceable today, CEOs admit 20 % staff over-capacity, Claude 4.5 outcodes humans, Harvard’s PopEVE cracks rare genes and more AI news for the week ending on Nov 28, 2025

26:40And while most of the research that I shared with you today is optimistic about the potential impact of this on the world economy. I don't really see how that's possible. Again, because I don't see that there's infinite demand for everything. This just cannot happen. So while the MIT research mentions 1.2 trillion in wages that can be done by AI and the McKinsey research talks about unlocking trillions in economic value, to gain those trillions, we actually need people buying those goods and services. And if people don't have money, then who's going to buy these goods and services. And if they cannot, then there is no growth in trillions.

27:18So I don't see how this can play out the way these research are showing. Again, I'm not taking anything away from McKinsey or MIT or PwC. They all have way more people and way smarter people than I am. I just think they're looking at this from a pure mathematical perspective and not from an economical viability perspective. And I haven't yet seen one report that talks about that aspect of it. So either I'm delusional, or these reports are very optimistic when it comes to what can be, at least from a short-term perspective, the impact of figuring out how to apply AI in a collaborative way or however else they want to describe it.

27:55But there's one thing that is clear from all of these reports, which is the big lack of understanding of how to apply AI effectively, both inside of these companies to define strategy, as well as the huge need for these skills when these companies are hiring. And it all comes down to training, education, and AI fluency. Inside the companies, it has to do with training people and leadership in order to figure this out. And outside of the companies, it comes to the individuals, you, to train yourself on how to use these tools effectively in order to get hired and get paid more money by these companies who are struggling to find people with better AI skills.

28:33So if you want some specific quotes from them in the McKinsey Global Institute research, they said, and I'm quoting, the outcomes for firms, workers, and communities will ultimately depend on how organizations and institutions work together to prepare people for the job of the future. Another quote from the same article, integrating AI will not be a simple technology rollout by reimagining of work itself, redesigning processes, roles, skills, culture, and metrics so people, agents, and robots create more value together. Again, it's all about redesigning role skills culture, which all comes back to training.

29:10Another quote says, demand for AI fluency, the ability to use and manage AI tools has grown sevenfold in two years faster than any other skill in US job posting. In the Bearing Point study, the quote is workforce planning, talent development, and organizational design will need to be rethought from the ground. You get the point. It's all focused about understanding AI deeper, finding talent that currently knows AI in order to reimagine the work, in order to gain the benefits from AI. This is what we have been focused on for the last two and a half years with company tailored workshops and open to the public courses.

29:49So if you are in a leadership position in a organization and you still do not have a plan on how to figure out AI for your company, for your industry, for your department, for your team, whatever the case may be, reach out to me. There's a link in the show notes to set up a meeting with me on my calendar so you can discuss your current needs and I can see if I can help you or at least guide you in the right direction. And if you are an individual and you understand the need that is coming in the market and you want to build your personal future, come and join our courses. We have the next cohort of the AI Business Transformation Course, which I have been teaching for over two and a half years now with thousands of business people who have transformed their personal well-being as well as the success of their teams and the people that they manage.

30:33The next cohort of the AI Business Transformation Course is opening on January 20th of 2026. It is literally the best way for you to take the first steps and accelerate your understanding of AI across multiple aspects of AI related specifically to aspects of business. So this is not a theoretical course. We're not going to talk about concepts of AI, but all of it is around practical use cases across multiple aspects of the business. And if you are looking for ways to improve your career success, this is an amazing way to do that. And there is no better time than the beginning of 2026 to make that next step.

31:11So if you want to see more details and join us in January. There's a link for that in the show notes as well. But since we mentioned Anthropic and the research, let's continue with Anthropic. Anthropic just released Claude Opus 4.5. Those of you who need some background, all the models that Claude released have three different variations with Haiku, Sonnet, and Opus. A few weeks ago, Anthropic released Sonnet 4.5, but they did not release Opus 4.5. Sonnet 4.5 took over the entire leaderboard, more or less for every single aspect of AI usage. But since then, several models came out, including Gemini 3 and Grok 4.1.

31:49Both of them took leading positions over Claude. And so now 4.5 Opus, which is the largest model, is trying to reclaim the throne. And to be fair, not completely successfully yet. That being said, it is definitely a better model than Sonnet 4.5. And as an example, they've let Opus 4.5 take a take-home exam that is designed for prospective performance engineering candidates, and Claude Opus 4.5 scored higher than any human candidate ever scored on that test within the two-hour time limit. This model outperforms Sonnet 4.5 on seven out of eight programming languages on different benchmarks, and it comes at a new pricing scheme that makes it cheaper than the previous Opus model with$5 per million input tokens and$25 per million output tokens.

32:39It is also built to run a lot more efficiently. Anthropic also added a new feature into the API calls for this model called effort parameter. This effort parameter allows to control how much Claude actually thinks in order to deliver different tasks. And what they're saying is that in the medium effort settings, the new Claude Opus 4.5 does better than Sonnet 4.5 using 76 % fewer output tokens. So it's not only that they made the tokens cheaper, they can achieve the same results by using significantly less tokens, which overall would make the effort significantly cheaper than previously possible.

33:16Now, while coding is a big deal because it drives a huge percentage of the revenue for Claude, this model is also supposed to be better than the previous model in vision, reasoning, mathematical skills, and overall outperform the Sonnet 4.5 across multiple benchmarks. Anthropic being anthropic and focusing on safety, they're saying it's that their most robustly aligned model yet with lower chances of prompt injection attacks can be applied to this model. So how is this model doing on the LM Arena board? Well, it is currently ranked number three on text after Gemini 3 Pro and Grok 4.1 Thinking.

33:51It is ranked number one for web development. And if you look at the overall table, it is ranked third after Gemini 3 Pro and Grok 4.1 Thinking, where it is sharing the first position across several different aspects and ranked lower on several other aspects. Overall, still reigning supreme with number one across the board on everything is Gemini 3 Pro. Another very interesting thing that Anthropic has released today is that they have developed a new methodology that allows AI to run effectively across multiple chats. Now, what is this coming to solve? The core challenge is obviously the context window.

34:29So those of you who don't know and have not been listening to this podcast for a while, every chat that you run has a limited memory that can be used in a single chat. It is called the context window. So as you try to run bigger and bigger tasks, you come to the end of the context window and then you have to start from the beginning. Well, what Anthropic has developed right now is a twofold solution with two types of agents. One is called the initializer agent and the other is a coding agent. And what the initializer agent does is it creates a summary basically of every chat in a JSON format that can be delivered to the next chat with clear instructions on how to continue from the previous chat.

35:04In addition, it gives very specific instructions to the coding agent on exactly what feature to work on at every single time out of the bigger tasks, being more like a manager versus just a deliverer of information. And what this new methodology enables is to run a process of developing multiple components in a single shot as the AI on its own will continue across multiple different chats, but continue the development in a structured and well-organized way. Now, this follows a very similar release from OpenAI that I reported last week. So the ability to run across multiple chats while continuing, if you want, the thought process between the different chats.

35:46This is just the beginning. But as this evolves in the next weeks and months, we might get into a situation where the concept of a limited context window will be a thing of the past. meaning AI will be able to complete tasks across multiple hours and maybe days and maybe even weeks if it can keep a coherent flow and continue working without basically the limits of the context window as it can deliver the data from one context window to the other and know exactly what to work on while keeping the process going in an effective way. This is currently built specifically both in OpenAI and in Anthropic for coding.

36:26But as they figure this out for coding, there's zero reason why this will not be effective for more or less everything else. What does that mean? It means a complete game changer to the level of sophistication, complexity, and length of tasks that AI can perform, connecting it back to the previous points of what type of tasks can AI perform today, what it will be able to perform in the very near future is going to be a very, very big gap between these. Because of this limitation that will go away, AI will be able to perform things that are far beyond what it can perform right now, which will have an even bigger impact on how companies can use it and on the overall economy.

37:04Sticking with new releases, but switching to a different company in a different field, Black Forest Labs has released Flux 2. So those of you who don't know, Flux is the best open source image generation model out there. It is extremely capable and it is very good across multiple different functions. Well, Flux 2 was just released and it is almost as good at almost everything as Gemini Dano Banana Pro. So what did they improve from Flux 1? It is significantly better at editing images, including editing high resolution real images and being able to make changes to them while preserving detail and coherence with the original image.

37:41It has multi-reference support, very similar to the new Nano Banana, meaning you can reference up to 10 other images simultaneously and still keep consistency of character, product, style, etc. from all the different images into a new image that you are generating. It is very good at generating complex typography, including generating infographics and UI mock-ups for entire websites, and it is much better at adhering to the structure of prompts and to brand guidelines if these are brought to its attention. I loved using the original version of Flux. I think it's an incredibly powerful model and it has the ability to train the model for your needs.

38:19It's called building a LoRa, which allows you to train the model on a specific style or a specific product or a specific person, which enhances the capabilities even further, which is a functionality that as far as I know doesn't exist in Gemini right now. Now in testing that I've seen different people do online. Right now, there's still an edge for the recent Nano Banana Pro over Flux when it comes to the fine details of the outcome, especially when it comes to multiple aspects of text, like creating an entire dashboard or a infographics. However, it is doing it significantly faster and way cheaper.

38:52So from a cost to quality ratio, Flux is in a much better position than Nano Banana. And in many use cases, it will be good enough. and will be able to deliver the work faster and at a much cheaper rate, which for many of us is the better trade-off versus the slightly higher capability. And it's just gonna depend on your particular use case. So the good news is within two weeks, we got extremely capable image generation and image editing models. One is open source, one from Google, and they are both at the level that enable actual real company work and can dramatically accelerate the professional aspect of image generation.

39:31Most people who use Flux use it through third-party tools and through their API. This is how I use it. However, if you want to just test it out, you get 50 free renderings per day on the Black Forest Lab playground website, which everybody has access to. Staying on the topic of new releases, Harvard Medical School researchers just developed and released an AI model called POP-EVE, or something like that, which stands for Population Calibrated Evolutionary Variant Effect. This is a complete mouthful, but what it does is nothing short of magic. It is designed to replace or address the extremely time-consuming effort of diagnosing rare signal-variant genetic diseases.

40:16When they tested the model, they gave it access to over 30 ,000 patients with undiagnosed severe developmental disorders. The model on its own was able to diagnose one third of the cases, again, that was not diagnosed by people before. Even more impressive is the model was able to identify 123 genes that are linked to developmental disorders that has not been previously known to cause such a disease. Now, since the paper was published, 25 of these genes out of the 123 found was verified and independently confirmed by other research labs. And these are the things that excite me the most about AI is its ability to really address real problems that are currently unaddressable by the technologies we had so far.

40:59being able to help people with really serious diseases by being able to diagnose exactly what's the problem with them at a scale and speed that is unmatched can really help us solve really big problems in the world, whether it's genetic illnesses or global warming or access to clean water, clean energy, et cetera, the biggest problems that the world has today. And AI should be able to, and hopefully will help us solve. Before we switch to some interesting insights from Sam Altman and OpenAI related to the release of Gemini 3. And before we talk about shopping, I want to share with you some insights from the interesting interview of Ilya Saskover, who was one of the co-founders of OpenAI and now is the co-founder of SSI, which stands for Safe Super Intelligence.

41:43He was interviewed by Dwarkesh on the Dwarkesh podcast about the current state of AI and the work that they're doing at SSI. This interview is way more technical than it probably should have been, which some of these interviews are. So I don't necessarily recommend you listening to the entire thing because there's a lot of segments that are highly technical. But that being said, they touched on a few very interesting points. One was the question of how can these models do so good at some of these evals? And then when you actually come to use them, they are either worthless or significantly less capable.

42:14And what Ilya suggested is what I have been claiming on this podcast all along, is that the models are trained to be good at the evals because this is one of the benchmarks that they can be tested on. So the reinforcement learning is rewarding them for being good at the evals. So they become very good at the evals. That doesn't necessarily mean they'll be really good at real life. So think about a high school student that is awesome at solving the test because he was being taught for the entire year on how to solve specific types of questions. That doesn't mean he's going to be really good at the actual application of the topic.

42:48And it is the same exact thing here. These models are rewarded to be good at the evals because this is one of the ways to train them. And so they become very good at that, but not necessarily at the more generalized, less rigid and structured version, which is real life of solving problems in a specific category. So this makes perfect sense to me. And this is what I assumed all along. And this just verifies what I thought about this topic. The other interesting aspect in the interview talked about the transition from the age of scaling to the age of research. What Ilya is saying is that in the early days of AI research, or not very early days, but let's say five to 10 years ago, The main focus was about research because nobody had enough compute and you had to invest a lot in research in order to come up with solutions.

43:30Then we went in the last few years through the age of scaling. All the companies were able to invest billions or hundreds of billions into more and more compute, and they were able to solve a lot of problems by scaling. Well, what he's saying right now is that this age of scaling is getting to the edge of its capacity, which will force companies and organizations and research labs to go back into research and invest in ways to better solve and better teach the models versus just brute force of providing more compute. And he's making a very interesting claim. He's saying that when you think about how humans work, humans are currently way more capable than these machines when it comes to generalized intelligence, despite the fact we have significantly less knowledge.

44:14So the idea behind scaling was let's take more and more and more knowledge in the pre-training phase, and that will generate better and better models, which was conceptually true and was true. And yet right now, these models have significantly more knowledge than humans, and yet they're significantly less capable than us across multiple different aspects, especially when you're generalizing intelligence. Or the way Ilya talks about it, he talks about robustness of human understanding much more deeply and exhibits this ability to analyze and understand things while having a tiny fraction of the data that AI has access to in its training process.

44:52So when Dwarikesh was pushing him to try to explain exactly what they're trying to work on and what super intelligence looks like and how they're going to release that in a safe way to the world, what they ended up with, I actually really like. What he's saying is that what he's trying to build is not necessarily super intelligence, but a super learner AI, basically developing an AI that learns much better than the way current AI learns. And he's saying that will put us on the path for AGI and ASI. And once they figure out how to make the AI a much better learner than it is today, they will start releasing versions of that to the world that will allow people to figure out how to use this in a safe way.

45:30I personally find this very interesting. Again, we are in an era where everything was focused around let's get more and more and more compute. And because it was obvious as we get more compute in a much larger data set, these models become smarter. But it is obvious that this is not the final way to move forward. And I find this exciting because it means that we will need potentially less resources, which would be less demanding on our planet in order to achieve better results with AI. And just the fact that he was able to raise billions to focus purely on research says that there's other people who believe that might be the right path forward instead of just billions investing hundreds of billions in compute.

46:09So overall, I think this is an interesting path, and I really hope this will yield successful results. Now, as I promised the impact of the Gemini 3 release on OpenAI in a candid internal memo that was obtained by the information, Sam Altman has alerted the employees of OpenAI to brace for a period of turbulence, or as he said, and I'm quoting temporary economic headwinds because of the success of Gemini 3. Altman was very open about the success of Google and how well they built the model with the memo stating, and I'm quoting, Google has been doing excellent work recently in every aspect. And it's specifically relating to their advancement in pre-training of models, which connects back to our previous point.

46:55They have found more effective ways to do pre-training, which saves them time and money and yields better results. This drove OpenAI to define a new project, which I'm not sure I'm pronouncing correctly. It's called Shallot Pete or something like that. And that new model or that new project is supposed to address the deficiencies they have in pre-training right now compared to Google with the goal to reclaim their superiority in the AI domain. Now, the report highlights something that I've been talking about on this podcast for a very long time, that it is not a fair fight. Google has a full vertical and horizontal integration in their approach to AI.

47:33They have a full stack of everything. They control their own chips, so they're running on TPUs, which they are developing specifically optimized for what they're trying to do. They have their own data centers. They have their own distribution platforms across the board from Android to Chrome to Google Workspace, et cetera. And OpenAI relies on a lot of other companies to provide most of these capabilities. In addition, from a financial perspective, there is a huge, huge difference with OpenAI making$13 to$20 billion this year. There's different rumors with different numbers, but this is the ballpark,$13 to$20 billion while losing a huge amount of money, burning through a lot of investor cash.

48:12While on the other hand, Google generated$70 billion of free cash flow over this past year. So Google has everything they need in order to keep on pushing in that direction without raising crazy amounts of money. And OpenAI will have to rely on other companies to deliver many of these capabilities, whether building the data centers or creating and delivering GPUs in their perspective, and obviously providing funding for all of this to work, where Google has all of that ready to go, including the ability to distribute that across multiple tools and gain the knowledge from people using it across all these different tools.

48:47So do I think OpenAI can beat Google in this race? I doubt it. But can they stay a significant competitor to Google? 100 % they can, and I'm sure they will stay in there as long as they can manage their cash flows compared to the amount of money they can raise in an effective way. And since we are getting into the most wonderful time of the year, at least from a retail shopping perspective, I want to share with you the two different companies, OpenAI and perplexity, have released shopping research assistance to their tools. So right now, inside of Chachapiti, inside of the selection of what you want to select, such as deep research, you can now select shopping research.

49:24I've actually used it to shop for a TV in this Black Friday week, and I found it to be very, very helpful. It is really good at understanding what your needs are and then curating results based on very complex ideas, concepts, and queries that you want to look for. So you don't necessarily have to look for a specific kind of a TV, but just to find what your needs are, what your budget is, and it will go and help you find the specific thing you are shopping for. It can even help you with ideas. So let's say you're looking for something cute for a five-year-old girl. It can give you ideas of what you can shop for, and then it can shop, and then it can help you find the information and compare different aspects of this to help you make an educated decision on what you are trying to purchase.

50:05I found this to be very helpful, and it saved me hours of doing research on my own. And OpenAI emphasized that the shopping results are organic and unsponsored, meaning they are ranked purely based on relevance to the user's needs and requirements and not based on paid placements like on most platforms out there, whether you're shopping on Google or you're shopping on Amazon, most of the stuff that you're seeing is sponsored. And here, this is not the case. They're trying to show you exactly what you're looking for based on your defined needs, which again, I tested and I found very helpful. A very similar thing was released by Perplexity.

50:41So if you are a heavy Perplexity user or if not an OpenAI user, you can go to the free Perplexity right now and use it to help you in your holiday shopping. Very similar concepts in both cases. You can check out on the platform if it is connected to a checkout process that they already have integrated with. And this gives us a great glimpse of the future that we're walking or running or sprinting into where AI agents are going to help us with more or less everything that we're doing and how we engage with the world and definitely with the digital world. In this particular case, just in shopping, but I think this will evolve over time to more and more aspects of our engagement with our job and our personal lives and how we engage with the world through digital interfaces.

51:21As I mentioned, my very first personal experience was very positive. There are many other aspects of this news this week that we're not going to dive into. Some really fascinating battles that are going on behind the scenes from a chip wars between Google and NVIDIA, new company and investment by Jeff Bezos, and many other good things that there's just not time to cover all of it. And if you want to see all of it, you can sign up for our newsletter. It has all the stuff that we didn't cover today or that we don't cover every single week. We usually include about 30 topics that we cover, maybe a little less, but there's usually about 50 that we curate every single week.

51:58And if you want to know the rest of them, you can just sign up for our newsletter. It also includes all the events that we're running, the training that we're providing, the workshops, the free stuff, everything you want is all in that newsletter. So go sign up for that. And while you're looking for the links for that, you can also look at the links for the workshops and for the courses that are upcoming, which will allow you to prepare yourself and or your company for 2026 and beyond. If you are enjoying this podcast, please hit the subscribe button on your podcast player so you don't miss any episode that we release.

52:30And while you're at it and you have your podcast player open, please rate this podcast on Apple Podcasts or Spotify. This helps us reach more people. And I actually read all of these reviews and get the feedback and try to give you more of the content that you are looking for. So if you have specific things you're looking for, please include that in your review. That is it for today. Have an amazing rest of your weekend and we'll be back with another how-to episode.

52:59you