253 | The AI milestones in 2025 and what it means for businesses, and the world in 2026

0:00Hello, and welcome to a special end-of-year episode of the Leveraging AI podcast. This week, we are not going to talk about the weekly news. We are going to talk about the summary of 2025 and my projections, or if you want, my thoughts about what's coming in 2026 based on the trends that we have seen evolving in 2025. In addition, I'm not going to do a news episode at the end of next week. I'm taking a week off with my family and spending some quality time together. There will be a Tuesday episode in between with a how-to episode as we release every single Tuesday, and we'll be back with a Tuesday episode in the beginning of 2026, and with a weekend news episode at the end of the first week of 2026.

0:45Today's episode is brought to you by the AI Business Transformation course, which is the course that I've been personally teaching since April of 2023. I've been teaching this course at least once a month, in many months, more than once, because I've done a lot of private versions of this course. So if you have not yet taken structured training about AI beyond listening to podcasts and following people on YouTube. It is definitely the time to do that. In the beginning of 2026, we are launching another cohort that starts on the third week of January and goes for four weeks in a row, two hours every single week, plus a one hour ask me anything sessions on Fridays.

1:24So if you, again, haven't taken any such training, this is a great time to come and join us and dramatically accelerate your AI knowledge when it comes to how to use it in a business context. Thousands of individuals and business people and business leaders from all around the world have taken this course in the past two and a half years and dramatically and have transformed their personal careers and or their businesses using the knowledge they acquired during that course. If you are a business leader and you're looking for structured training for your entire team or your entire company, please reach out to me through my email or LinkedIn.

1:57There are links to both of them in the show notes. Most of what I do is private training for organizations through different kinds of workshops and courses and so on. So reach out to me for that. We're also launching the second cohort of our more advanced introduction to workflow automation with AI in which we build on top of what people learn in the first course and teach you how to apply this and connect this to your entire tech stack and build actual automations that can do stuff in your business. This will start immediately after the first course ends. So you can take both of them back to back, or if you have the basic knowledge, you can join just the second course.

2:32And now let's dive into the summary of what happened in 2025 for a short little while, what is probably going to happen in 2026, or at least the trends that I can see right now, and then a little bit on how you can prepare for what's coming. So just like the previous decade has laid the foundation that made Generative AI an option for us, such as the transformer architecture, such as distributed into huge data centers, such as fast internet, and so on. All these things played a role in enabling the AI we know today in the same exact way, the things that happened in 2025 are laying the foundations for what's most likely coming in 2026.

3:12So the first thing that I'm going to talk about in 2025 is reasoning models. While they were technically introduced in Q4 of 2025, with introduction of ChatGPT 01, they became a lot more available in the middle of the year when ChatGPT released GPT-5 and baked the reasoning model into the regular model that everybody uses. Because before that, based on OpenAI themselves, less than 10 % of the population actually used their reasoning models. So if you've been listening to this podcast and you have been following me and others talking about this, you might have been using reasoning models at the end of 24 and the beginning of 25, but the vast majority of the population did not use any of them before the models themselves started picking the reasoning capability as part of their workflow.

3:56And so the jump in usage of reasoning models and the jump in the understanding of the labs on how to use them more effectively has jumped dramatically in the second half of 2025. And that's going to play a very big role into 2026. Another big aspect that happened in 2025 is the rise of China when it comes to AI capabilities. For now, mostly software, but hardware is on the rise as well. This started in the beginning of 2025 with a deep seek moment when suddenly there was a model from China that was at par-ish with the Western leading models and really placed China on the map. Since then, we have multiple really powerful, highly capable models from China that are not only really good, they're also in most cases much cheaper than the Western models.

4:45The two models on the Western side that actually broke that equation to an extent is the most recent Gemini 3 Flash and the one before that, Gemini 2.5 Flash. Both of them are extremely capable models. Gemini 3 Flash, more or less as good as Gemini 3 Pro and better or at least equal or at least somewhat equal to the leading models out there, but for a fraction of the cost. But there are many more like that from the Chinese side of the world, which has led to a global competition between the US and China when it comes to global domination in the world of AI. More on that later. The third big component that I want to talk about when it comes to trends in 2025 is agents.

5:26So while the conversation in 2024 was about better and better models, the conversation about agents intensified and became a lot more real in 2025. And in addition to the fact, more and more tools and more and more platforms enable building, developing, and deploying agents. And beyond the fact that many more companies, most of them in the enterprise level, started testing and some deploying agents, three more big things happen that will enable the most likely agent explosion we are going to see in 2026. One of them is the introduction of MCP. We've talked about MCP multiple times in this show, and we even showed a few use cases on how how to apply it, but MCP stands for model context protocol, and it was invented by Anthropic, but then open sourced so everybody can use it, and it was widely adopted.

6:15What it enables to do is it enables to develop an interface to an existing data set and or an existing tool that already exists and connect it then to any AI tool seamlessly in a few lines of code that you can copy and paste, which means you can connect to your ERP, to your CRM, to your email platform, to your marketing platform, to any other tool that you want, and to relevant data sets just by developing the MCP once. And then that connector to that particular tool, let's say Salesforce, is now available to everybody to use across any AI tool in a matter of seconds. So this is a huge deal because it saves a crazy amount of work across the board to many, many different companies, which drove adoption for agents and their connectivity to existing tech stacks.

7:00The other development that happened that will enable a lot of growth in 2026 is what's called A2A or agent-to-agent protocol, which was also open sourced. And it is a protocol that defines how one agent effectively talks to another agent, regardless of who created them on what platform, which tools, etc. So it standardizes the communication between agents, which then enables a lot of the stuff that we're going to see in the future. More about that shortly. Another huge deal that happened late in 2025 is the ability to go beyond the current context window limitations. So context window is the amount of data you can put in a single chat.

7:39And it has grown dramatically in the past few years. And the leading models right now have usually between a million to two million tokens context window. What is a token? It's the way these models work. They don't actually know words. They know tokens, which is about 0.7 or 0.75 words, which means in the most advanced tools today, the context window allows you to bring in between 750 ,000 words to about 1.5 million words. That is a lot, but it is not enough to bring in huge context, such as your entire database or your entire code base or the entire data set of everything in your company. However, both OpenAI and Anthropic has launched two capabilities right towards the end of the year, OpenAI with what they call context compaction that came out together with GPT 5.2 just a couple of weeks ago.

8:27And what it enables the AI to do is to effectively compact the context from a previous chat and push it to the next chat so the AI can continue working on a much longer context without breaking the knowledge that it had in the previous conversation. Once this process is perfected, It basically allows it to run indefinitely while rolling forward information from one context window to the next and working in huge amounts of data. Right now, it is geared specifically for coding and looking at huge code bases, but this will change and will most likely be available to any kind of long context, long duration tasks or entire projects that individuals and companies will want to do.

9:09A few weeks before that, in the launch of Opus 4.5, Anthropic released what they called agent harness as part of their multi-session SDK that practically does the same thing. So it works in slightly different ways. In this particular case, the way it's working is that there are two agents. One is an initializer agent and the other is the actual coding agent. And the initializer agent sets up the persistent environment and logs all the different actions and steps that are being taken so the coding agent can continue independently regardless of the actual context windows and how many times it was changed in between the different steps.

9:46Regardless of the way it actually works, the outcome is the same way. We now have practically the ability to run extremely long sessions with keeping the context so the AI agent can know what happened over a very long period of time. Again, in both cases, it was developed for coding. in both cases, it will most likely go beyond agents and transition to any other type of knowledge work. The final big change that happened towards the end of the year that will also enable these systems to be significantly wider spread is a research from the Cognizant AI Lab and University of Texas, Austin, that published a paper called Solving a Million-Step LLM Task with Zero Errors.

10:30And they developed a concept they called MAKER, which stands for Maximal Agentic decomposition, K threshold error mitigation, and red flagging. That's a really long mouthful, but what it actually does is it takes a task and then the AI breaks it into very, very, very small segments of the tasks. But then it lets several different agents solve the same task in parallel, and then it compares the results. The process then uses basically a voting scheme where the agents vote on the most likely correct answer based on statistical parameters. So it weeds out the mistakes that specific agent makes because it compares it to the results of other agents.

11:12So if you have five agents are solving a problem, four get to the same conclusion, and one gets to a different conclusion, it will continue with the four agents. Now, because they're breaking it into really, really small tasks, and because every task is evaluated separately, they were able to run a process that has over a million steps and run it all the way through the end with zero errors. No hallucinations, no mistakes, no formatting problems, et cetera. Now to make it efficient, they're using GPT-4-1 Mini in their test, which is a cheap and fast model. They could have now probably used Gemini 3 Flash, but any of these models that are smaller and fast and efficient will do the task well.

11:51And because they're much cheaper, you can run all these parallel agents and still make it cost-effective, especially that you're guaranteeing correct and accurate results over a very long, complex task. So what does all of these components together give us? They give us the capability to now run tasks and entire projects that could be hours or potentially days still in 2026 while delivering accurate results. Again, right now, many of these tools are built around writing code, but this will go way beyond that into marketing and sales and strategy and operations and so on. meaning we will have agents that will be able to act over much more complex environment and much more complex tasks and projects independently of humans.

12:37Now, to be fair, we are not 100 % there yet, especially not with the regular setup that people have. And so one of the things that I'm sure will start happening in 2026, we will start seeing companies higher for people that will evaluate the outputs of AI, or they will define specific time windows or part of job descriptions of other people that are currently working in the company to evaluate the output of AI in their area of expertise. Now, this may sound counterintuitive. We're adding work because of AI. No, the reality is you're not adding work. You are saving five hours of work that the AI is doing for you and then investing one hour in evaluating the work.

13:15So you're still saving four hours and getting to better results and faster. Another big thing that happened in 2025 that will connect to all of this is real-time context awareness. And this comes in several different shapes and forms. The first one is connectors to your entire tech stack. So even out of the box, inside your Chachapiti, Claude, Gemini, etc., there are connectors to many tools that you use in the day-to-day, but there are many other connectors not straight in these tools that can go far beyond that and connecting that with MCP solutions from different providers tells you that the AI can now get access to your real work data regardless of where it is stored.

13:54That's obviously not across the board. There's still a lot of issues. The data is siloed in different places and different schemes and all databases. I'm not saying it's solved. I'm saying we've made a huge step forward in 2025 to giving access to our real business data to these AI tools. The second thing that improved dramatically is live mode, the ability of AI to have a live conversation with you based on a video feed and or viewing your screen, which is now available built in to most of these tools. I use it in both ChatGPT and Gemini all the time across the board from fixing stuff at my house to solving complex automation problems that I'm building on NA10 or other automation tools.

14:35And these tools are extremely helpful in doing that, and they will keep on getting better, but they fall under the category of real-time context awareness. The third component is agentic browsers, which are becoming a big deal and now are a part of every offering of all the major companies. So you can click the Gemini button on the top right of Chrome. There is now a Cloud Chrome extension that I've tested in the past couple of weeks, and there are several agentic browsers with the leading ones being Comet from Perplexity and Atlas from ChatGPT. All of these allow AI agents to view your screen in real time, take over your screen, and basically perform most tasks, which you can perform in the browser, which means most tasks you can perform in the digital world.

15:18All of these come with serious warnings of security and prompt injection and so on. So be aware of that if you're planning to use them excessively. But going back to the point of real-time context awareness, these tools are a very big deal and I'm sure we'll see more and more of them while solving the security problems that they bring right now. The next component is true multi. The models in 2025 are real multi-modal tools, meaning they can generate, see, and understand video, images, text, etc. all in a single model, which gives them a much better understanding of our real world and how it works.

15:57And the next step of that, which is going to be a huge trend in 2026, is world models. So one person who's been talking about world models for a very long time is Yan Le Koon, who recently left Meta in order to start a company that focuses exactly on that. Another person that's been a big supporter of this is Fei Fei Li, one of the founding fathers of the modern AI. But in the recent and fascinating interview with Demis Asabes from Google DeepMind's podcast themselves, he talks about that the next breakthrough will most likely come by combining world models with the models that we have right now.

16:32And the last person that talked about learning in real time from the actual world is Ilya Saskover, who is one of the co-founders of OpenAI and currently the founder of SSI, Self Safe Super Intelligence, who also talked about models that can learn over time from the world and from everything that they're experiencing. So all of these people are talking about world models. What are world models? They're models that can see and experience the world and not just live inside the box of a computer. They have access to cameras. They start to understand physics. They start to understand the real world around us.

17:06And they learn like babies and kids learn by experiencing the world around them. This will allow the models to be a lot more grounded with the realities we live in. And it will be a much more solid baseline for robotics, which we're going to talk about shortly. Another aspect that I just mentioned is continuous learners. Both Demis Asabes and Ilya Saskover and many others are talking about models that will continuously learn on their own. So instead of doing it the way it's done right now, meaning let's collect all the knowledge in the world and run a training run, which will probably continue to exist as the baseline, you develop a model that is really good at learning.

17:43And then it is learning on its own continuously, meaning you don't need to do additional training runs as the model learns as it goes forward, which dramatically accelerates the learning. Combine that with what we talked about last week that became a big deal in the past few weeks, which is recursive self learning AI systems, meaning AI systems. And as we spoke last week, hardware as well, that improves over time on its own in a closed loop without the input from humans. And you understand how AI models and AI hardware can grow faster and faster while being continuous learners and understanding the world better.

18:19And you can develop much more advanced tools, which are the path based on all these experts to AGI and beyond. Now, from a practical perspective, what we have seen is we have seen all of these AI capabilities being more and more integrated into the tools and the hardware that we use every single day. So now you have AI capabilities inside of Microsoft Office, inside of Google Workspace, inside of Salesforce, Notion, etc. More and more tools in your tech stack have AI built into them. And it's getting integrated between these tools as well through MCPs and other connectors. But it's not just software.

18:55We have seen AI being integrated into hardware as well. And that will continue moving forward. So right now, as an example, I'm driving a Tesla. I have Grok inside my Tesla. Initially, it was just a fun way to have a conversation while you're driving and learn what's the weather going to be at the destination. What were the results or asking trivia questions with my kids. But right now, it is starting to get integrated into the car system, such as navigation and so on. And very shortly, Grok will be your interface into the car. And I'm sure Tesla is not the last company to do this. We have Microsoft Copilot plus PCs, in which AI is built into the actual hardware of the device.

19:31You have Pixel 10 Pro from Google, that is a device that has many AI functionality built into the device. And there is the so far non-existent, but hopefully will change in 2026. Apple Intelligence, which will enable AI to run on the device. Connecting the dots for a second, think about AI like Gemini 3 Flash, which requires significantly less resources and still is a very capable model. That trend will keep on happening and will enable running high-end advanced AI on device, which provides a lot more security, safety, and speed, which is suitable for many more use cases. Going the next step from just devices that we have today and combining them with AI, robotics has made a huge jump in 2025.

20:13We now have multiple companies in the humanoid and quasi-humanoid robotics race that are developing amazing robots and starting to deploy them at growing quantities. And the prices are dropping as well. So while the top models still cost hundreds of thousands of dollars per model, many of them are now in tens of thousands of dollars and smaller variations of them are now in single digit thousands of dollars. So as an example, Unitary, which is a Chinese company, which was one of the most advanced humanoid robots out there, started selling G1, which was their first small robot. It's only four feet tall, but it can still do a lot of stuff around the house or in the business for$16 ,000 to$40 ,000, depending on the variation.

20:53And now they have an even cheaper version that's called R1 that they're selling for$6 ,000. Now, can it do everything the big models can do? Absolutely not. But it is showing you the trend of highly capable robots that can do more and more things that humans could only do before. And obviously beyond that, because they're stronger and more capable across multiple different things, such as accuracy and vision and wavelength that they can see and the amount of capacity that they have and the amount of hours, obviously they can work. And you understand combined with the dropping prices that we're going to start seeing robots more or less everywhere, starting with factories, then going into service providers across different things, then in coffee shops and restaurants, and shortly after that in our homes as well.

21:36There's obviously currently safety concerns and other issues, but the trajectory is very clear. And the last two components that we have seen as trends in 25 that will continue into 2026 is model personalization and skills. So let's start with model personalization. The first company that provided us a glimpse into that is ChatGPT with memory, but then with the ability to create your own custom instructions to the model. And now it is becoming even more and more customizable as the year continued. The same thing is now happening in the other models as well. So once long-term memory was introduced, these models become more and more custom to who you are.

22:14They understand your universe. They understand your context. They understand your company. They understand your family. They understand your needs. And they provide more personalized answers. The same thing will happen with the personality of the models that, again, already exist in ChatGPT on a basic level and will probably continue happening into 2026. What does that mean? It means that if you and your coworker ask the same question from the model that you're using, you're going to get different answers because you provided it different context over time and it knows how you like to receive your information, what kind of data you're looking for, how to connect the dots for you in a way that is helpful for you, which is different than somebody else.

22:51This has huge benefits from the value that AI provides and huge risks from my society perspective when there is no unified truth or processes for doing things and every single individual is seeing their own version of the truth and the process. So there's good and bad in that, like many other aspects of AI. And the last component before we dive into what would be the impact in 2026 is scales. Scales were introduced by Anthropic and they just in the last 10 days, open-sourced skills as well, just like they did with MCP. I think skills are incredible. So skills are the ability to teach the AI how to do something specific, a skill, and package it as a package that the AI knows how to pull only when it needs it.

23:33So it doesn't consume your context window and the AI can still use it as needed when it needs it. It can connect to specific tools, it can have specific data, and so on, and it will perform and enhance what AI can do across the board. And I anticipate now with them open sourcing it to see more and more companies starting to use scales. And that will become a second nature and almost an afterthought, probably by mid or the end of next year, where these will just happen in the background and we won't even think about it. Now I'm going to connect some of these dots to some key aspects that Demis Asabes, the CEO of DeepMind from Google, has shared in the recent interview he just did as part of the DeepMind podcast.

24:14So he mentioned several different things that we need to address in order to go to the next level with AI and that he's expecting in 2026. One is solving what he's calling the jagged intelligence. And what he means by jagged intelligence is the fact that AI can show PhD level of knowledge in specific things, but fail miserably in some other stuff that a 12-year-old can perform. So that inconsistency in its ability to understand, analyze, or deliver data is something that has to be resolved. The broader aspect of this is what he calls the shift to reliable agentic systems, meaning agents that can consistently deliver correct results.

24:55And he's talking about the fact that the balance will shift from raw power to consistency and accuracy, meaning instead of investing more in growing bigger and better models, which are already really, really good, the focus will become on forget about making the model better. Just give me reliability. Give me 100 % consistency on the tasks that I'm doing today. Forget about what it can do three years from now. I don't care. Just let me solve the current immediate problems. And hearing from Demis tells you that's the focus in DeepMind, which tells you that that's most likely the focus in the other major labs as well.

25:28So where is all of this going to lead us in 2026? The first thing is agentic explosion. we are going to see agents more and more everywhere. First of all, in large enterprises for internal processes, but over time in 2026, it will grow into more and more complex tasks because of all the capabilities we talked about before, much longer running agents that can take much longer tasks and work for, again, hours and potentially complete days on their own on specific tasks. We will see either no or much lower level of errors coming out of these models, significantly less hallucinations, either because the models themselves will learn to self-correct or with mechanisms such as I mentioned earlier from the research that was just published.

26:16And from a concept, it means that we're going to go from a tool that can get some data in and provide some answers out or generate images or generate anything else to true team members, actual AI collaborators that will work together with humans as part of a team inside companies and organizations to achieve the tasks and the goals of the organization. Now, this requires a complete mindset shift and a complete technological shift and a complete workforce strategy shift from what we know today. Because in the beginning, people will be working with one or two agents, but over time, people will manage multiple agents just like they manage employees right now.