207 | ChatGPT Agent is now live which changes everything, AGAIN!!! What does this mean for you and your business?

0:00Hello, and welcome to the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage GI to improve efficiency, grow your business, and advance your career. This is Isar Maitis, your host. And I'm actually on vacation right now, and I was not planning to record an episode this week. It is my dad's 80th birthday, and the whole family got together, and the nieces, and nephews, and cousins, and my kids, and everybody. And it's a lot of fun, and it's really, really great. And I'm very excited for him and for everybody else for getting together. But because of that, I was not planning to record an episode today.

0:28And yet, a few really big things have happened and I could not leave you in the dark when really big things are happening. And so I decided to record at least a short episode today. We're going to focus on the really big things and dive into what they mean. And we're going to touch about how the world is turning agentic and what does that mean for you and your business? What does it mean for you and your life? What does it mean for the world? And so there's a lot to talk about, even though we're going to dive into only one or two topics. So let's do this.

1:00It all started early in the week when the Chinese startup Moonshot AI released Kimi K2, which is an open source model with one trillion parameters in the backend and 32 billion active parameters. And it is a really good model. This is an open source model, so it's scoring higher than most or all open source models across multiple benchmarks, and it's also scoring better than some of the leading closed source Western Hemisphere models. In addition, it is an agentic model. And on top of all of that, it is really fast and really cheap. So the trick here is a few things. First of all, it's a mixture of expert style model, which is most of the recent models are, which means it has different areas of the model that are specializing in different things.

1:40And you can train each and every one of them separately. And the agent knows how to call them separately. In addition, they have developed a new way to train the model, and they are calling it Muan Clip Optimizer. Now, this may sound like Chinese, pun intended, but what they're claiming, and I'm quoting now, it enables stable training of a trillion parameter model with zero training instability. What is training instability? Well, training instability has been maybe one of the biggest issues when training large models. And what basically happens is that it creates issues with really large training runs, forcing companies who try to run them to either restart the run or implement very costly safety measures and accept, in many cases, suboptimal performance in order to avoid the model crashing while it's being trained.

2:23Either way, it's a very big tax on training large language models. And with this new methodology, they're able to avoid that altogether, which allowed them to train the model much cheaper than any other model of its size and of its capabilities, which in return makes the model itself much cheaper. So what does much cheaper means? Well, if you use it through the API, it's 15 cents per million input tokens, and it's$2.50 for every million output tokens. If you compare that to GPT 4.1, GPT 4.1 is$2 for every input token, so more than 10x, and$8 for every output token, which is almost 4x. If you compare that to Claude4 Opus, it's$15 for every input token, that's 100x more expensive, and$75 for every output token, which is 30x more expensive.

3:10Now, is it as good as these models? Maybe, maybe not. Either way, it's close and it's much, much cheaper. Now, the way I always go to measure things is how much people are actually using it, especially through the API. So I went and checked one of my favorite API tools that is called Open Router that allows you to actually connect to one API and through that get access to more or less every model on the planet. They take a little bit off the top, but you can do one implementation and then get access to all the different models. I've been using it for a very long time. And over there, there is a dashboard that shows you how much people are using from different models this week.

3:41And on the programming side, Kimi K2 is now number five, meaning more developers are using it through the API this past week than Cloud 3.7 Sonnet, Cloud 4 Opus, Gemini 2.5 Flash Preview, DeepSeek V3, and GPT 4.1. All are very capable models. The only models ahead of it are Cloud Sonnet 4, Gemini 2.5 Pro, Gemini 2.5 Flash, and Grok 4. That's it. Now, does that mean it's better than all these other models? No, it means it's more cost-effective than all these other models, which is what really matters. It means it's good enough for the tasks these developers are using it for at a fraction of the cost, which is an important part of the game when you're building applications around APIs from different large language models.

4:23Now, in addition to all of that, it is an agentic model, meaning it knows how to autonomously use different tools and define its own instructions and pave a path to complete the goal that you set for it, whether it's browsing the internet, writing and executing code, et cetera, et cetera. It's a very capable agentic tool for a relatively small cost. So I was very excited to report about this, but I'm like, you know what? This could wait maybe another week. Nothing will happen. I will report about it a week later. But then towards the end of the week, OpenAI introduced their version of the same thing.

4:54Then something bigger happened. So every time I reported about these generic agents that can do a lot of stuff based on instructions you give them, like GenSpark and Manus. I had several different episodes where I talked about them. Specifically, if you want to check one episode where we talked a lot about these tools and how to run them safely, you should check out episode 196. It was labeled how to safely run powerful AI agents like Manus and GenSpark with no risk. But in that episode and in many other episodes, I said the same thing. I said these tools are incredible. They are the future or the present if you're a geek like me and they are changing everything we know because they're significantly more powerful and capable and with knowing very little, you can generate a lot.

5:30But what I said is a lot of these tools are A, more geeky and B, a little riskier. And all of that is going to change if OpenAI are going to issue their own agentic model that does the same thing because OpenAI has the trust of about 800 million weekly users right now with all the respect to Manus and GenSpark and other tools like this. They probably all combined drive less than 10 % than traffic that OpenAI gets on a single day, maybe less than 5 % of the traffic that OpenAI gets on a single day. And so that event happened on the 16th. On the 16th, OpenAI announced that they're releasing OpenAI Agent, which is their version of a generic agent that can do a lot of things.

6:10And before we dive into what the model can do, let's talk about what it is. What OpenAI did is they took several different capabilities that they've developed before, and they combined them together into a very powerful tool that on its own can decide which of those capabilities to use. It's a combination of deep research, operator, data analysis, and coder all combined into one agentic tool. And what this tool knows how to do is it knows how to research stuff online in its own little browser, which is actually really cool. So the approach they took is instead of having to use your browser, which generates a lot of risks, because as an example, your browser usually has access to all your saved passwords and sometimes credit cards and so on.

6:48Instead of that, it has its own little browser within the OpenAI interface that pops up and then it runs things within that browser. So it can browse the web, it can research the web in similar ways as Deep Research. It can also operate web pages like Operator does, meaning it can click on things, fill out forms, et cetera, that Deep Research does not know how to do. So just this combination on its own is extremely powerful. But in addition to that, it knows how to analyze data. So the data that it brings from these different sources, it can write Python code, put it in spreadsheets and analyze it in multiple ways that gives it an even better, bigger benefit.

7:21And on top of that, it knows how to write code and it has access to terminal and even code execution. So the combination of all these things make it an extremely powerful tool in completing more or less every task that you can imagine, because it can do the research, it can figure out what it needs to do even deeper, it can define its own process, it can write code, it can analyze data, and it can do all these things very, very quickly. A few additional cool things that we've added is the ability to interrupt the model in the middle of work. So you said something and then you watch the model doing its thing and that gives you an idea or just you suddenly think about something you forgot to add, you can add it while the model is working on the thing that it's working on and you will take that into account.

7:57It's very conversational, meaning it will stop and ask you questions if it's not sure about stuff, just like a employee hopefully would. And it can connect to existing data sources that you currently connected your ChatUpt account to, such as Gmail or SharePoint or Google Drive and so one, which makes it even more powerful. Now, it also, because it can write code and because it knows how to analyze data, it can create spreadsheets that can be exported to Excel or Google Sheets, and it knows how to create PowerPoint presentations, including generating the images for the presentations. And so what we are getting is an extremely powerful tool, similar to Manus and GenSpark and these kind of tools, only it's coming from ChatGPT.

8:36ChatGPT comes with a lot more trust with the population. It definitely has a bigger footprint and distribution with the broader population, meaning we're going to have more and more and more people using really advanced agentic tools. Now, the new agentic capability is going to be rolled out to everybody, including Pro, Plus, and Teams users. The way it works is just like you pick all the other modes like Deep Research. There's just agent mode as part of that in the drop-down menu on the left of your prompt box. I still don't have access to it. I have the Plus license, but I assume it's just rolling out.

9:08And just like everything else with ChatGPT, it will take a few days and everybody will get access to it. There's very different limits. You get 400 runs of this tool in the Pro license and you get 40 in the Plus and Teams license. But to be fair, 40 is more than one a day, which for the average user should be way more than enough. And if you need more, that means that paying the$200 a month for the Pro version makes perfect sense to you because you're using this a lot more. Now, what can you do with this? You can do the examples they gave in the actual launch, which I highly recommend watching the video.

9:37We're going to drop a link to that in the show notes. They've shown stuff like doing research for shopping and planning for a wedding. They also shown an example of how to allow the tool to measure itself and bring information about how well it's doing compared to other models. And you can also do a lot of work-related stuff with it, obviously, like market research, scheduling, analyzing resources, deploying different things, preparing for presentations, et cetera, et cetera. There's probably thousands of business use cases where this tool will be extremely helpful. Now, before I tell you what I think about it and where I think this is all going or what's the impact of that, I want to share one more aspect, which is rumors that are coming from several different reliable sources right now, which is OpenAI is planning to take this to the next step and really build around it a tool that has a suite of workspace tool, just like Microsoft 365 or Google's G Suite, which means they are going straight after the main driver of business for two of the most successful software companies ever.

10:33And they're literally going after their bread and butter, the things we use every single day. Now, are they planning to replace the Office Suite or are they planning just to complement it? I'm not exactly sure. I think time will tell. But it is very clear to me that if I had the choice, if there were really solid tools within ChatGPT that are fully integrated with all the other things that ChatGPT does, I will reconsider my usage of Google Suite. Now, will it replace everything? Probably not in the beginning. Can you replace everything over time? Absolutely. can you do it better than what these tools are doing right now?

11:04Right now, it seems that that's the case. Will Microsoft and Google catch up is a very big question. From what it seems right now, Microsoft is doing a pretty poor job of implementing ChatGPT within its workspace as co-pilot. And Google, while they're doing better than co-pilot, they're still not doing great. And I've said that multiple times on this show. I think that both these companies have an incredible opportunity. I thought they will capitalize on this opportunity before the end of 2024. I was obviously wrong, but they need to get their act together and bring together a model that actually looks into everything in their ecosystem.

11:37I don't want Gemini for slides and Gemini for sheets and Gemini for docs and a Gemini for Gmail and the same thing with Copilot. I want just one Gemini that connects to all these tools, that knows everything that I'm doing and has access to all the information within that universe, whether it's my G Suite or my Microsoft environment, including everything that comes with it, whether it's the Microsoft 365 Office Suite, whether it's SharePoint, whether it's Dynamics 365, et cetera, literally everything Microsoft I wanted to know. And I wanted to understand which of the tools to use when in order to be most helpful to me, because that is how they're going to win against OpenAI.

12:11And right now, it seems that OpenAI is doing it to them because OpenAI now has connectors to many of these environments, and you can turn them on and off in order to prevent it from going to the internet and focus it wherever to get the data. And now they have these agentic tools that can generate outputs that compete directly with the outputs that are generated by the Office suite or G Suite. And so I think this is a serious wake-up call for Microsoft and Google, and I'm very curious to see how quickly they respond and how well they respond to this very big threat. Now, if you think that's the last component, there are even more rumors that OpenAI plans to integrate a payment checkout system straight into ChatGPT, and that will lead to several different things.

12:50The first thing that it will lead to is you can use the agent tool to do everything that you want it to do, to go and research a specific topic, to compare different options, to pick the right options, and then to actually go and purchase that option for you. That could be a trip. This could be clothing. This could be food. This could be booking a place at the restaurant. It could be anything you can imagine, including doing the checkout for you. Now, what OpenAI shared in their launch is that when it comes to payments, you can decide to put your payment tool straight into ChatGPT, or you can ask it just to send you the link to do the checkout on your own.

13:19I think over time, as we give it more trust, it's going to be a no-brainer, just like today we're used to saving our credit cards on Google or other sources. We will probably do the same with ChatGPT, which means you may want to see what it's about to buy for you, but once you trust it completely, maybe that's even going to be redundant and you're just going to allow it to do your shopping for you across the board. This kills many, if not all, or at least most of e-commerce websites. So if I was Amazon, I would be thinking very, very hard right now, how do I counter this new threat? If I am Shopify, well, Shopify made the right move.

13:52They've done a partnership with ChatGPT, and that's going to be the first big partner where you'll be able to go and shop things across all the Shopify stores, which makes perfect sense to Shopify, makes perfect sense to Shopify users, and makes perfect sense to ChatGPT, and makes perfect sense to OpenAI. Now, in addition to providing a great service, OpenAI will take a cut off the top. So, they will take a few percentages out of every transaction happens, which will give them another very significant, potentially, revenue stream. Again, if you think about 800 million weekly active users. And if you think about there's an opportunity to convert all of them to shoppers as well, because this will help you shop across multiple platforms, find the best price, compare different options, read the reviews, basically find the best option for you.

14:30This is way better than any other option out there today, which means more and more people are going to do it, which means less and less traffic to traditional e-commerce websites. So why is this so important? And why did I decide to step away from my entire family to record this episode? This is a complete game changer. From my perspective, the release of an agent by ChatGPT, as I said all along, is a new GPT moment, meaning it's as big and as important as the release of the original ChatGPT because it changes the way we interact with computers and with the data around us. It allows us to do significantly more with significantly less effort.

15:05And if you still don't understand the difference between that and a regular chat with ChatGPT, in a regular chat with ChatGPT, you have to give it very specific instructions on what it needs to do. step by step, one by one, monitor what it's doing, correct as you're doing it. Also, when it needs access to a tool, if it needs to write a document, if it needs to browse the web, if it needs to do different things, in many cases, you need to do it for it. Meaning you need to take the data and now do the research and bring it back. You need to take the output and create the document. You need to do all these things and now you don't have to.

15:33It does all of that for you. It figures out and corrects as it's doing the process. So you need to do a lot less for getting significantly more. The stuff that I've done with Manus and Jenspark that now I'll be able to do with Chachupiti is mind-blowing compared to the amount of investment I had to put into them. Now, the impact of that on everything we're doing is profound. First and foremost, we will be moving ourselves one step or a few steps further away from the actual tasks. So if right now you have to prompt step-by-step the AI to do things for you, which removes you from some of the steps in the task, now you won't even define the tasks.

16:05You will define the goal and the tasks themselves are going to be defined by the AI, which means you don't even know what the AI is doing. Now, yes, you can look right now on exactly what it did and you can follow what it's doing and you can stop it and change it and fine-tune it at any given point. But once these tools evolve and you will consistently see that they're delivering the right results, you will stop doing that altogether, which means you won't really know what the tools are doing. You will just give it an input, you define what the output needs to be, and you will then use the output.

16:29This will be true for university students, for high school students, for our personal lives, and definitely for the day-to-day in our businesses. Now, is that scary? Yes, probably to most of us. Is that going to be very helpful? Well, it's going to be very helpful if we know how to use it, A, effectively and B, safely, because otherwise it's a terrible opportunity for really bad things to happen because we remove ourselves from the process. Now, the other thing that we need to ask ourselves is how good is this tool right now? Is it currently a cool demo tool like the demos that OpenAI have done when they launched it?

16:58Or like, I'm sure we're going to see thousands of demos online within the next few weeks. Is it good enough for basic tasks? Or is it at an enterprise grade deployment level? I don't know. If I had to guess, I would say that right now it's probably somewhere between a good demo level to a good enough for basic tasks level and depending on the tasks. But the reality is, it doesn't matter. It doesn't matter because of two different reasons. Reason number one, once you open this Pandora box, there's no going back. You cannot put it back in the box. Once more and more people understand a genetic capability, they will want more of it because it is really magical.

17:31The other reason that it doesn't matter whether it's there or not is that it's going to get there. In the very short term, companies and individuals are going to find workarounds for the big issues that are stopping them for using it at a wide level. And yes, it will not be able to do everything, but it will be able to do a lot if you know the limitations and you can work around them. Think about our kids in high schools and universities having the opportunities to basically do the work of an entire course in a few minutes by just giving the right prompt and letting it run through the entire content of the course, summarizing all of it and creating whatever report they're supposed to create.

18:01I don't see any student doing anything else unless he's sitting in a classroom with a pen and paper and needs to do it by hand, which has its benefits, but it definitely does not prepare that young individual to the future of doing that at the workplace or in the society. So there's a lot of questions to be asked and a lot of unknowns when it comes to the future of these systems. And then the last reason is obviously OpenAI will now have access to huge amounts of data of actual real-life usage, which will give them more information on what is working and not working with this tool and allow them to upgrade and update the tool in order to make it enterprise-grade tool that can be used for more or less everything in our work.

18:36Now, as I mentioned, I don't have access to it yet. I started seeing people that do have access to it, but based on my experience with Manus and GenSpark, which are similar tools, I can tell you it is a complete game changer. And within the next few weeks, we'll start seeing more and more examples from more and more companies and individuals sharing how they're using the tool, how it works, what are the limitations, and so on. But then there is the final question that is related to this, which is how good of a prompter do you need to be in order to actually enjoy these tools? So what this tool does is it opens an even bigger gap between the people who know how to use AI versus the people who do not know how to use AI.

19:09I meet these people every single week. This is what I do. I teach courses. I teach workshops to different companies. And there are more people right now who do not know how to properly use the basic AI tools like Chachapiti, Claude, Gemini, et cetera. And they're using it in a very superficial way without having deep knowledge on how to do this. And this new functionality is just going to wider the gap between the people who know what they're doing with AI to those who don't. If you know what you're doing, that puts you at a very significant advantage, both from a career perspective, as well as from a company-wide perspective.

19:38If you are in a leadership position and you and other people in your company know and understand how to use these tools, you can run circles around your competition. And that's going to be even more dramatic now with access to this tool. If you are not one of those people, if you're not one of these companies, your competition might learn that first and then they will run circles around you. Now, if you want proof, OpenAI themselves just shared that they've developed OpenAI Codex, which is their cloud-based coding agent, in just seven weeks from scratch. So one of the most advanced coding tools that was developed using AI by people who know how to use AI was developed in just seven weeks.

20:14By the way, Codex itself generated 630 ,000 pull requests, PRs, which is a process in a code development in just 53 days. That's over 10 ,000 PRs per day, and the numbers are just going up. This is outpacing traditional coding teams by 50%. 50%. That means the people that are using codecs are creating code 50 % faster than people who are not using it. It is very similar to what we hear from other sources. If you look at the recent information from Microsoft, they have generated 600 ,000 PRs using GitHub Copilot. And so similar numbers, a huge spike, and they're reporting 30 % faster code reviews and better code reviews than they did manually with people before, which means they can develop the next version faster, which means they can now deploy it, which means they can now develop it even faster and so on.

21:03We're getting to systems that are basically accelerating their own development and becoming better and better. And the same thing will happen to companies who figure it out versus companies who do not figure it out. Now, if you need a little more proof that agents are the next big deal that you have to learn or you will stay behind, Butterfly Effect, the Chinese startup behind the viral agent Manus, which again, I've been using for a while. It started as a Chinese company. They then opened another office outside of China, and now they're closing down their Chinese-based team and moving 100 % of their operations outside of mainland China.

21:38They have relocated all core 40 engineers to Singapore, and they've established a headquarters in the US, all of that in order to attract US investors and US users and disengage themselves from the scrutiny of being a part of the Chinese economy. Part of it is just the way they want to be seen, and part of it because the US government is restricting AI investment in quote-unquote countries of concern, China being one of them. So where are we? We are at a point where as of the next few days, 800 million weekly users will have access to very powerful agentic capabilities that, by the way, did not exist at all by any tool four months ago.

22:14So the mannest moment happened in March. It's just very, very recent. And again, it probably has tens of thousands of users, maybe hundreds of thousands of users, but Chachipiti has 800 million. The problem with that is that a huge part of the people around the world, even people who somewhat use AI, are completely not ready for this. They don't understand how the tool works and they definitely don't have the skills. And the knowledge gap and skill gap is just gross. As I mentioned, I meet these people every single week and most of them barely knows how to prompt properly a basic AI tool. So what are we doing?

22:48We're basically taking somebody who's learning how to drive and giving them access to a Formula one car. That is not a good idea. Now, I know what some of you are thinking. Some of you are thinking that agents, because they're so sophisticated and they know how to define their own tasks, may reduce the need to know how to prompt. And in the long run, I would probably agree with you. But in the immediate future, I think there's going to be a huge difference between the people who know how to use these tools properly and the people who don't. And to be fair, I think the people who don't are actually going to waste more time than the time that they're going to.