Speaker 1: New feature within ChatGPT. A feature that since the month of December, during this period of announcements that OpenAI made, we knew it was going to come out. We are talking about ChatGPT Tasks. The ability to configure in ChatGPT reminders and tasks that are going to be executed periodically. Hey, do you want ChatGPT to make you a daily summary of the most important technology news? Well, you configure your task and every morning when you wake up, ChatGPT will have made you a daily summary of the news. And this feature, we knew it was going to come sooner or later. We know it was leaked, that it was going to come out in the month of December, but in the end, for some reason, they delayed it. And yesterday afternoon, OpenAI finally put this new feature in our hands. So let's try it. And to try it, the first thing you're going to need is to be a paid user. This is a feature that for now is only for Plus, Pro and Teams users. Although OpenAI has already indicated that in time it will be available for all users. If you have access to the functionality, you will see up here that a new button appears in the avatar menu that indicates tasks. And it tells us that it is in beta phase, which is very good because after trying it for several hours, the tool, and you are going to check it in this video, still has many bugs. So much so that if I click on this button now, you will see that in my case it does not let me access. Now I explain exactly why. So, being so, the alternative to be able to create a task will be to come to the model selector and select the new model that appears, which is GPT-4 with programmed tasks. What you don't have down here, ask ChatGPT to do a follow-up later. And here again, I get angry because I don't understand why OpenAI continues to insist on putting functionalities, which are complementary tools, to the model within the model selector. I understand that at the level of how things work, it is really a different model, it is a fine-tuned GPT-4 to be able to work with tools, if I understand that. But most users who are not technical do not have to understand that. And it is easier to add this functionality as a button down here than to have to put it in the model drop-down, which is already quite confusing. In fact, if you remember, I made this same complaint with the Canvas tool that in the end they ended up placing right here. So I don't understand why they did it like that, but hey. We select the GPT-4 model with programmed tasks, we click here, and now we have the same ChatGPT interface, but where we can ask it to configure different tasks. I want you to make me an exercise table every morning to stretch after getting up. We hit enter, and in this case we will see that it has created, well, look, a task. News. Morning stretching table. Daily at 9 in the morning. This part is the most interesting part of working through a language model that can understand the context of what we are asking for. This is not the typical reminders that we configure with Siri or with Google Assistant, that you have to specify exactly what you want, but here the language model, when I said I want an exercise table for the morning, when I wake up, it has assumed, well, surely it is because it does not know that as a YouTuber at 9 in the morning it is not yet time to wake up, but it has assumed that this reminder has to be configured daily at 9 am. At the same time, we see that a banner has come out below that tells us to activate the desktop notifications. Think that this is now going to be a recurring task that is going to be executed in the background, with which, in some way, ChatGPT has to notify us that this task has been carried out. And this can be done in two different ways. One can be through the email, through the email, which I do not recommend for now, because after trying several different examples, the email is filled with a lot of emails from OpenAI, but they are also emails that do not provide you with anything different, right? There were people who said, well, this is interesting because I can configure ChatGPT Task to make me, every night, a story of the news that has happened during the day, that I get the email and therefore I have a reading that I can access and see, right? Notice that another email has just been updated. But if we enter, we will realize that really many of these messages are going to be covered. It is simply a brief reminder that reaches you to the email and that gives you direct access to the ChatGPT conversation to be able to see the entire context. With which, in one way or another, we have to end up accessing ChatGPT. With which, being so, my recommendation is that you come here, Configuration, Notifications, and activate only the application notifications. The ones that will take you to the ChatGPT application itself, both on your desktop device or on your mobile. And if you are not happy with how ChatGPT has configured the task, we can enter this menu here and we can change the name, we can give it more instructions, we can make the prompt more complex if we want, and we can change the periodicity or the time of the task here. You can pause them, you can delete them, and well, typical functionalities that one would expect from a system like this. So look, it's interesting because we can configure two different types of tasks. For example, we can make a reminder, a task that we only want to occur once in a while. Remember me in a minute, drink water. We hit enter. And in this case, ChatGPT has created the task and has sent us an error message. This is what I was telling you, that it is still noticeable that we are in beta phase, and in particular, this is an error that many people are reporting, where when we specify a task with a specific time, for the next 20 minutes, for example, then the following error pops up. In the application console we see that there is a specific error with react-rendering, an error with the time value, which may even be with the application itself in Spanish. Even so, although here an error has popped up with the configuration of this task, this task has been created, okay? I have already checked this on other occasions. If I go to the mobile application, I get that there is a water reminder task, and we can see, indeed, that here it appears, drink water today at 1.31 pm, okay? Where we can pause, if we want, or continue the task itself. Well, in this case, I got two different conversations. On the one hand, there is the task conversation, and we see that the conversation of remember to drink water has been generated. For the first time, and this had not happened to me before, this reminder has been done to me in a different conversation, which is a different dynamic from the one I was seeing so far, because normally this reminder did it to you within the conversation itself where the task was specified, with which, well, a little strange behaviors. But we can see, indeed, how notifications are coming to my mobile, which tell me, precisely, that I have to drink water, that I have to stretch my back, and the different configurations that I have done before. With which, that is one of the possible tasks that we can configure specific reminders in the future. But I think the other thing is more interesting, the recurring reminders, that you can say, hey, every so often I want you to execute this prompt so that it solves a task for me. And of this type, for example, just, well, look, I just got the notification of outstanding news of today's IA, that this is a recurring task that I have defined as make me a report of the IA news of the last few hours every 20 minutes. And here, ChatGPT, well, it configures the task for you, it tells you, of course, I will inform you every 20 minutes. And here we can see, within the same conversation, not as it has happened before, well, that every 20 minutes, ChatGPT is making a summary of us, as if we had executed the prompt, it searches the internet, because it has access to the search, canvas tools, all the tools that we have integrated in GPT 4.0, it executes the task and every 20 minutes it sends us a report of the news that we have asked for. Things that I don't like when trying this tool, or that I don't know if it might clash with how I expected it to work. When I ask ChatGPT to make me a report every 20 minutes, well, a little bit of what I'm looking for is for the conversation to continue coherently. In this case, if you notice, each report of the news it makes me every 20 minutes, that in the end the news don't change, well, it's a copy-paste of the previous one. They are the same news, with the same structure, repeated over and over again. It doesn't change anything. So, in this sense, if I would have liked this to work so that ChatGPT wouldn't execute the prompt and turn it into the same conversation, but that ChatGPT would intelligently say, ok, if I have informed you of this news 20 minutes ago, in the next 20 minutes I could look for other news, or at least answer you that there has been no news regarding the previous news report. In fact, I wanted to try this and I have tried with another prompt. I want you to do a search every 20 minutes of the latest news about it and compare it with the news you have shared with me previously to see if there has been any update. That is, your task now is not to give me the news, but to compare it with the news you have told me previously to see if there is any news. Just in case there is any news, I want you to tell me. I just clicked here, it just failed. Well, ok, it's in beta. And notice that the execution of the task has come out a little weird. To begin with, it answers in English and this is a problem that I am verifying that happens with GPT-4 models which is that when a task is assigned, when this fine tuning is done, this culture of maintaining the multilingualism is being lost a bit and suddenly it responds in a very usual way in English. And also, notice that the answer, without having given any previous news report, does not include anything, but it is telling us that there is no news. Then, recurrently, it reviews again. There is no news either, but it has not given me any news of the first ones. And then, in the end, there is a problem and it fails. Failure that, although we cannot see it now because it has disappeared, was represented by a label that said delayed, as if the task had been delayed. Many, many failures. Notice that the idea of the tasks is good and surely all these errors end up being fixed, but I am very surprised that OpenAI continues to release products, updates, in addition to the fact that it has taken its time to be able to clean it up from what it had in December and yet, when it comes out, it is so full of errors. Another error that I have found very curious is the following and it has to do with the limits of how many tasks we can run simultaneously. If you counted on having an army of tasks running with hundreds and hundreds of tasks, remembering at every moment that you stretch, that you drink water, counting other news, you have to know that for now there is a limit of 10 active tasks. A limit that seems quite, quite tight to me. 10 active tasks are not so many. Personally, it does not occur to me as many as I want to configure, in my case, but I could imagine many users who do, but then there is another limit that has to do with periodic tasks. If I ask you, for example, to remind me every 5 minutes that I drink water, let's see that ChatGPT tells us that there is a limit where you can only run a periodic task 4 times per hour. Okay, well, as a maximum of 4 times per hour, let's say that it reminds us not 15, but 16 minutes. So we make sure that it is less than 4 per hour. And even so, we see that the same error pops up and here I don't know if I'm doing the math wrong or what, but this seems to me that it is also poorly configured. So what we have for now is a fairly promising tool. I think it can be very interesting, but for now it still has a lot of work to do to solve many of these errors. Even so, many of these errors will surely be solved in the next few days. In my case, I have reported to OpenAI everything I have been finding, these errors that are surely affecting, especially the Hispanic community. I have tried things like this, they seem pretty cool to me, right? Well, for example, I want you to suggest an intermediate-level Python programming challenge that I must solve every 20 minutes. These exercises will be related to some topic of the daily tech news. You can look for some relevant news on the internet, taking advantage of the search capabilities, the tasks, all together. And in this case, we see that every 20 minutes adds to the conversation a new programming challenge. Challenge. News analyzer on autonomous vehicles. It has searched the internet, it has seen that the topic of autonomous vehicles is a popular news, and invites us to implement a system that collects data from at least 10 recent news articles using Python, okay? And it tells us how we could address the problem, what resources, what libraries we could use. 20 minutes later, it proposes to us again energy consumption analyzer in data centers with Python. ChatGPT has realized that a relevant topic in tech is the topic of energy consumption and proposes to us to do an exercise of analyzing the trend of electricity consumption in the data of a CCV file. In addition, we could ask it in the reminder itself to implement a solution for this challenge that we could even open in the ChatGPT canvas itself to study the code, edit it, execute it, well, all integrated in a single tool. To finish, two more reflections that I want to add with the publication of this tool. One in positive and one in negative. I'm going to start with the negative. I'm a little worried to see that OpenAI continues to add tools that continue to fragment the space of functionalities that have to give support every time a new model is updated. We see that we have models that have not yet arrived, that some have some functionalities activated, others deactivated, and we see that they continue to bet on functionalities under this premise of doing fine-tunings to models that learn to use tools. Well, there is a GPT-4 model that uses the Canvas tool, there is a GPT-4 model that uses the Task tool, and this worries me because it seems that a very strong bet is being made to cover all that space and give support to all these tools on the GPT-4 model, something that sounds strange if we were really at the door to see an update of the GPT-4 model to more functionalities, more fragmentation, and more complexity. But in positive, this new step that OpenAI has taken with these tasks has to be celebrated as the first time we see ChatGPT having autonomous agent attitudes. And I know that a functionality as stupid as this, which is simply adding a stopwatch and executing a prompt, is overloaded a lot when we call it autonomous agent. But it is true that it is the first time that we are seeing ChatGPT doing things without us having to be on top of it, that in the morning, when you wake up and make yourself coffee, it will already be writing that summary of news that you have asked for. On a technical level, this has nothing to do with autonomous agents, but it is a first step in that direction. This is celebrated with the same pride that when you have a child and you tell him for the first time to go to the corner bazaar to buy bread and come home with the bread. It is celebrated, it is not a complex task, but hey, it is a first step of a functional adult that in the future will be doing more complex and more autonomous actions. And in that sense, seeing ChatGPT taking these first steps, even when the product and the interface have all these errors, let's remember that it is in beta phase, well, guys and girls, it is celebrated. With this, this is the analysis of the ChatGPT tasks. If you liked the video, share it. Here you have a couple of more interesting videos about artificial intelligence. And see you with more news in the next video. I'm going to give it to the camera, I can't reach it.