OpenAI Dots: testei o assistente pessoal integrado ao Codex

Author: Rafael Quintanilha – QuantBrasil · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: OpenAI Dots: testei o assistente pessoal integrado ao Codex · Channel: Rafael Quintanilha – QuantBrasil · 39:43 · Watch on YouTube

Topics: AI agents, AI models, Multimodal models

Summary

The video reviews OpenAI Dots, its personal assistant integrated with ChatGPT and Codex, and compares its features and maturity with Grokbot. The presenter demonstrates shared conversations across devices, access to a local computer, cloud coding tasks, and voice interaction, while noting confusing UX, slow responses, and occasional model failures. He finds Grokbot more mature and responsive, but says OpenAI’s voice sounds more natural and its assistant does not use the main plan limit. He sees promise in using an assistant to discuss a product or analyze data and then dispatch work to a cloud agent.

Key points

Dots is OpenAI’s always-available personal assistant, powered behind the scenes by GPT-6 Astra according to the presenter.
At the time of recording, Dots was available only on Pro and Business plans, and the first Dot did not consume the main usage limit.
The assistant uses a single-agent model, unlike Grokbot’s earlier multi-bot approach.
A Dot conversation is shared across ChatGPT, Codex Desktop, and mobile, but the presenter finds the ChatGPT interface confusing.
Dots can access a persistent cloud computer and, in the demonstration, operate the presenter’s local computer through Codex.
Codex Cloud can run coding tasks in a configured repository environment, including tasks dispatched through the Dot.
The presenter encountered slow task handoffs and cases where the model stopped without completing the expected next step.
Voice interaction improved during testing and sounded natural, but pauses and slow responses made it less useful for conversation.
The presenter considers Grokbot more mature and responsive, while praising OpenAI’s voice naturalness and more direct answers.
He sees a useful workflow in discussing product data with an assistant and then sending a related task to a cloud agent.

Chapters

Start	Chapter	Summary
00:00	Introducing OpenAI Dots	The presenter introduces Dots as OpenAI’s personal assistant and frames the video as a comparison with Grokbot.
02:05	How Dots are positioned	He describes Dots’ always-on, learning assistant concept and compares it with ideas already present in other tools.
05:25	One assistant, plans, and usage limits	The presenter discusses proactive assistance, the single-agent design, plan availability, and Dots’ separate treatment of usage limits.
09:38	Interface and local computer access	He demonstrates how Dot conversations sync across clients, critiques the interface, and shows a task operating his local computer.
16:52	Codex Cloud and task dispatch	The presenter explains Codex Cloud environments and compares their setup and capabilities with Cursor’s cloud agents.
21:00	Testing parallel cloud tasks	He dispatches website color changes through ChatGPT and Dots, then observes both the parallel execution and a model’s difficulty completing a task.
26:13	Voice mode in practice	The presenter tests voice requests for analytics and coding work, finding the voice natural but the responses slow and sometimes awkwardly delayed.
32:57	Comparing voice assistants	He tests Grokbot’s voice mode and compares its responsiveness and visibility into actions with Dots’ more natural-sounding voice.
36:02	Verdict and closing	The presenter summarizes the trade-offs, sees potential in connecting product discussions to cloud tasks, and invites viewers to share their preferences.

Quotes

your main DOT does not use up your Coldex limit. — 03:08

regardless of the client you’re using, whether it’s Codex Desktop, the web, or even here on mobile, you’ll have the same conversation. — 10:09

the cloud concept here is exactly the same as the Cursor cloud — 20:28

Naturalness without agility hinders the conversation. — 31:25

I personally think Grockbot is a more mature product. — 36:02

Concepts and entities

Dots, OpenAI, Codex, Grokbot, Hermes, GPT-6 Astra, Cursor, ChatGPT, HostGator, agentes na nuvem

Full transcript (original language)

00:00 After much anticipation, OpenAI finally unveiled its newest product, the Dots, at Dev Day 2026. Their own take on personal assistants, heavily inspired —to put it mildly—by Grokbot. So, in today’s video, I want to tell you everything about the Dots, these personal assistants highly integrated into the OpenAI ecosystem, and show you what they are and examples of what you can do. I will compare them to Grokbot and answer the question: is OpenAI ready to win the race for personal

00:31 assistants? Folks, those who were following OpenAI’s social media before Dev Day already realized they would come out with an alternative to Grokbot , right? People were really anticipating a new model, especially after GPT-6 Sol fell far behind OPO 5.5 , and they did indeed come out with GPT -6. A Sol, which was indeed an advancement, even if still inferior to its competitor. I’ll leave the link here for you to watch the review I did

01:04 of GPT-6.1 Sol. In this video, I want to talk to you specifically about the Dots and about personal assistants, a subject I always enjoy. I have already spoken here on the channel several times about various other personal assistants. For example, I compared Muse, Meta’s personal assistant, with Jam Night Park, which I will also leave here if you want to watch. And I’ve brought Grokbot to the channel plenty of times, which is my favorite personal assistant today and for which I’ve already shared several use cases. These

01:35 use cases are quite relevant to our video today because everything you could do in Grokbot, you will be able to do here in the Dots, because they really are very similar. So check out the Grokbot use cases after this video, okay? But let’s head straight to the launch of the Dots. So what did OpenAI tell us, then? The Dots are agents that are always available, always on, and considerably capable, built to manage

02:05 anything. So it’s a new way of working with AI, one that understands what matters to you, is always working in your place, and takes the work off your shoulders. It takes work off your plate so you can have more time and attention back. Well, a new way of working with AI isn’t exactly true; this has existed for quite some time. Personal assistants have been around, you know, for a long time. And this same idea of an assistant that understands you,

02:36 right, is also nothing new, because back with Hermes, we already had this idea of agents that learn, generate skills, or have that “dreaming” functionality, where the model understands what you did during the day and consolidates it so it doesn’t make the same mistakes, and so on. So, it’s basically OpenAI, you know, catching up to the competition instead of bringing something necessarily new, okay? So alright, the Dots work with the frontier of intelligence. This is very interesting because behind the scenes

03:08 they are using GPT-6 Astra, which is a model I’ve also reviewed here on the channel. I liked it for some things, but it’s very expensive. So the fact that Dots run on GPT-6 Astra is interesting, to say the least. And even more interesting is the fact that your main DOT does not use up your Coldex limit. Which is something quite interesting, we’ll talk about that in a bit, okay? So alright, the Dots have their own computer in the cloud, right? Just like Grockbot, they learn over

03:40 time, just like Hermes, and can work on your goals 24/7. Through OpenAI’s plugin ecosystem, they can quickly connect to over 4,000 applications, you know, giving them the necessary tools to solve any task. So this is a really interesting thing, the fact that you can easily connect to your tools is something we also have in Grockbot, right? So your DOT is always at your disposal, whether in ChatGPT, Slack, or

04:11 Teams, okay? You can ask questions, explore ideas, provide feedback, or simply make a voice call if you want to chat with it. This here was something that was heavily pushed during OpenAI Dev Day and it just went completely wrong, right? Yeah, I’ll even ask the editor to put it in there because it was actually kind of embarrassing. All

04:53 right. Um, how’s it going? Yeah, they clearly tried to lean quite heavily on the demo to show off voice mode and it simply didn’t work. And honestly, in my initial tests here with Dot, it was still pretty bad, okay? There were several times when they weren’t responding to me. So I don’t think it was just a one-off thing during the demo where the Wi-Fi was bad. Their system really wasn’t able to handle it

05:25 well. Alright. Continuing here, the more you work with your DOT, the more it learns your preferences, how you think, and what works best for you. So, naturally, you will need to test the dot more so it gets accustomed to your use case, okay? And there is an interesting thing they mention here, which is that one of the magic parts of the Dot is when it brings work already done for you without you even asking, okay? So, it is that idea of the agent being more proactive. This is something

05:57 I haven’t really seen work much yet; if this were perfected by the OpenAI side, it would be interesting, but I confess I haven’t seen it work very well yet, no, okay? Now, look here, everyone, at the moment you only have access to your primary dot, that is, your main agent. Coincidence or not, Grockbot launched Primary Bots on October 2nd, meaning that over at Grock, you will now also have a single bot. And why does this make a difference? Because at OpenAI,

06:29 you have the idea of talking to one agent. And over at Grock Bot, the idea was for you to have a set of bots where each bot would work on a specific function, and then you would have this routing between bots, okay? even being able to create a conversation with multiple bots to solve a specific task, okay? So this is a design change. OpenAI is betting on a single bot. Grockbot started by betting on multiple

06:59 bots and now they seem to be shifting to the primary bot idea too. I confess that I still don’t know which is the best architecture. I understand that having just one bot is simpler from the user’s point of view because you always interact with your own assistant, right ? Now we need to see how this will be in practice, because having more specific bots clearly makes it easier to better solve that task. How will the context handle things if you are working on many things? This is something we will need to explore, right, more to know what the result

07:29 will be like. It seems the industry is still adjusting to what the best proposal is, okay? Now, the not-so-good news is that at the moment I am recording this video, Dots are only available for the Pro and Business plans, meaning you have to be spending at least some dollars there to use the dots. This isn’t necessarily a surprise , because when Grockbot first launched, it was only available for Cursor Ultra, which was a much more expensive plan. And eventually, as they understood the costs and the compute required, Grockbot was added to the $ 20 Cursor

08:01 plan as well, okay? So I think this will possibly be available to ChatGPT Plus users soon, too. Now, if it does come to the Plus plan, it likely won’t be Astra, right? They don’t say, but we imagine it’s Astra on reasoning-low, which powers the dots nowadays. And maybe when it comes to the Plus plan, it will be GPT-4o or even GPT-4o mini, which has very good availability; we’ll

08:33 have to wait and see, alright? And why is this relevant, guys? Because look, your first Dot is included in your Pro or Business plan at no additional cost. So you don’t use up your limit when interacting with your Dot, which is very interesting and different from the approach Cursor took. If we go here to the Cursor dashboard and search for “ spending,” you’ll see that we have three limits for a given person. We

09:04 have the limits for Cursor’s proprietary models, the Cursor models, in this case, Grock and Composer. We have a second limit just for other models, models that aren’t from Cursor, like Opus, for example. And a third limit for Grockbot alone, okay? So I’m on the $ 60 Pro plan here. It lasts quite a while, but it is a separate limit for Grockbot. In OpenAI’s case, they simplified it; it’s a single limit , and your Dots, or at least your first Dot, don’t consume that limit. So

09:38 that’s quite interesting, especially considering it has GPT-4o behind the scenes. Great, everyone. So now I want to show you my Dot in practice, okay? I’m here in my desktop app for ChatGPT and up here there’s a little button to configure the Dot. I’ve already set it up. I’ll click here and I’m in this conversation, and you’ll see it’s a single conversation, okay? No threads, no sub-agents. It’s just a single agent working for you 24/7. Got it? So that’s the idea. That is the mental model for

10:09 the Dot. So, I don’t know, I’ll send a message here. Hi, how are you? And one thing I want to show you is that this same message will appear here in my ChatGPT interface. So, if I come here, let me select my dot here in the ChatGPT web UI, see? I’m here on chatgpt.com. I’ll select the same pen here, which is my dot. And look there, the same message appeared here. So this turned out really cool, because regardless of the client you’re using, whether it’s Codex Desktop, the web, or

10:40 even here on mobile, you’ll have the same conversation. This is something quite interesting, okay? Now, one thing needs to be said, because the ChatGPT interface, as it is today, is extremely confusing, guys. If we look here at standard chatgpt.com, you’ll see the two tabs we have here, which were already confusing, right? The Work tab and the Chats tab. And then Codex is kind of lost. Here, see, under these three dots is where you’ll find Codex, you click here. And there you go. Here

11:11 now you have Codex Cloud, which I’ll talk to you about in a second. So here you can dispatch tasks to the ChatGPT cloud, very similar, extremely similar to Cursor’s cloud agent; any resemblance is no mere coincidence. And yes, here you have the possibility to speak directly to your dot. So look, look at how many ways you have to interact. You can interact here normally, through the normal tab, you can come here to Codex. and start a

11:41 conversation in the cloud. And you can come here with your Dot and also have a conversation in the cloud. So, it ended up being a very confusing interface, a UX I didn’t like very much, and it seems the folks at OpenAI received quite a lot of this feedback. So much so that there’s a tweet from today by Tibo saying that, well, they are focusing. The only things being worked on at the moment are simplifications, more efficiency for more usage, features that make a difference, or new models, right? Basically, he said it

12:13 all here, but what I wanted to highlight is that the first thing he mentioned was simplifications, meaning they are looking to simplify because it really became a bit confusing and unintuitive, I’d say, for most users. But okay, let’s go back here to the dot and analyze the interface, okay? You’ll see that the dot has its own computer here, which is a computer that is always available, 24/7. And if you click, you can see exactly what that computer has access to and even take

12:44 control here and start operating that computer. Now, don’t be surprised to see that this is very similar to the Grokbot interface. If you click here on Grokbot, you’ll also have a computer, and you can also open it here and control the computer. So it really is the same mental model; it really is the same thing you can do with both Grokbot and Dot. All right? But one thing I really liked here about Dot is that, besides the remote computer, it has

13:14 access to my computer here. Those who follow me here know that I am a big fan of OpenAI’s remote control. For me, it is by far the best way for you to control agents from your phone. You have to have your computer here, right, turned on, but you can, you know, operate your computer remotely. This is very good, but with the Dot interface, you can do it from your Dot. Which is very interesting. So, I’m going to do the following: I’m here at chatgpt.com

13:44 and I’m going to say the following: “go through my MacBook and open Chrome at programe aí.” So, look, I’m here on the web interface interacting with my Dot and I’m asking it to go to my local computer and open something locally. This is to show you that it has access to my local computer as well. So, okay, it already understood what it needs to do. We can see here, look, that the agent is thinking. And here comes a

14:16 main piece of feedback that I didn’t like much about Dots, which is precisely the fact that we get a bit lost. When we are working, you know, in a thread what the model is doing, if it is heading in the right direction, what it is opening, what it is thinking about doing. And here with the agent to the user, and we end up not seeing it. But look how cool this is. You see that it opened a thread in Chrome. I’m even going to go back here to the Codex

14:47 because we can click here, look, and see that this is a thread my threads here in the Codex. Open “programe aí” in Chrome. This is running. I can come here and see exactly what is being done . Look at the message sent by the agent . This is controlling my local computer . And the coolest thing is that this was dispatched through the ChatGPT Web interface. Look how cool. We can see

15:17 here that it will possibly resolve this quickly. Again, for those on the interface. Look at that, it’s operating on its own here. Oh, how wonderful. Oh, how wonderful. For some reason, it opened the campaigns page; I don’t know why it keeps operating my computer here . Let me tell it this here. I want you to go to the homepage. But one thing that is a bit annoying here is that you’ll see it’s a bit slow, okay? Because it has to think about what it’s going to do. Then it has to dispatch it to that local conversation in Codex.

15:49 Then Codex has to receive that, and then Codex has to go to the home page, okay? So, if we click here, you see, it’s still thinking. I don’t think it has even received the request yet. We come here to Codex, let’s see if it has sent it already. Now it has received it , look. User clarification. I want you to go to the homepage. So it will go there and understand what needs to be done. Now, one interesting thing, okay? It’s running here, the 6.1 extra high. Yeah, I didn’t configure this at any point, so I think it uses the default I have on my machine, which is GPT 6.1

16:20 sox far. So that is a cool thing, okay? So alright, it has already understood what to do. Let’s go back here and wait for it to enter the page. It works, folks. Now it must be said, this process is still a bit slow, okay? This operation is not the most straightforward in the world. Now, look at that, it’s controlling the page here and done, finally it entered the homepage without me touching anything. Would it have been faster for me to go directly through the Codex app in Remote Control? It would. So, at the moment, as much as I think this is cool

16:52 , I think for most people who are already used to using remote control, it won’t change much, okay? Perhaps what changes most is you being there in conversation with your assistant and then having the need to dispatch a task to the cloud. That is what I want to show you next. Alright, folks? A second piece of news I wanted to bring to you today is the fact that OpenAI now has a new cloud, the Codex in the cloud. This already existed, but it had fallen way behind other competitors, notably

17:22 Cursor, you know? And now you can dispatch tasks, right, in the cloud within your Codex or even within ChatGPT or your personal assistant. The cool thing about this is that it runs 100%in the cloud and you don’t need to have your local computer, you don’t need to have internet, etc., etc. How does it work? Well, the first thing you need to do is set up an environment. So , when a new conversation starts here, you noticed it appeared here for me,

17:53 look, on the cloud, but I can select this computer, meaning local. You will see that I can also choose the remote option here. And then, it automatically selects my HostGator VPS, which is my main choice for VPSs and is also a partner of the channel. So, if you are interested in a VPS with 24/7 support in Portuguese, my suggestion is that you check out HostGator. And the link for you to check out the HostGator VPS is both in the description and in the pinned comment. And the VPS is

18:24 interesting, why? Since we are talking about Dot’s personal assistant and making a comparison with Grockbot, we can’t forget that Hermes, which is an even more customized personal assistant that allows you to do much more, can also be installed on a VPS quite easily , okay? So if you are interested in having more control over your personal assistant, I also recommend using a VPS like HostGator to host an Hermes instance there, right? But what I want to show you today is that you can

18:55 execute a task from the cloud. And here you see that I already have the program’s environment configured and I can click here to create a new environment. And it will allow me to create an environment based on a repository I already have connected here on GitHub, letting me dispatch tasks to that repository in the cloud, okay? Since I already have the program one selected here, I’ll show you how it works in practice, right? So we have this environment here and basically you will configure its environment

19:26 variables so that when you initialize the environment it has all the necessary variables and also what the security policies are, whether you will allow internet access, which domains you will allow, and so on, okay? So here you have the installation, you know, the scripts and models are now capable of doing this on their own, okay? If you’re going to allow internet access, the environment variables are here too. Coincidence or not? Again, coincidence or not, this is very similar to what we have in Cursor,

19:56 right? If you come here to Cursor, look at this. You also have an option here to select the computer. You also have the option to select the cloud, which is basically the same thing, the same environment, running on an agent in the cloud. And you can also select a remote machine here, like in the case of Host Gate. The difference here with Cursor is that they have a slightly worse interface for remote control. You have to start the conversation by selecting remote control, which I think is worse, right? In Codex, you don’t need that.

20:28 You just start a local conversation and you can come here on your phone, for example, and control it from there, as long as your computer is set up for it, okay? So, the interface in Codex is superior for remote control, but the cloud concept here is exactly the same as the Cursor cloud that I’ve already featured on the channel and mentioned several times is possibly the best one out there, okay? That’s why I make the caveat that OpenAI might not have reached the level of Cursor’s cloud agents yet, but Cursor is indeed the

21:00 winner for me in this area today, and I imagine OpenAI will quickly get there too. But what’s cool here, guys, is that I can also dispatch the task here from my computer, you know, locally, I’m here on the desktop, or I can dispatch the task from the regular ChatGPT on the web. So I can come here, look, to Codex New Chat, and the same cloud environment is here too. The only difference is that here I can only run it in the cloud, right? Obviously, I’m in the cloud, I’m not on my local computer here, so I can only dispatch tasks from the cloud. So let’s even do

21:32 a test. Let me switch this here. The GPT6 model to high. I’ll switch it here to fast mode. And then I’ll give it any task here. Please change the primary color of the website to red. Got it? And then while it works here, I can go to my personal assistant. Let’s open the pen here and do a test with it as well, okay? I want you to dispatch a task in the cloud for me for the

22:03 program AI project. And I want you to change the primary button, the color, the primary color for the site to yellow. Wonderful. So, first we need to see if the Pen will correctly dispatch a task to the cloud, but the big advantage is that by running in the cloud, I can have two tasks in parallel , even if they’re modifying the same thing, right? Theoretically, both are modifying the button there, but I can make them work in parallel without one interfering with the other, because they are running in separate

22:35 environments, which is excellent, okay? So let’s see if the Pen will be able to dispatch this task to the cloud. And we can see here, look, this is the one we dispatched, asking to change it to red. And it already said it’s going to start here, look: make AI primary yellow. If we select it here, here is the Pen’s request, look. The user explicitly asked for a task in the cloud, okay? So let’s wait and see if this task will appear. And look at this, guys, coming

23:05 here into my Codex, I can see both tasks, look. Change the color here to red and here change the color to yellow . So this was dispatched by the pen. Got it? OK. Both are running in the cloud. We can see here, look, the little cloud symbol. So it has already started working on the other one here. Supposedly it finished, but look at this, it finished here but stopped, right? It didn’t do anything else. Let me tell it. Where is the code? This already seems like a model problem to me, right? Because it made the change

23:37 in the local workspace. And then, okay? What are you planning to do this? And look, guys, how sometimes the model isn’t ready for the functionality, right? I asked: “Okay, you made the change, what do you expect to do after that, right?” Then it said: “Oh, the next step would be for you to review it in the browser.” Then I said: “How can I review it in the browser if you are in the cloud?” And then it’s now like, whoa, let me check what I can do here, right? So he’s talking, saying you can’t, and then he stops and does nothing. That’s really a problem with

24:07 the model that’s a bit annoying, right? Okay, so, please just create the PR. So I’ll ask him to create the PR here while the other task is also running in the cloud, okay? Look how cool, you can dispatch a task to the cloud not just from the web or locally, but from your own dot. Now, just to be clear, this is something you can also easily do with your primary Grock Bot, okay? I did a test earlier here, where I asked him

24:38 the same thing I asked Pen, look. Please, dispatch a cloud agent to the program there to change the color, right, to green instead of blue. And so he went, worked, worked, and generated it for me, with the difference that the Grockbot was much more detailed, right? He sent me a screenshot of how it worked before and how it looked after, right? The buttons here, the blue and green path, they changed and everything else. So, it was actually much more complete; he sent a grid of all the changes with the new colors. So, it

25:10 really did something much more comprehensive, even beyond what I asked . But again, this is something you could already do in Grockbot, with the advantage that in Cursor you can control your agents, okay? So, if you come here to the Cursor dashboard, we have, look, cloud agents. And here, cloud agents is the same idea, folks. We have our environments here that we can click to edit and see the same thing, look. The installation script, the secrets, environment variables, whether we allow internet access or not

25:42 . So, it’s exactly the same proposal that Coldex made. And the difference is that we can come here, look, and change the models, because in Cursor there are several models we can use. Here, in this case, it was set for me to Claude Opus 3.5. Medium, alright? So, the same solution that Codex presented, Cursor already had. So, this “novelty” in Codex wasn’t a novelty for those already in the Cursor ecosystem. The advantage now is having this in the OpenAI ecosystem, which many people like to use in their daily routine.

26:13 Wonderful. Guys, before finishing this video, I want to show one of the main features that OpenAI insisted on highlighting at DevDay, which is being able to operate through voice mode. So you can see here on your dot, there is a little button for you to call. And in fact, this call even simulates as if it were a phone call. Some people liked it , others didn’t. I tested this several times, guys, and at the beginning it didn’t work, it wouldn’t answer me, later it started answering, okay? So

26:44 let’s do this test together here. Let’s see if OpenAI answers me this time. Hey , nice to. >> Hi, why are you talking to me in English? >> Oh, my bad. Let’s do it in Portuguese, yes. What do you need? >> Look, I wanted to know from you how many people accessed the program in total? How many unique users accessed the program during the launch we held last week? So, we held a launch on

27:16 Tuesday, Wednesday, and Thursday. We had three lives. I want to know what the total number of visits and unique users was in that period. >> Good. I’m pulling the data for the launch period from September 28th to October 4th. And checking the total number of visits and unique users. >> No, it’s from Tuesday to Thursday. Tuesday, Wednesday, and Thursday. >> Oh, I see. Just Tuesday, Wednesday, and Thursday, adjusting for this timeframe now. >> Well, guys, so here I’m showing you a use case that I consider quite cool, okay? Which is having your personal

27:47 assistant and you make there Thursday, September 29th to October 1st, there were 8,076 pageviews and 170 unique visitors to the EAI program. Each person only counts once in total, even if they accessed it on more than one day. >> Right? And can you tell me then how many classes were completed on Friday and Saturday? >> Good question. Let me check this Friday and Saturday period for you. On Friday, there were 130 class completions made by 55 people. On Saturday, 99

28:18 completions by 35 people. Adding both days, 229 completions. >> Is that right? Thanks. So guys, I showed you here an example that I consider quite cool for you to do with a personal assistant, which is this idea of brainstorming, you analyzing your platform, your product, right, your company, talking, asking your assistant to pull data. Here in this example, it was much faster than it was before, okay? If you tested it right after launch, you might not have liked it, just like I didn’t. But in this demo with you here, it has really

28:50 improved. I think they are solving their compute problem, you know? Actually, guys, it worked so well here that I even want to run another test. Let’s try to do something a little more complicated here. Go. I’ll call it again here. Let’s see. >> Good evening. You came back quickly, huh? >> Yeah. Yeah, I see here that I just received a message saying that the task to change the color to yellow is done. Right. >> That’s it. I see it was implemented with yellow#FAC15 and the dark test,

29:21 and the tests passed. >> Okay. Can you send me a screenshot of how this page looks and create a pull request for me? >> Sure. I’m taking care of that. I’ll request the screenshot and the pull request now. >> Right. Uh, while you do that, I published a video this Sunday. Can you confirm to me what position it is in, like 1st of 10, 2nd of 10, etc., on YouTube? Is this video currently where I am speaking to you right now? >> Yes, I will check the current position for you. >> Look, guys, this is a test I want to

29:52 run, which is: its ability to work in parallel, and besides, this second YouTube check I asked for will possibly need access to the studio, which it will only be able to do locally here on my computer, okay? So I have some doubts about how well it will be able to do this and the speed, okay? Because although it impressed me here with the first task in terms of speed, many of the criticisms we saw regarding the dots was precisely the fact that it wasn’t that fast, okay? And when you’re on a call, >> a quick update: the screenshots were generated, but sending the images

30:23 failed due to an authorization error. I’m looking for a supported way to deliver this to you. >> I think if there’s an authorization error, can’t you just ask the agent on the local machine to take a screenshot and send it in the chat itself? >> I can, yes. I’ll request that now. Or you could ask it to put the screenshot as an attachment to the GitHub pull request. >> That works too. I just need to ensure the screenshot shows the yellow version of the task in the cloud, because the published site hasn’t changed yet. >> Well, back here, guys, the big issue is that when you’re working and the model

30:54 goes silent for too long, like it did, especially in a voice message, it feels a bit weird, right? So, >> perfect. I’ll include that option to have the screenshot attached to the PR. Meanwhile, I’m still waiting for confirmation of the attachment, the PR link, and the video’s position in the studio. >> Yeah, it’s still much better than I imagined. I think it’s slow for a conversation that would be, you know, something you’d do day-to-day, okay? But it’s better, it’s progress, and the voice is very natural, isn’t it? You’ll see, we’ll do a comparison here with the Grok Bot too, and you’ll see how

31:25 OpenAI’s sounds much more natural, although it’s slower; it’s no use being natural and not working, right? So, see? The amount of time it takes for it to get back to me here. It just stays silent, right? And for a voice conversation, that’s really bad. Let’s wait here to see when it replies. >> Fair. Makes sense. I’m still trying to show the screenshots in the task chat and check the video’s performance. I still don’t have the confirmed result, but I agree with you. Naturalness without agility hinders the conversation. >> Look at this, guys, how cool. It

31:56 appeared here, in the agent’s output, a home desktop. If I click here, it’s exactly that yellow color, see? So, it’s possibly generating there in that cloud conversation the screenshot with the yellow buttons to put in the pull request. So, it works, right? You see that it works well. The big problem is that it’s very slow to do via voice. So , it’s not very useful to do this by voice; it won’t help you check the studio right now at 9:14 at night. Your GPT 6.1 vs Sonnet 5.5 video is fourth

32:26 out of 10 by views with 11.7 k in the first 7 hours and 29 minutes. >> Yeah, it got it right, it managed to go into the studio and get the information . But it took quite a while, it took, I don’t know, 4 minutes to manage to go, open the studio, see it, and get back to me, right? So, it really is quite slow, okay? It looks like it will work, so it’s just a matter of speed. I’ll even hang up here so we don’t waste too much time, but you saw more or less what the experience is like, all right? If it were faster, it would be better.

32:57 I believe that the moment they can bring a more interesting speed to this, we will have a much better workflow. Going back here to Grokbot, let’s try to do something similar in their audio mode, okay? Let’s go. Hi, how are you? Are you there? >> I’m here. Great. >> Great. Can you tell me how many classes were watched in total in the program? On Friday and Saturday? On Friday the 2nd, there were 84 distinct people with 313 plays and 55 completing at least

33:28 one class. On Saturday the 3rd, 50 unique people watched at least one, with 187 class starts in total and 35 completions. I want the ranking by class of the >> No, just answer what I asked you. How many people watched the class in total today? Searching for today’s people. >> Yeah, we have it here on the Grok interface, you can see it down here, right? >> This is just the beginning of the work. As soon as the numbers come in, I’ll pass them to you. >> Look, connecting to the post, we can see that it’s fetching information from

33:59 >> Today’s data is still partial. On that side, it’s still Sunday, October 4th. So, there’s no Monday number yet. So far, 35 distinct people started at least one class, 117 plays in total, and 19 finished, totaling 55 completions. Since the day isn’t over, these values can still go up. >> Okay, can you look up the pull request that’s open in the program’s repository for me? There >> I’ll pull the open PR, it’s in progress . Do you want me to detail any specific

34:30 part when it arrives? No, I just want you to tell me which PR is open. You see that she talks a lot more, she doesn’t leave much silence, does she? And even when there’s silence, we can see better here what Grokbot is doing, right? Now it seems it disappeared. >> The open PR was 56 Fix Green Primary Buttons, but actually it’s closed, not merged. It closed today around 3:13 PM Brasília time, still as a draft. >> But isn’t there any other open PR? >> Checking if there’s any other one.

35:00 Nothing new around here yet. Let me know if you want me to force a new look . Yes, there is. Number 57 Fix change website primary color to red. Opened tonight around 9:02 PM Brasília time. >> How did you say there wasn’t one and a second later you said there was? >> My bad, the rush got in the way. The result had just come in when I spoke. >> Right. Do this for me, you can close this pull request too. >> Closing 57. >> Hey guys, let’s see here. Grockbot will close it, it’s open. >> Just the beginning. When it confirms,

35:31 I’ll let you know. >> It’s what we created in that Cloud task , right? And it closed. It managed to operate. >> Done. 57 closed without a merge around 9:19 PM in Brasília. The repository is left with no open PRs. >> Wonderful, guys. So, what did I want to show you here? I wanted to show these two personal assistants working side by side, because their proposals are really very similar. I personally think Grockbot is a more mature product. OpenAI still needs to improve the UX a

36:02 bit. Now, in terms of voice mainly, OpenAI has a clearly superior model, but again, the UX is not that great yet . One thing I didn’t like much about Grockbot’s voice, which is much faster and more responsive, is that it gives too much information, right? That’s not even a voice issue, it’s an issue with the model itself. So you see I asked one thing, right? How many classes were completed? Then it said how many plays were made and that the day hasn’t ended , like, it couldn’t be as direct. I thought OpenAI’s model was more direct, okay? Now, regarding speed, it seemed

36:35 to me that Grockbot’s voice model is more interesting. And also for those who want to control what the agent is doing, seeing the connectors there, seeing the integrations happening, I think it’s really cool, okay? Who do I think this is for, guys? I think this is really cool for you to interact, as I said, with your product, to brainstorm, to cross-reference information. And then you cross-referenced information, saw something that could be improved, you immediately dispatch a task to the cloud, for example. This is something

37:06 you could already do with Grockbot and it was a very interesting workflow. But Grockbot will operate within the Cursor agent unless you use, you know, a way to dispatch it via VPS, as I did in my original Grockbot video. So here I have an engineer, I have a chat that is an engineer, which is basically a Codex that runs locally on my VPS machine at Hostgator, okay? So, when I wanted to dispatch a more advanced programming

37:38 task, I would do it locally via this engineer, which is my chat connected to the HostGator VPS. When it was a simpler task, I would dispatch it via cloud. Now, the advantage of being in the OpenAI ecosystem is that you can now, by talking to your DOT, also dispatch tasks from the Cloud Agent, but specifically from Codebase, or locally via Remote Control, as we were already doing, with the advantage of having a single interface. So, these products have become quite comparable.

38:10 OpenAI, in my view, still needs to catch up a bit to reach where Cursor is , which is natural since Cursor got there first, okay? They launched GrokBot earlier and also launched Cloud Agents earlier. But if OpenAI gets this right, I think it could be very interesting, especially because your DOT doesn’t consume your main limit. Of course, today it’s only on the paid plan, but if this happens on the free plan, I think it will be a very cool thing for those who want to use AI

38:41 without spending much of their limit. Alright? But now I want to know from you, which of these two personal assistants did you like more? GPT DOT or Grok Bot? Or maybe you would use Hermes instead? Tell me here in the comments. I want to know your opinion and the community’s opinion. Also tell me what use cases you see, what you would use these personal assistants for , and if it makes sense for your use case to use them via voice. I also want

39:12 to know what you would do in my place. If you liked this video, I kindly ask you to like it and share it with your friends who like personal assistants. Don’t forget that in the AI Program, my AI engineering platform, where we launched the AI Engineer track last week, we have an open waiting list, where one of the courses is precisely the personal assistant course. So, if you are interested in me adding a personal assistant course to the AI Program as well, go there and sign up.

39:43 It would be my top choice to do in 2027 . Alright? I thank the folks at HostGator for being a partner of this channel. Don’t forget, if you want a 24/7 VPS with support in Portuguese, go to HostGator. If you want to host your Hermes, go to HostGator as well, and it will be a success, alright? I am incredibly grateful to you for being with me until now, and see you next time.