CONTEÚDO

Vale a pena? Suscripciones DESTROZANDO GRATIS está

There's no cage more comfortable than one that convinces you you'll be safer inside. The main artificial intelligence tools seem to sow more fear than tranquility.

👀 Confira só, clique aqui!

Destaques

• • This is Paraquet version 3, and it works quite well.
• • IC analyzed over 170,000 MCP tools and found that action tools saw their usage jump from 27% to 65%.
• • In fact, it already tells me here that it's running, meaning it took about 5 seconds.
• • To get yours, I'll leave you this link here where I recommend the KVM2 plan.
• • I just pasted it here, and the text already tells me I have over 300 models available.

Fear of reaching their limits, fear of our privacy, or fear of their costs. They're getting more and more expensive. They say it's a bargain, and subscriptions exceeding three figures are becoming the norm.

However, in the shadows and without advertising, there are alternatives capable of doing everything. All the tools we'll be looking at have something in common. They seek to escape dependence on Big Tech, and what they offer can be used completely free of charge, privately, and without limits.

Well, let's look at the context. Anthropic, one of the leading artificial intelligence companies, called for a potential global pause to halt the advancement of AI, but at the same time launched Cloud Fable 5, based on myths, a model they claimed was banned.

Yaki suggests that many of the warnings are merely marketing ploys to increase the reach of their new products. However, Stanford also states that the gap between the United States and China has practically closed, shrinking to just a 2.7% advantage this year.

But it's important to note that the real battle isn't whether the US or China's cloud infrastructure is better, but rather working with a closed cloud versus an open and flexible one.

Nearly 40% of interactions with AI tools already contain sensitive information, and open models are currently only about four months behind closed models by 2026.

Closed models cost an average of six times more than open models, meaning that shifting demand to open models could result in savings of $25 billion annually. And to discover which useful tools we can use without being tied down, let's move on to the first use case.

Here we'll finally be talking about one of the tools I've been using that you've asked me about the most. It's a platform that was created as a free, local, and open- source alternative for dictating text without paying for subscriptions.

This is an alternative to well-known tools like Whisperflow, except that its creator, CJ País, made it after breaking a finger and being left without any simple, free, and private way to efficiently transcribe by voice.

To see what it's capable of, let's jump to Handy, the platform I'll be describing below. Once you enter the platform, you'll see that it makes its purpose very clear: it's a platform for speaking within any text field.

But unlike other tools, this one is free and open-source, so we can transcribe anything. Also, keep in mind that despite being open source, its installation is just like any official application. So here it would be as easy as clicking on the one that corresponds to our device.

Once we have it downloaded, we would click on it, drag it to install, and when we open it, it will go through an installation process that I didn't have to do, since I've actually been using it for a while now.

But keep in mind that we simply had to give it access to our microphone, as well as accessibility so that it can type in any text field what we're going to transcribe.

And finally, it will recommend installing any transcription model, since this tool can work without Wi-Fi, as those artificial intelligence models will be installed locally.

Regarding the models, You'll find a lot of alternatives, but since you're probably watching this video in a language other than English, I recommend you select your language directly. In my case, it would be Spanish, so this will show me all the models that are suitable for that language.

Another aspect I recommend you take a look at are these bars here. The fuller the first one is, the greater the accuracy, and the second one indicates the faster transcription speed . I actually recommend the one I've been using, as it's the one recommended by the platform itself.

This is Paraquet version 3, and it works quite well. Once you've selected your model and granted the permissions it requests initially, you can start using it.

I'll share a couple of tips and tricks for this tool now, but first, regarding the general settings, keep in mind that you can change the key you use for transcriptions here .

In this case, I'm using the French C key, which you can change, since it was a key I never used before, and now I'm putting it to good use with this tool.

Once I have this set up, when I open any tool, like this one here, if I hold down that key I've configured, the animation you're seeing now will appear. And when I release the key, you'll see how it transcribes everything perfectly. In fact, I just released it here; we're not going to cut anything.

And you can see that it transcribed everything perfectly and in record time. These kinds of tools are really key, because AI tools like GPT chat or other models often perform better when given text input, but that input needs as much context as possible to work properly.

And this type of tool really simplifies the whole process. Furthermore, if we return to the platform, I want to highlight a couple of features.

These are words that used to give me trouble, but now, simply by adding them here—and keep in mind that to add them, I would just have to type any word here; I'll type "prueba"—it would be added here, and I wouldn't have to do anything else.

I, however, am going to remove the test version and now I'm going to test these three words I have saved. Even though no other words have capital letters in the middle, the dictionary would still learn it.

And if we go back to the beginning, we also see how both "Alejabi" and "Artificial Academy" would be perfectly spelled. With this, we can say goodbye to wasting time typing and on that proofreading that is sometimes necessary when we have to transcribe something for artificial intelligence models.

And keep in mind Typing on a keyboard, we can get about 45 words per minute, but with this transcription tool, we can exceed 200 words, making us significantly more efficient, not only in AI output but also in the input we provide.

Keep in mind that Google is heavily investing in open-source, free, unlimited, and proprietary software. And they've also released their own transcription product. It 's called Eloquent.

If we jump to their platform, which I'll also link to in the description below, you'll find a tool that's no longer just available for computers ( in this case, only for Macs), but also for iPhones. It even has a slightly more user-friendly interface.

I have it installed on my device too, but I have to mention the downside: unless you speak English , this tool, despite its capabilities, currently only allows you to use it in English. IC analyzed over 170,000 MCP tools and found that action tools saw their usage jump from 27% to 65%.

This surge is also driving major tech companies to migrate their chatbots, such as GPT chatbots, to agent platforms like Codex. Odissius presents itself as a workspace similar to GPT or cloud chatbots, offering chats, agents, MCP, research, memory, email, tasks, and more.

However, unlike the previous examples, it's free, unlimited, and completely private, ensuring that data remains between you and your computer. To learn more, let's dive into Odissius, the platform we're discussing.

And if you're very active on the channel, you probably already know them inside and out, since I already did a detailed tutorial on this tool in the video before this one .

But I didn't skip this part because I'm going to discuss some key points that no one talks about, not even me in the previous video, and that you'll really find interesting. But there's an easy solution. To find that solution, we need Docker.

This allows you to run an application with everything it needs inside an isolated package so that it works just like any computer or even a server. It also makes installation on any paid server much easier .

To see this, I've just switched to my VPS, that is, my own private server, where I have Docker installed. And notice that from here, if I click on Docker Manager and go to the catalog section, with just one click I'll have a multitude of tools that I can run privately 24/7.

In fact, to show you which ones I have installed, if I go to the projects section, you'll see that I have different agents like Hermes, Disius, Lama, and other tools.

Going back to the catalog, let's look at its installation, since here it's as easy as typing the tool we want to install, in this case, Odisseus. Then, simply clicking "select" and giving it a username, which I'll call Alejabi, and a password to remember it.

And now, if I click "deploy" without even having to open the terminal as we saw in the previous tutorial, all the configuration will be done automatically in just a couple of clicks. In fact, it already tells me here that it's running, meaning it took about 5 seconds.

And if I click "open" from that new server, the Odisseus platform would appear here. Then I just had to enter the username I mentioned earlier, in this case, Aleji. And now I would enter that as the password. The password I had written.

And when I click on sign in, I'll have the Odisseus tool here , which isn't installed on my own computer, but on a private server, which will be active 24/7 so I can now automate anything without having to depend on my computer.

Notice that the URL I would have from here would be different, since it would actually be in a private cloud. To get yours, I'll leave you this link here where I recommend the KVM2 plan.

And when you choose it, if there's a coupon where it says, click on it. Thanks to the fact that the hosting is only sponsoring this video, if you enter the code Alejabi when you click on apply, it will be even cheaper. Keep in mind that the longer the period, the better the price.

And now, if I go to the settings section here, I can use it via, for example, Open Router, which we can use for free. For this, I just created an API key that I'm going to copy. I just pasted it here, and the text already tells me I have over 300 models available.

By adding it, I'll have virtually all the existing models directly from here. Among them is the latest Fable cloud model.

This one does have a cost, but if I type the word "free," I'll find a bunch of completely free models. So, I'm going to select one of them. And if I ask it anything from here, like, "Hey, tell me, what could the Odysseus genetic platform do for me?" How many things could it automate?

It would already be perfectly transcribed here with that handy tool we saw earlier. And if I click "send," it would start processing it at this incredibly fast speed you're seeing.

And we can see all the things we'll be able to do from this platform without any limitations, while still maintaining control of our data and privacy. With this, we're no longer talking about free alternatives, but about building your own artificial intelligence command center on Odysseus.

I already have a complete tutorial, which is the one I recommended earlier. So, now that we know this, let's move on to the next section. As we mentioned at the beginning of this video, Google is really pushing the boundaries in every aspect, including all its open-source models.

In fact, they released Yenma 4, with 12 billion parameters, that runs on laptops with around 16 GB of V RAM or unified memory, offering vision, audio, and reasoning capabilities .

The key point isn't that Will has released a larger model that can run on mid-range laptops, but rather that it's bringing multimodal and agentic AI to the average computer. With this, we can say unequivocally that paid AI isn't going to disappear, but it will certainly cease to be mandatory.

By learning a bit more about this model, we can see how, with these 12 billion parameters, we can surpass the previous YMA 3 model with 27 billion parameters, which could only be run on a very powerful machine. But this one is usable by anyone.

In fact, to see how easy this is, let's jump to Olama, the platform that allows us to install any open-source model in the easiest way possible. But in this case, we'll go a bit faster, since I already discussed it in detail in this video. So, we'll focus on this new Yemma model.

To do this, if I click on the models section, notice that when I type the word "Yemma," all the Google models will appear. Among them, we have this latest version 4. And from here, scrolling down, look what we find: all the variants are just a click away.

The smaller or higher this number is, the smaller the model we'll find; that is, the model we can run on a less powerful device, but obviously, the model won't be as powerful as one with more parameters. This is where I'll be focusing.

This one has 12 billion parameters, which is the one they just released. And keep in mind that, luckily, we have it optimized for both Windows devices and Apple Silicon hardware.

For this, note that we have the MLX models here, which simply refers to all those Mac computers with processors from M1 to M5. In my case, it tells me that it was already downloaded successfully, since I downloaded it previously after its release.

So now, from the Dolama application, if I select the model we just chose and input something, like "hello," we'll see how, despite being a very large model, it can actually start reasoning quite quickly. In fact, it would already have that response in less than 10 seconds.

And if I ask it something now, like what it can do for me, when I click "send," it will automatically start working and reasoning about that answer without any waiting. And a few seconds later, I can find all those blocks that tell me everything the model could do for me.

But in addition to this, the most interesting thing is that we can integrate it with all the tools we've seen previously, such as Odisius.

And from here, look how I'll now have that model available, since the video that Odisius recommended showed how we can link Olama and Odisius, and all the new models we add will automatically appear. I already have 12 billion parameters in SD here , and now I'll be able to work from here.

For example, if I ask the same question now: "Hey, could you tell me what Odysseius can do for me?" regarding process automation, as soon as I click "send," it will start reasoning almost automatically. And look, I'm recording and I have a ton of folders and tabs open.

This diffusion process generates refined text simultaneously, enabling it to generate up to 1000 tokens or more per second when running on consumer graphics cards.

This allows it to generate text four times faster, achieving not only significantly faster speeds compared to other models, but also improved output quality, which is comparable to those more standard models that take longer to provide results.

In this final section, we'll move beyond the text-based aspects to see how everything related to creative work can also be done this way. To do this, we need to explore Ideo Gran's new model, as he has just launched his first open-source foundation model with 9.3 billion parameters.

The most interesting aspect of this model is that it's not just another image generator; it allows for real design, legible context, layouts, colors, and control by Jason.

And since its launch, Idograman has never focused on creating images per se, but rather on professionally designing any concept. To see this, let's jump to Idegran, the platform you'll also find in the description below.

And from there, once we land, we'll encounter a new The interface is noticeably improved, especially compared to those Ideogran tutorials we discussed over two years ago. The most important thing here is that Ideogran 4.0 allows us to create images with up to 2K resolution, and without limits.

Keep in mind that using the official platform does have a limit, as it requires a certain number of credits that are renewed daily. I'll explain how to install it locally later.

The first examples shown here are use cases like posters, and we can see several examples—very professional, and it's hard to believe they're made with artificial intelligence.

Similarly, we have other examples such as designs for t-shirts, logos, and more, as Ideogran truly excels at integrating letters into advanced creative concepts.

In fact, for a quick example, I've come across this design, which I find very interesting, where it plays with letters that appear to be submerging into a sea. Then I'll send it, and it will start processing here.

Notice here that it's used four credits, since each image would actually cost one credit. I 'm generating four at once, which is why it has that cost. Keep in mind that from the official platform, without installing anything, you can create up to 12 images per day.

As I mentioned, this has already been generated. So, I would have opened the one I think I like a bit more here. And look here at how we would have those letters as if someone had actually written them, and the J would be melting into the sea.

Let's see how we can now use it without limits, and for that we have a platform called Confi UI, which I already discussed in detail in this video here, so I'll go a little faster.

So, when you open the platform, if you click on the templates section, notice that by going to the image section, you'll find this IDE version 4 template. If you click on it, an optimized workflow will appear so you can modify the program you want it to run, even in JSON format for complete detail.

And from here, keep in mind that once you enter, Confi UI will directly show you which template you need to download and install; simply click the button, and it will do everything automatically. We could do the same thing with video format.

As you can see, so far in this video, we've seen models created by the United States, like Google, and others from Europe, like Dicius, created by the creator of PewDiePie. And now we can also find Asian models like Bernini.

So, as we're seeing, this is n't really a war between countries, but rather about finding the best alternative for us .

For example, we could upload a video like this one, where there's an empty temple, and tell the program to add a couple of people directly to the video with a slightly more golden filter. And look at how good the result would be!

Similarly, we could upload another video of a city and tell it to appear in winter with snow, and the result would be applied perfectly. Likewise, we could apply different background changes to any video or even change the The entire perspective of a video we've previously generated is enhanced.

The most interesting aspect is that we can also work with multiple inputs simultaneously. For example, we could have this reference image and a video showing other flowers, and upon generating the video, it will automatically replace the flowers.

Similarly, we could have a photo of this helicopter and a video of a landscape, and ask it to integrate them naturally to achieve the desired result. We can also insert anything into a video.

For example, we could add this image or another reference, like this video here showing several Chinese characters, and upon generating the video, we would have that image perfectly integrated.

We could do the same with more elements, using two videos, for example, here with the Bernini logo in motion, and another video here showing a kind of film set in an auditorium. And notice how it integrates both videos perfectly, making it appear as if this were the original result.

This model can truly work with any type of input. In fact, we can upload this image of the statue, another image of some headphones, another of a Bernini t-shirt, another of a floral swimsuit, and a landscape with a bench and a beach in the background.

Well, we'd have this video you see here, with everything integrated. With this, we're no longer just talking about creating videos; we're really looking at a kind of nano-banana where we can include different sections and objects we want in a video, and it will take everything into account.

Similarly , elx also launched its own desktop version. It's called LTX Desktop and can run on Windows, Linux, and Mac, although if you want to run the models locally, you'll need a minimum of 16 GB of VRAM. That way, you can use everything locally, although you can also experiment with the APIs.

The platform itself is this one here, called elx Desktop. A completely local artificial intelligence tool that doesn't run in the cloud, is completely free, and everything can remain private. Once we download it to our computer, we'll see an interface like this.

When we start a project—in this case, I'll call it a tutorial— we click "create," and from here we can create any type of video. For example, I'll create a video of an older man doing extreme mountaineering in the snow. Notice that, since it's transcribed, I could add up to 4K quality.

I could also use up to 50 frames per second and select a duration of up to 20 seconds. However, I'm going to leave it with these settings, and if I click "generate," it will start processing.

A little later, I'll have this video generated, which, by the way, also has sound effects, since this LTX model can generate videos with audio.

But the most interesting thing about being able to generate high-quality videos, whether from an image, audio, or a program, is this video editor section, where we can experiment with all the AI-generated video effects.

Here, in fact, it would already appear in the timeline, and from here I'd be able to work with different clips as if I were an editor, but one with artificial intelligence. We could fill in the gaps with other videos that it reinterprets based on the concept of the video we're working on.

And here, as we've seen, some tools like Handy, Odissius, and others could have been just another one on a closed platform with limits on usage and monetization, but it's not about being just another one; it's about doing more than just one.

Fonte: Assistir vídeo original
Baseado na transcrição original