Transcription
Google has rolled out several updates. I've highlighted the most interesting ones from my perspective and will tell you what you need to know right now.
First, NotebookLM finally has point-based presentation editing. Second, the Gemini core has been updated to version 3.1 Pro. Developers have massively improved its logic and capabilities. I've tested this on my own tasks and will show you what came of it. And the third new feature, which few have heard of, is that you can now generate unique music directly in the Gemini chat with just one request. We will analyze all these updates in detail now on screen, and at the very end, I will show you my personal method for teaching NotebookLM to generate powerful prompts for itself. Watch until the end, it might be useful for you. Let's go!
Let's start with the Gemini update. As you can see, the post about this was on February 19th, and let's see what has changed here. We'll focus on this table and try to understand it. I'm not a super AI engineer, so to understand this table, I asked Gemini itself to help me and highlight the most important points.
So, one of the main points is ARC-AGI 2. This means how well the neural network understands complex relationships and solves non-standard problems. And just look, this value has almost doubled. This is a test of so-called abstract logic, whatever that means.
The second parameter that is significant for us is Apex Agents. Here too, the increase is almost twofold. This parameter is responsible for performing long professional tasks. This means that when I feed it a huge chunk of data and ask it to write a long script, for example, for a new YouTube video, the model will no longer lose its train of thought and will keep our goal from beginning to end.
And the third important parameter is GPQA Diamond. Don't be fooled by the fact that the increase here isn't that big. The thing is, this parameter is responsible for precise scientific expert knowledge, and therefore, even such an insignificant increase is still a pretty cool result at the current time.
And to avoid being unsubstantiated, let me show you how I decided to test this. To be able to switch between models, the easiest way is to open Google AI Studio. And here we can change the models. So, instead of Gemini, it's Gemini 3.1 Pro. The thing is, when we are in the chat, we cannot switch between models. That is, if we choose the Pro mode, it will be 3.1 by default. And while in Google AI Studio, we can choose this model.
And I asked Gemini to come up with a prompt for me. The complexity of this prompt was in the correct display of text. I'm not some programmer, I won't be comparing algorithms or anything like that here. I decided to compare how one model and the other would handle it using my examples, like writing a script.
And I took the same prompt that Gemini helped me write, and I inserted it, looked at the response, what it would generate for me. Then I changed the model, inserted the same prompt, and generated the response again. I copied these two responses, pasted them back into Gemini, and asked it to compare them and see what the differences were, which model performed best.
To be honest with you, for my tasks, there is absolutely no difference. The only moment we caught was that here, when comparing neural networks for video generation, there was simply an error in the facts. That is, 3.0 told me that there is no built-in audio in Veo, no integration with Audio FX 3, that is, it simply stated that there isn't, and that's it. But in version 3.1, in conjunction with Audio FX, that's the truth. So, 3.0 made a factual error, and everything else is identical.
Whether to consider this a big leap, I don't know. Again, I repeat, for my tasks, there is absolutely no difference. In one of my videos, I explained this using the example of cameras in mobile phones. When these cameras first appeared, everyone boasted about who had four megapixels, then who had eight, then twelve, then twenty-five, and then someone was photographing the moon, and so on.
And now the time has come when everyone really doesn't care how many megapixels their phone has. For example, I honestly don't even know how many megapixels my camera has. I have an old iPhone 11, it takes pictures, and I'm happy with it. And there are many such users. It's the same with the power of all these models for ordinary tasks like writing scripts, text processing, and some simple logical tasks. Even a ten percent increase, as you can see, is not felt.
And in general, on the internet, specialists and programmers compare these models on a slightly different level. Their main attention is now concentrated on pure programming. For example, in tests like Geekbench, the new version scores an astronomical eighty percent and is closely approaching Claude Opus 4.6.
And everyone on the forums is actively discussing so-called "vibe coding," which is when neural networks understand not just dry syntax, but the design concept itself, the style of the future application, plus strict technological requirements. They come to the conclusion that Google has greatly improved the hidden limits. Now the model has three levels of reasoning instead of two, and the new middle level thinks as powerfully as the highest level in the old version. The file upload limit in the API has been increased fivefold to 100 megabytes, and direct support for YouTube links has been added. And the maximum response generation volume has increased to 65,000 tokens.
And most importantly, why everyone is so happy: the price for developers has remained unchanged. Well, let's be honest, I don't use Gemini as much as these hardcore programmers. For me, for the most part, all these giant volumes of code and vibe coding make no difference.
Now let's move on to what many NotebookLM users have been waiting for for a long time: presentation editing has finally been added. This means we can now edit each slide individually. Before, even if we didn't like a simple element, like the font color, we had to redo the entire presentation. But now we can redo only one slide and only that one element.
To demonstrate this, I'll show you on my notebooks. Let's go into my most frequently visited notebook in Anti-gravity, where I learned to use this tool. And here I have different presentations. I've honestly tried it here, but let me show you how it works. For example, we go into a presentation, say this one, select the first slide, and for example, on this slide, I don't like the font color. It was white here initially. I asked it, let's try to go back. Here, a "change" button has appeared. We click it and can dictate what we want to change: "Change the font color in the slide title to white."
And here, pay attention, if I click "Create Presentation," it will start re-drawing the entire presentation. That is, it will all be unavailable. And if I need to make changes to several slides, these changes need to be dictated like this first. So, look, I can switch to the next slide and also enter some change here: "Create a different design for the flowchart on the main slide on the left side." And as you can see, they are highlighted in blue. So, we make all the changes, dictate our own changes to each slide, and then click "Create Presentation."
Clicking it, as you can see, a new version is being generated, and we can no longer access it. So, don't generate one by one; dictate all of them and then click "Generate." So, the presentation is ready. I don't know, it took about a minute. Let's go in and see what we have here. So, we asked for the font color to be white. Done. By the way, it would be convenient if it wrote somewhere what changes were made. There are complex presentations sometimes, it would be convenient to have some history of all changes. Well, okay. And here on the third slide, yes, we asked, yes, here it is, the MD file, we asked for a different flowchart. Well, super, in the same style, another one. Well, as you can see, it coped quite quickly, not bad, in the same style. So, I think if it redraws, changes the style, and changes the color, then I think the problem with word errors is already solved. You can just rewrite, redo the slide, and it's done.
And another interesting feature regarding presentations is that before, we could only download in PDF format. Now, as you can see, you can download in PowerPoint format. I don't know who might find it useful to download in presentation format. It doesn't offer any advantages. This entire slide is just in image format. If it could break it down into layers, elements, and extract the text separately, that would be great. But as it is, well, it exists, and that's that.
And so, a new function in our tools is music generation. I don't know if anything can be played here, like copyright issues or something. But let's just describe our track and create our new one. I've prepared this request. Again, you understand, I didn't write all of this. I asked Gemini, "Create a prompt for me to generate music." That's all. The other details, as I understood them, are tempo, mood, instrument description, duration. Thirty seconds is the maximum, as far as I understand. I also specified "without lyrics" because I know that if you don't ask, it will insert its own words. And studio sound quality.
We click "Generate." So, it generated. Not on the first try. You know what, I removed "thirty seconds." Perhaps that was the mistake, or maybe it just so happened that I removed it and it managed to generate. Well, let's listen. Overall, it's quite good, in my opinion. And the coolest thing is that you don't have to worry about copyright at all. That is, we can play it at full volume, and as you can see, no one will knock on our door for it.
There's also a cool feature here that they wrote about in the specifications: you can insert a picture, and Gemini will recognize the mood of the picture, what is depicted in it, and can pick music for it. Well, let's try it. I found this picture with a beach and the sea. Well, I don't know if I need to write a description or if it will guess itself. Let's do it in a few words: "Create a track that fits this picture." "Oceanic high-fi." Even with such a title. I click, enjoy the atmosphere. Well, let's try. It even generated a picture, but we can look at this one. Does it fit? Yes, it fits. By the way, if we want to download it, we can download it in two formats: video, that is, an audio file with a cover like this, and only an audio file in MP3 format. Cool, cool, quite convenient. By the way, the track is thirty seconds. Why didn't it want to generate when I initially asked for thirty seconds? I don't know.
And for the sake of interest, look at this picture I found. What will it sing to me about it? Let's just send it without a description, like this. "Track: Carp Hip." I can't even read it. Based on the atmosphere. Wow, a cool picture. Let's listen. Uh-huh. So, this is how it is. If the picture is with scrap metal, the music fits. According to Gemini, this is what it is. Well, it plays not metal house. Uh-huh, I understand. So, metal house.
"Keep the same track, but in Russian. Among the mountains of metal and rust, I found my path. From a pile of junk, I create; in this lies my essence. We collect scrap metal." So, I think if someone, like me, has never used tools like Sora and all these music generation tools until now, now you don't need to go anywhere. Everything is right in our chat. Be careful, it's an addictive thing, you can get lost for three hours.
And now, the promised life hack from the beginning: working with NotebookLM, that is, how to make it generate cool prompts, and then you can use these prompts either right here in NotebookLM or in your other tasks. I saw this on YouTube, and I'll show you briefly now. I've used it a couple of times, and it works quite well.
Let's create a new notebook. Let's name it "Prompt Generator." Now, let's just open a new tab and enter a query related to how to create prompts correctly. Something like: "Ten best practices for prompts," or "Best prompts," or "Best solutions," or "How to write prompts better." It's clear that it's best to write this in English, as there will be more sources and they will be of higher quality.
And open all the sites that appear here one by one. In a new tab, open in a new tab, open in a new tab. You understand, right? So, you see a resource from OpenAI, like a help center. You see from Claude. And I have this extension. It's called "NotebookLM Web Importer." I click on it. This is the prompt generator. I select the notebook I want to add to, and I click "Add to Notebook." That is, I don't even need to delve deeply into reading what it's about. I see that these are some tips on how to write prompts. I click "Add to Notebook," and that's it.
So, let's go back to our notebook, refresh the page, and here are our sources. You understand, right, that you can insert not four sources here, but fifty, for example. All the articles you can find, and several videos from YouTube on how to create prompts. And now we have our own database that tells us and tells the neural network how to create cool prompts.
And here we simply ask for prompts for some other tasks. We need a prompt for working in some other notebook, and we collect a separate notebook for prompt generation. Similarly, for example, if you want to write instructions in some notebook on how to do design correctly, you can create a separate notebook with some design rules, generate these rules, and then use them separately from this notebook.
To summarize today's analysis, Google has stopped making just a smart chat. Or rather, it is, but it's almost imperceptible now. But what we do feel and see is the development of the same ecosystem. It's clear that work is being done on NotebookLM. The tools now have a new music generation function. It's all part of the same trend: the one who wins is the one with an ecosystem, not just a pretty chat. At least, Google is pushing for this.
I won't give you the prompts I used during this video today. I don't think anyone needs them. But I did generate a cool table with a description of all possible combinations when composing a prompt for music generation. So, if you don't want to dig into it, just go to my Telegram channel, take this file, copy the styles, instruments, and quality from there, paste it into your prompt, and have a disco. No copyright issues, you can generate, upload to YouTube, wherever you want. In short, go ahead, take it, use it. The link is in the description.
Well, if this analysis helped you and if you want more of this kind of concrete information without fluff, then subscribe to the channel. It's very important to me. Also, there are still a lot of views without subscriptions. Subscribe, support, like, comment. And if you want to support me more than with a comment, all the tools are also in the description. This will greatly motivate me and help me. Thank you for watching. Value yourself, value your experience, and goodbye.