Everything You Need to Know About Gemini 3.0!

0:00Gemini 3 is here and you can now use one subscription for every task. They launched with two models, Gemini 3 Pro and Gemini 3 DeepThink. Of course, it's available in the Gemini app for everyday use, and then it's also available for developers in the AI studio, as well as their new developer platform called Antigravity. And it's also become available for enterprises. Gemini 3 performed well over almost all of the different benchmarks, including Humanity's last exam and VendingBench. Some of the improvements I've seen with Gemini 3 are its ability to do vibe coding and even agentic coding. It's increased multimodal understanding between images, videos, audios, text, and even code.

0:36And overall, I've seen the improved reasoning where it'll think deeper and understand the nuance in my questions. I've been trying to learn a little bit more about the basics of coding as I start to dive into vibe coding. And I asked ChatGPT and Gemini the same question. Create a five-step plan to learn how to code. And I thought Gemini's response was much better overall. While ChatGPT's response is still helpful, step three kind of tells you to build tiny projects as you learn and gives you some small tips. I thought Gemini's response was much more helpful as it walked me through a problem that I may face, which is kind of following tutorials but not actually starting to write.

1:06So it kind of gave me a project progression to avoid what it calls tutorial hell. And it gave me an easy three-step project progression, which I will definitely follow. Gemini also really excels over long horizon tasks. For example, on a benchmark that tests this, FendingBench 2, Gemini performs much better than any of the other models, including Claude Sonnet, Grok 4, and ChatGPT 5.1. Vending Bench 2 tests models' ability to stay coherent and successfully manage a simulated business over the course of a year using a vending machine as an example. It gives LLM the primary goal of maximizing profits while giving it control over things like pricing and the different items that are in the vending machine.

1:44I've been planning on going on a road trip along the west coast from San Diego all the way to Canada. This is a complex trip, involves multiple stops, and I want to do this trip over the course of a week. I thought that having Gemini 3 plan this trip would be a great way to test its abilities in reasoning and also multi-step tasks. So I gave it the prompt to plan a one-week road trip along the US West Coast from Mexico to Canada, planning all stops, including scenic or landmark stops, gas stations, food, and hotels. Let's see how it does. You can see here using the thinking model, it doesn't just start right away.

2:13It'll start by outlining the structure and defining the itinerary, and then break down the problem in multiple steps. Something I love about the response is it doesn't go overboard and give me too much information. it really lays out the information that I asked for in a neat way. And also during its research, it found that the Pacific Coast Highway is closed in the Big Sur region until 2026, so it rerouted the trip to accommodate for this. Overall, I love that it was to the point and understood the regions that I'm going to and didn't overload me with information. It even tells me some pretty specific details about how you must get the clam chowder at Splash Cafe in Pismo Beach.

2:46A few days ago, I made a video and I wanted to make a thumbnail using Gemini's help. With its increased multimode understanding, it's able to take in a video and understand its context like never before. So I'm going to bring in the video and ask it to help me find scenes of the video that would make for a good thumbnail, whether it's the reaction that I'm making or the context that is on the page. Just by giving it a quick prompt, it gave me 3 great options for a thumbnail. It was able to understand what was going on in the video at different timestamps, mentioning the anime race car driver at 5 minutes and 7 seconds or the guitarist on stage at 8 minutes and 30 seconds.

3:16Gemini was previously not able to understand video content like this. You're also able to just upload videos of you playing sports and Gemini will be able to output detailed instructions on what you're doing wrong and what you can work on, giving you a detailed plan to get better. Gemini is now able to create interactive visuals directly in the chat. As a guitarist, I've been wanting to learn a little bit more about different chord structures and how they sound different when compared to each other. So I told Gemini I want to learn about different chords in music from major to minor to major 7th to diminished and I wanted to use a piano as the main interface, walking me through understanding different chords and use sound to help me understand this as well.

3:52After selecting Canvas in the toolbar, I sent in this prompt and it created a chord laboratory that had an interactive piano, an audio engine, a chord selector, and visual theory as well, which can help me learn more about chord structure. Let's see what I came up with.

4:22This is really wild. It also gives you a theory and a list at the bottom, and a listening guide to help you better understand the chord structure. You can even ask Gemini to help you learn about complex topics by just typing into the chat interface or uploading complex PDFs. And it'll generate code for an interactive visual which then you can play with right inside of Gemini to further understand what you're trying to learn. For developers, you can now use Gemini 3 in third-party tools like Cursor, GitHub, or Replit, and in Google tools like AI Studio and their new agentic development platform called Antigravity.

4:51A great place to start exploring new Gemini models is Google's AI Studio. AI Studio is a browser-based development environment where you can build, test, and even deploy different AI-powered applications. I actually used AI Studio in a previous demo where I made an app that used NanoBanana, and I've noticed some crazy improvements since then. We can see some of the differences here. One of the main differences is Gemini 3's ability to create beautiful UI. In Google's AI studio, you want to make sure you head to build. In here, you can browse some apps that were made using Gemini 3. Compared to the app that I made using Gemini 2.5, these apps have insane design.

5:22Just looking at this page, you would think it would take thousands of dollars to get your website to look this good, but you're now able to make websites at this level in just one prompt. I asked Gemini to create a website for my vibe coding business where I wanted a white UI, soft shadows, minimalist typography, kind of matching that Swedish design aesthetic. And while what it created here may not be as insane as some of the examples, I'm still really impressed. Just by giving it that simple prompt, it created this full website filled with customer reviews, a curriculum with great animations, a daily vibe section where you can generate new tips, and overall I just feel like it is a pretty great UI.

5:56I definitely think I can prove it over a little bit more prompting, but being able to create a design forward website with one simple prompt is crazy. Another way you can use Gemini is to create games. Like this one, Tempo Strike even uses your camera to track your hands.

6:18This one, Chader Pilot actually uses sounds while you control some sort of plane.

6:33And the sounds are even interactive with whatever button you're clicking. So if I click the up button, it seems like the tone kind of rises. And if I click the down button, the tone kind of gets a little lower.

6:49Pretty insane. I asked Feminade to create a game where you can control a plane in a 3d environment. Let's check it out. So it gave me quick directions where W is throttle up, S is air brake, and then you can kind of control the pitch as well. It even gave me a mission briefing where I'm supposed to clear for takeoff and then navigate through the obstacle course.

7:12This game is really hard, almost impossible to play, but it is really cool that Gemini was able to create a game like this. You can see that this plane gets up to 2000 kilometers per hour, which makes it almost impossible to hit any of these achievements, but still pretty fun. While that game wasn't amazing, you can see the possibilities that are available with Gemini 3. And a trick that I can do using AI Studio with Gemini 3 is actually bring in a screenshot of a site that I really like and have it use that website as a template for my site. I use Whisperflow's website design, which is really cool to help me create a website of my own.

7:50And while I would really never just steal someone else's website, using it as inspiration for your own is a great way to use this tool. In my prompt, I asked Gemini to use the photo I uploaded for inspiration and create a website for my AI tutoring company. You can see that it basically took the typography and the general design from Whisperflow and just made it its own. While it doesn't look as good as the Whisperflow website, it really did a good job copying all the different typographies, colors, and just general design. One thing that's insane though is that it added an AI voice coach directly in the app without me even asking.

8:21Let's check it out. So I click the button and just ask, how can I get better using AI? That's a fascinating topic. What particular applications of AI are you most interested in exploring? I would love to learn a little bit more about prompt engineering. Great. Prompt engineering is a really important skill with AI. Are you working with text-based models or something else? That's so insane. It even seems like you train the model that I'm talking to to understand that it's an AI tutor. And as a developer, you can even take some of this code and implement it into your own projects. And Google also just released its Segentic developer platform, which I'll dive into deeper in a different video.

9:01Thanks for tuning in to This Week in AI and I'll see you next time.