Meet Google Gemini experiment 126, a model that might deserve the hype. Today, I'm running two big tests in Visual Studio Code, one using this model with reference code and one without. So let's see if this model is a big deal or just another failure. There are

rumors going around about the new model from Google. The Gemini experiment 126 is being called the Gemini 2 Flash. It's right now completely free, both on the Google AI Studio and AI Studio. But I found this other tweet from Tick Ki, I think that is his

name, saying this is not the Gemini Flash 2. When he used cursor, it registered as a premium model, which means it's the Gemini Pro model, not the Flash version. And after I tested it, I feel like it's more of a Flash version, not a Pro version, from

this brief that I got when coding with it. So if you head right now to the Llama Arena, AKA Chatbot Arena, and open the leaderboard and select the category of coding, you'll find that the first model right now is the Gemini experiment 1206.

And seeing a Gemini model at the top of the coding leaderboard is a bit weird for me, especially since Gemini models used to be bad at everything except mathematics. So I had to test it and confirm whether this model is actually good or not. So to set up this model inside

your Visual Studio Code, until now, the client and other composers aren't working great with it. The only option that's working 100% right now is Code GPT Chat. You can install it from the Extensions tab, select your AI, select the provider, select Google AI.

Studio, and there you will have multiple models—all of the Google models—you will find them there. Select the 126, and it will automatically connect you after you put your API key. You can get your AI key from AI Studio.google.com, and there you can get or create a new AI key.

You don't have to set up anything else, and it will work automatically on your Visual Studio Code. Before we start generating new code, this is the current project that I have. I created it a while ago using Klein and the .5, but it lacks a few pages; for example, the

category pages aren't working, and the About Us page isn't working. I just need to create these two pages and improve a few things inside this project to let me know if this model is good enough as a coding assistant or not. So let's get started. So I'm going to

start by creating a new page inside a route that I want. I named it `categories` for the first page, and I created this prompt for Code GPT:  "I want to create an HTML page to display all posts for different categories in a blog website using Tailwind CSS and GSAP, similar to the

style." The style I'm going to give it is inside this page. I'm going to copy and paste it, and don't worry about the amount of tokens; this model can handle 2 million token counts, which is impressive. I'm going to hit send and see

how it will handle it. The most annoying thing about this kind of Code GPT assistant is that it doesn't create files or folders or execute commands inside the terminal. Basically, it's not like VS Code or Klein, which kind of sucks, but it's very good when it comes to connecting to any

New model so far. Client or continue? Can't work with this model. I have tested both of them and also another composer, which is connected to another, didn't work also with this model. So now it's creating the fake data for this page and it's creating components, so it's not

creating everything single in one page. It's creating multiple different files or different components. Okay, it's creating even other pages like category details, the slug, and how to run it. And it did give me key improvements of what it did and what I need to do to enhance this kind of code.

This is really decent from Gemini, especially last week I did a video using it and it was awful, like really awful. This experiment, this is the result that I got from the model, which is honestly impressive. I put the code inside the files by myself, which is the only annoying

thing about it. But first of all, it did give me fake data to use inside a symbol file and it guided me where to put it. The second thing that it did, it created three different components: header and post card. Then in the end, it created

the page, aka the category page, and put all the components together inside this page. And the fake data used the navigation, and I just have to switch from using route to use param parameters because for some reason it didn't pick up on the what version of Next.js I'm using,

which is kind of a fault, but the total experiment so far is one of the best models from Gemini so far. It's not only after it's done giving me the code, it told me what I need to improve, but actually it went to give me a detailed

List for future development: what I need to do from one to 100. I'm not joking, this is like one prompt this model has been working on already, an existent project. What if I created a project from basically zero? Then the best thing to

use is this: it is either itself. So, I'm going to try this right now. So in a completely empty folder I have here, I created a new file called `env` and I put one thing inside it called `either-model`, and the name of the models I want

to use. Go ahead, put your API key over here and save this file. We're going to use AER with it, and if you don't have AER on your machine, you simply need to install Bison 3.1 and using Bison to install either.

Should you be using this command? I'm going to leave this in the comments down below so you can use it if you want. In your terminal, hit `either`. It will automatically work and it will ask for creating a new repository. I will hit

yes. It created a new repo. Now it gives me a warning about the Gemini experiment 126, which doesn't have any information like the cost or the Windows size of it. So basically, it will tell you to be careful with this model, but we're going to go ahead

and say no, I don't want to see the commendation, and it will start creating the project using it. So here I will give `either` this prompt: `Create a Next.js app using the App Router, not TypeScript, and build a modern e-commerce page with`

A clean design, dark mode, and creative UI, and it should run. Now it started to create this project's project structure and what it will look like, what the global style and dark mode will look like, they're with components, the header components, the theme toggle components.

Okay, it's creating two different themes here: dark and light. Product data, or it's planning everything at first, but I don't see anything being executed. I first have to create the project before I start anything, because right now it tells me I want to add this file to a folder that

doesn't exist. So I'm going to first create a new project for Next.js, and this is the result that I got. As I expected, if it didn't have some sort of reference to our quiz, it will start acting like this, which is extremely basic and ugly,

I have to say. But the planning and the setup is really neat, like creating the multiple components, both the components inside the beige... it's very good. I can use it in a different way by creating the main layout. It's using the HIO model or the Sonic 3.5, and the main theme

on the style. I asked Gemini to create components using it. It's not disappointing, to be honest; like the first result that I got was very good. I have to give it that. But when you give it a blank page and you want to build in

it something that is really good, it will start acting like this. But I can improve the Gemini prompt a little bit by giving it more direction, using what kind of images, using different icons, making the style in a certain way. But usually, Sonnet and Hau handled this for us, so at the end

Sonnet 3.5 is still the coding king, and just to double down on what I just said, go ahead to the live bench.  You will find exactly what I said. If you sort it by the coding average, you'll find that actually the Sonnet 3.5 is still the

king of coding, and after it, exactly, is the Gemini BMAN 1206, which is what I just said. This model is good at coding—extremely good at coding—when you give it a reference, basically a code snippet to use, or also in planning. I noticed in

planning it's extremely good. Also, the speed of generating code is insane, like compared to the other models that I tested from the Gemini experiment. This one will take the cake, and I'm pretty sure when this model came out in its real name—eka Gemini 2, Flash, or maybe the

Gemini 2 Pro—and integrated inside Klein, I think you will see its full potential unlocked, like the Sonnet 3.5 or the Claude model or the GPT-4. So for now, I will keep my eye on this model,

especially because it's free and it has very good performance in coding. So that's it. If you found this video respecting your time and providing you with valuable information, please hit the like and subscribe if you're not, and leave a silly comment

down below in the comments. Let me know what you think about this model. I'm really interested in this model because of the whole free thing—you can use it in an unlimited way, but you don't have to break the daily limits or the requests per minute—and the whole reference code

is really good. I'm going to use this model even more when it comes out as an officially named model. So thank you guys for watching and see you on the next video.

Is Gemini EXP-1206 a Sonnet 3.5 Killer?....The Winner is CLEAR (VS Code + Aider)

Codewello

Introduction to Gemini Experiment 126 and initial tests

The video introduces Google's Gemini Experiment 126, a coding model, and plans to test its capabilities with and without reference code.

Discussion on the model's name and potential confusion with other Gemini models (Gemini 2 Flash, Gemini Pro).

The presenter's initial impression suggests it's more like a 'Flash' version than a 'Pro' version.

Gemini Experiment 1206's top ranking in the Llama Arena coding leaderboard is highlighted as surprising.

The video explains how to set up Gemini Experiment 126 in Visual Studio Code using Code GPT Chat.

Testing Gemini 126 with an existing project

The presenter describes an existing project needing two new pages and improvements.

A prompt is created to generate an HTML page using Tailwind CSS and GSAP.

The model's ability to handle a large token count is mentioned.

A limitation is noted: the model doesn't create files, folders, or execute terminal commands.

The model's creation of multiple files and components, along with suggestions for improvement, is praised.

The presenter manually places the generated code, highlighting this as a minor drawback.

Testing Gemini 126 with a new project from scratch

The presenter tests the model's capability to create a project from scratch using `either`.

Setup instructions for using `either` with Gemini 126 are provided.

The process of creating a new Next.js app using `either` is shown.

The prompt for creating a Next.js e-commerce page with specific features is given.

The model's performance is less impressive when starting with a blank project, creating basic and less visually appealing output.

Comparison with other models and conclusions

Sonnet 3.5 is confirmed as still leading in coding benchmarks.

Gemini 126 excels with reference code and in project planning stages.

The presenter anticipates Gemini 126's full potential upon official release and integration with other tools.

A summary of the video's content and a call to action for viewers are given.

The presenter expresses enthusiasm for the model's free and unlimited usage, along with its strong performance with reference code.

Is Gemini EXP-1206 a Sonnet 3.5 Killer?....The Winner is CLEAR (VS Code + Aider)

Is Gemini EXP-1206 a Sonnet 3.5 Killer?....The Winner is CLEAR (VS Code + Aider)

Scripts