Google AI Studio Masterclass: A Comprehensive Tutorial by Humanitarians AI Student Fellows
A student fellow tutorial covers Google AI Studio's core features, from code optimization and video timestamp analysis to fine-tuning Gemini on a custom dataset.
Plenty of AI platforms promise to do everything, but Google AI Studio's actual value shows up in a specific set of concrete tasks: writing from an image prompt, solving a genuinely complex math problem, pulling timestamped answers out of an uploaded video, and fine-tuning a model on a custom dataset without writing training code by hand. This tutorial, built by a Humanitarians AI student fellow as part of the program's learn-by-teaching approach, walks through each of those in turn.
What sets Google AI Studio apart
Google AI Studio is described as an advanced platform for exploring, experimenting with, and building AI-driven applications, built on top of Gemini, Google's own model environment. That distinguishes it from tools like Claude or Perplexity in a specific way: Google AI Studio isn't just a chat interface, it's positioned as a full environment for building, training, and developing models, including an AutoML option that lets someone fine-tune a model on their own data without needing to write code.
Core features at a glance
The platform's key features include code optimization, where a user can input a code snippet and get analysis of its performance, complexity, and potential bottlenecks. It also handles video analysis and Q&A, letting a user upload a video and ask questions about specific moments, with the platform returning answers tied to actual timestamps in the footage. Pattern recognition is another feature, where the platform can take a sequence and predict what comes next. And model tuning and experimentation round out the core feature set, letting users adapt Gemini's underlying model to their own specific use case.
Writing from an image
The first practical demonstration uses a simple, concrete prompt: write a short, engaging blog post based on a provided photo of a meal, including a description of what's in the image and a discussion of a personal meal-prepping journey. Given just the image and that instruction, the platform produces a full draft blog post, which doesn't have to be used exactly as generated, it can be reviewed, reformatted, and customized further, but it removes the burden of writing an entire post from a blank page.
Solving genuinely complex math
Google AI Studio also handles more demanding mathematical tasks, demonstrated here with a request to derive the equation of two lines, L1 and L2, from a fairly complex setup. The platform doesn't just return a final answer, it provides a detailed walkthrough of how it reached that answer, and the same capability extends to other problem types like trigonometry.
Video analysis with real timestamps
One of the more distinctive features shown is the ability to upload a video, including a tutorial video, and ask the platform questions about specific content within it. In one example, a StatQuest video was uploaded with a request to explain reinforcement learning, and the platform returned an answer tied to a specific timestamp, along with a breakdown of which time ranges in the video covered decision trees versus reinforcement learning specifically. That same capability was demonstrated on a Vercel tutorial video already posted on Bear Brown & Co., where the platform broke down which minute ranges covered which topics, letting a viewer skim a long tutorial rather than watching the entire thing to find a specific answer.
Generating more content
Beyond a single blog post, the platform was also used to generate three separate food-themed blog posts complete with recipes, along with suggested image prompts that could be handed off to an image generation tool like Stable Diffusion, using a detailed scene description, for example an overhead shot of a baking sheet filled with roasted chicken. The quality of the resulting image prompt, and by extension the generated image, depends directly on how specific and detailed the original description is.
Fine-tuning Gemini on a custom dataset
The tune model option lets a user adapt Gemini's model to a narrower, more specific purpose than its general capabilities. The example used here involves sentiment analysis: Gemini can already classify a simple sentence as neutral, positive, or negative reliably, but a sentence like "the car was running at 2500 RPM" is genuinely ambiguous in sentiment terms. For a use case built specifically around understanding car-related statements, a user can build a small CSV dataset, for example labeling "this product is amazing" as positive, "it was an average experience" as neutral, and "the service was horrible" as negative, import that dataset through the data source option, and tune a Gemini 1.5 model against it. During training, a visible loss value drops as the model learns from each data point, starting high and decreasing as understanding improves, and once tuning completes, the resulting model becomes available directly in a chat interface for testing.
Key takeaways
- Google AI Studio is built on Gemini and functions as a full environment for building, training, and tuning models, including a no-code AutoML option.
- Core features include code optimization analysis, video analysis with timestamped answers, pattern recognition, and model tuning.
- A detailed image-based prompt can generate a full blog post draft, and the platform can also produce suggested image prompts for tools like Stable Diffusion.
- The video analysis feature lets a user upload any video, including their own tutorials, and get answers tied to specific timestamps within the footage.
- Fine-tuning Gemini 1.5 on a small custom CSV dataset, such as sentiment-labeled sentences, lets the model handle ambiguous cases outside its general-purpose training.
Who this is for
This masterclass is aimed at students, educators, and anyone new to Google AI Studio who wants a practical walkthrough rather than abstract feature descriptions. It's part of the Humanitarians AI Fellows Program, where student fellows deepen their own understanding of AI tools by teaching them to others.
Full transcript(auto-generated, with timestamps)
[0:03]Hi all, this tutorial is about Google AI studio. So what is Google AI studio? Google AI studio is an advanced platform which is designed to explore, experiment and build AIdriven applications. Now you might ask what the difference between Google AI or tragic or perplexity or claude is. So Google AI is based off of Gemini which is again Google's environment. It is a whole Google environment and most importantly Google AI is used to build, train and develop models and it it provides a AutoML model where you don't need to know how to code. All developers can just give their data sets and tune fine-tune models and with the
[0:57]Help of Gemini or Google AI it can enable and fine-tune and you can tune those models to represent or give your answers and outputs exactly. So the first thing first let's go through the key features of Google AI studio. It helps with code optimization. Like I said, the allows users to input code snippets and analyze their performance. It provides insights into complexity, bottlenecks and areas of optimization. You can also do video analysis and questionnaire. that is you can upload any tutorial, any video and based off of the time and the information that is provided on in that video you can ask questions and Google AI will provide you with an answer for
[1:46]It. It can also do pattern recognition. Um it you can give it sequence patterns and it can predict what the next or the next pattern or sequence is. And like I said, you can do model tuning and experimentation. Now, how to set up Google AI? You can, as any other website, you can log in, you can sign up. Uh I've chosen to login because I already had an account and I logged in. And once you set up your workspace on the left side, you can you can see you'll have create prompt. You'll also have a starter apps. there's a tune model uh tuner model uh option that is enabled and you can do
[2:29]More. So the first thing first I used it to create a blog. So I gave it an exact prompt. I said write a short and engaging blog posted on this picture. It should include a description of the meal in the photo and talk about my journey about meal prepping. So how this helps is I just provided it with an image and it can accurately write a blog for me. It helps save me some time. I don't really have to write the blog myself, but I can go through it, format it, and change make specific customization and changes depending on my needs. So the next thing is it works great with
[3:10]Mathematical questions. For example, you can see how complex this question is. So I have in this in this example I've asked it to create a equation of the two lines L1 and L2 and it has given me a detailed understanding of the problem and the detail answering to that question and how it reached to the answer. It can also solve trigonometric problem solvings and more. So this is the part where it does video analysis. You can upload like I said you can upload any video. You can even upload this tutorial and ask Google AI at what moment at what time and you know this was said or what
[3:54]Was this video about and more. For example, I uploaded a stat quest video and then I asked to explain reinforcement learning for me. So you can see along with the time frame at what time at at O42 it said this and from this time to this time the video explains about decision trees from this time to this time it explained about reinforcement learning and more. So you can see how accurate Google AI is able to do that for you. Now this is how Google AI looks like and this is the interface and you can ask multiple questions. You can create a prompt, you can ask it anything. Create
[4:39]A blog for me about food creation. Like for example in this I asked about food. I asked to create three food prompts for me, three blogs and asked it to give me the recipes for it. So as you can see it created the blog post and it it even suggest image prompts like for example I can use deli or even other modules like stable diffusion and more to use. You can use um stable diffusion and more to give this particular description of the image and then generate. For example, I can say overhead shot of a rectangular baking sheet with filled with roasted chicken pieces more and so
[5:24]On and so forth. Basically how accurate your prompt is, you can uh generate the image as such. Now uh this is the example where I was talking about the video. So I uploaded a video here of the versel tutorial that is uploaded already on bear brown and co. And then I said I don't understand this tutorial. So help me understand with time frames what is said when. So as you can see with the time frame it said from this second to this second we talked about this. From this minute to this minute it talked about this. So with a proper split of the time frames and the timestamps, it
[6:03]Was able to tell me at what particular time what was said. So this helps me break down the video tutorial into parts that I will be able to understand and work with. It helps you you know in if you don't have some time to watch the video you can just give the video up to Google AI studio and ask it to create a whole tutorial out of it for you and you can just skim through it and understand. Now tune a model option is where you use this particular Gemini's model to tune it to your needs. So there is another video on fine-tuning what fine-tuning is
[6:42]That is already put up on bear brown go please go through it and then you will understand what fine-tuning is so in this model you can fine-tune Google AI's Gemini model for your purposes for example um you want it to understand a particular sentence like Gemini already understand sentiment analysis if you give a particular sentence it will be able to tell you that you know um if this person or the sentence is neutral positive or negative but for what about examples or sentences like the car was running at 2500 RPM and this is okay that is okay so sometimes it is difficult to understand statements like that if it
[7:33]Was neutral positive or negative so if your application is to understand the car's need or more about that you can fine-tune this model by giving it data sets. For example, I can create a data set a CSV file out of uh car related sections and then I can give it to this. What I can do is I can click a data source. I can import and then I can import the file here whichever file that I need and then I can say tune this model once so the models that it's tuning is Gemini is 1.5/0.01 01 fine tuning. So once this is tuned, you can use that. For example, here I've
[8:21]Tuned my um I created a data set which is a very basic example. I created a data set where you can um I created a data set where it was just classification of certain sentences into neutral and positive for like this is the data set. This product is amazing so I built it as positive. It was an average experience so I said it was neutral. The service was horrible so it was negative. So these are the various things that is just a basic example of just what Google AI can do. You can train, you can change your data set based on your specific needs. I understand that the data set
[9:07]That I showed you right now is already what Gemini 1.5 can do for you. But that was a basic example. So like I said, you can change that example into how into your particular needs. tune it to your needs and give that upload that data set and fine-tune the Gemini 1.5 model. Here you can see how as and when it was learning from each and every uh data point how the loss has reduced. Initially there was a high loss. It was not able to understand. Then it started reducing indicating that it was able to learn. And then when you click on use in chat you will be able to use it here.
[9:45]You can type in your prompt based specifically on your particular model what you fine- tuned it for and ask a question based on it and Gemini would be able to understand learn from your model and give you the right answer. The more data set you provided, the more training data you provide, it will be able to give you more accurate answers.
More videos
2:08Bridging the Pixel Gap in Browser Automation.
2:23How One Narrow Safety Rule Can Make an AI Less Safe Everywhere Else.
2:04Why splitting a chunk from its document makes it retrieve for the wrong question
4:20Three You Can Take Back. One You Can't.
2:21Why a 50-turn agent pays for the same screenshot 35 times unless it caches the pixels
1:53