AI Teading - Hands-on AI experiments and lessons

This is an application that adds a machine learning library to Scratch 3. Machine learning functions are organized in the form of blocks, so they can be used similarly to existing Scratch blocks.

A DQN-based self-driving car simulator. This simulator leverages DQN to help cars learn how to drive themselves.

This is a game that removes frost from images captured by a camera. The game was created using a hand recognition model.

This application can recognize objects commonly seen in everyday life in 3D. It estimates and displays the pose of objects learned with specific data in real time.

MediaPipe provides a solution for hand and finger tracking. It can track the movements of 21 key parts in 1-4 hands.

MediaPipe provides a facial recognition solution. It estimates 468 points on the face and converts them into a 3D shape to fit the face.

MediaPipe provides a pose recognition solution. It estimates the positions of 33 major parts of the entire human body and connects them to display them like a skeleton.

Integrates MediaPipe's hand, face net, and pose features.

MediaPipe's facial recognition solution quickly detects the presence or absence of a face and displays six key parts. The functionality is simple and the processing speed is very fast.

MediaPipe's solution for removing people or backgrounds from real-time video. This feature can be used for video compositing, etc.

This is a face recognition app created using TensorFlow. This app guesses and displays information such as expression and age from the recognized face.

Provides a feature to remove people from the background in real time using TensorFlow. This feature recognizes people and removes them from a scene through the process of learning the background.

This is an app that recognizes faces and poses from still photos and uses them to create character animations.

This is an app that recognizes people's poses and faces in real time from captured videos and turns them into moving character animations.

You can create character animations using MediaPipe and TensorFlow. This feature recognizes a person's face or the entire person and makes the character move along.

This is a game that learns from QuickDraw's hand drawing dataset to recognize what the user's drawing is. The game uses CNN to learn and recognize drawings.

This is an app that uses CNN to learn QuickDraw's hand drawing data and recognize the learned drawings. This app can recognize 100 drawings.

This is a game where the user draws a picture presented as a problem and the learned artificial intelligence guesses what the picture is.

This is a drawing app that predicts what the user has drawn and presents icons with a high probability of matching so that the user can select them.

This is a continuous drawing app created using RNN. It learns the QuickDraw data set and has a function to continue drawing like a hand drawing when the user draws.

This is an app that demonstrates the process of learning how to avoid a ball using DQN. The agent receives information from sensors and receives more rewards the longer it stays alive.

This is an app that demonstrates the learning process of avoiding enemy planes using DQN. The agent receives information from sensors and the longer it stays alive, the more rewards it receives.

An app that demonstrates the process of performing reinforcement learning using TensorFlow. The goal of the process is to balance the rod for as long as possible.

You can see the process of learning the snake game using DQN. Reinforcement learning was carried out so that you receive a negative reward when you hit a wall or your own body, and you receive a positive reward when you eat fruit.

You can see the process of learning the 2048 game using the MCTS algorithm. If you provide a large number of simulations in one move, you can reach 2048 with a high probability.

We used a supervised learning algorithm to teach Gorilla how to play the game. We used a simple neural network built with TensorFlow.

You can see the Flappy Bird game learning process using neural networks and genetic algorithms. As the birds pass through generations, they become better at flying based on information from previous generations.

Using CNN, we created an artificial intelligence that learns the game of Tetris. It is repeatedly trained to place blocks similar to the match data of champions.

This is a game where you find objects around you that match the presented emoticon. Images are predicted through CNN, and the prediction matches the correct answer (emoticon).

It is a web-based machine learning learning tool designed to allow anyone to quickly and easily create machine learning models. You can create image, audio, and pose models.

Using the Performance RNN model, a song is created by referring to the frequency of use of key signatures. The model learned musical structures and patterns using LSTM.

This is an app that utilizes the MusicVAE model to generate repeating melodies. MusicVAE can create harmonious melodies by smoothly connecting multiple melodies.

We created an arpeggio pattern to play using Magenta and Improv RNN. Improv RNN conditionally generates a melody based on the basic chord progression.

This is a drum composition app created using a pre-trained drum RNN model. If you provide a seed pattern, it continuously generates the next pattern based on it.

This is a game where you can control your character through motion recognition using TensorFlow and a camera. The game character follows the user's actions and attacks the opponent.

This is a Pac-Man game that allows you to manipulate direction through image recognition. A pre-trained MobileNet model was used to recognize motion in four directions.

Keras is a deep learning library that helps you quickly and easily implement artificial intelligence. In the demo, you can check out Keras' learning process and performance in your browser.

ConvNetJS is a library that allows you to train deep learning models in the browser. Through the demo, you can visualize the learning process and results of the neural network.

You can take a step-by-step look at the rock-paper-scissors image recognition learning process using TensorFlow.

Load deep learning models pre-trained with TensorFlow and output them as 3D visualization scenes. You can visualize the processing of each model.

A program based on TensorFlow where the Tyrannosaurus character learns how to avoid obstacles through several types of training.

This is a program implemented to help users easily understand the inner workings of CNN. You can visualize the image recognition process step by step.

This is a program that allows you to examine the learning process of GAN in real time with an interactive visualization tool. The operation process of the generator and discriminator for 2D data is presented as an example.

This is a service that allows you to animate landscape photos using the AnimeGAN model.

This app learns from various MIDI examples and plays the next part according to the user's performance. When you play a few notes, the next part is played accordingly.

Using LSTM, it remembers the strokes of letters written by the user and displays predicted letters based on them.

This is an app that uses MediaPipe's FaceMesh to recognize faces and create virtual avatars based on them.

This is a GAN model that creates a face image with selected attribute values. You can create the desired character image by changing the model type and attribute values.

You can play the 2048 game with image recognition using CNN.

This is a service that uses the Pix2Pix GAN model to receive input source images such as sketches and convert them into other types of images.

An autonomous driving simulator built with WebGL and Three.js. The simulator is designed to avoid static and dynamic obstacles and reach the destination.

This is an autonomous driving demo visualized using the KITTY and NuScenes datasets using streetscape.gl (a tool for visualizing autonomous driving data).

This app shows the operation of a genetic algorithm through a parking learning process. Using a genetic algorithm, a car is trained to park on its own.

This is an app that shows ethical issues that can arise during autonomous driving in the form of a game. There is no correct answer, and you can check other people's choices.

You can train a self-driving car by applying various reinforcement learning algorithms. You can check each reinforcement learning process and its results.

OpenAI's flagship conversational AI. Powered by the latest GPT-5 model, it offers multimodal capabilities to chat with voice, analyze images, and draw pictures, in addition to text.

An AI image generation tool that can be used for free without logging in. Formerly famous as 'DALL·E mini'.

Gemini (formerly Bard) is Google's latest multimodal AI model. It can understand and process text, code, and images, and provides a variety of information in conjunction with Google services.

Microsoft's AI assistant based on GPT-5 (formerly Bing Chat). Combined with the Bing search engine, it answers based on the latest information and draws pictures via DALL·E 3.

Microsoft's image creation tool powered by DALL·E 3 (formerly Bing Image Creator). It converts sentences entered by the user into high-quality images.

Choose a Pokémon and chat with it using an AI model, providing a fun and educational experience

A Korean AI portal service where you can select and use famous latest AI models from around the world, such as GPT-5 and Claude 3.5, for free in one place.

Image generation AI provided by Adobe. You can experience various features such as text-to-image and generative fill on the web.

Currently the most popular music generation AI that creates high-quality songs including lyrics, vocals, and accompaniment just by entering text prompts.

A conversational AI that is a strong competitor to ChatGPT, with excellent Korean writing skills and the ability to understand long contexts.

A service where you can chat with AIs having various personas like real figures or anime characters. It offers a highly immersive experience as if talking to a real character.

An opera created by four Blobs trained with machine learning. Adjust pitch and vowels with your mouse to create harmonies in real time.

A project by Meta AI Research that turns your character drawings into animations. Upload a drawing and watch it dance and run.

One photo is real, the other is AI-generated. Test your ability to distinguish between real and fake images.

A word association game powered by machine learning. Test how well the AI understands the meaning and context of words.

Play along with an AI bird trained on cello and violin performance data. The AI generates bowing motions and pitch in real-time based on your mouse movements.

Draw on the canvas, and music plays based on the position and instrument. A synesthetic AI experiment turning visuals into audio.

A motion capture AI that uses your webcam to make your doodles animate and mimic your movements in real-time.

See how closely your singing matches Freddie Mercury's voice. AI analyzes your pitch, timbre, and melody to give you a score.

AI imagines and plays sounds that match an image. You can wander around Google Street View and hear the 'imaginary sounds' of that location.

A Google Magenta experiment that converts a sound you sing or bang into the sound of instruments such as flute, saxophone, trumpet, or violin. Each instrument model was trained on a 10-minute recording of a real performer, so breathing sounds and bowing noises come through as well.

Google's first AI-powered Doodle, released on Bach's birthday in 2019. A machine learning model harmonizes a two-measure melody you compose into Bach's signature style. It uses Coconet, a model trained on 306 of Bach's chorale harmonizations.

An experiment from Google Creative Lab that arranges thousands of everyday recorded sounds into a map by similarity, then lets you build drum beats out of them. The raw material is ordinary noise rather than instruments.

A web tool that uses AI to split an uploaded song into vocals and instrumental. You get two tracks: a karaoke version without vocals and an acapella version with the voice isolated. The site says processing usually takes about 10 seconds.

A Google Arts & Culture experiment that uses machine learning to harmonize a melody you enter in the style of Johann Sebastian Bach. You can replay the result with a different instrument or tempo, or share it by link.

A free tool that reads text aloud in a human-sounding voice and turns it into an audio file. It offers many languages including Korean along with a range of voices, and states that the generated audio can be used commercially.

A Magenta experiment where you build lo-fi hip hop by clicking objects in a small virtual room. The TV in the middle is wired to MusicVAE, which blends melodies, and the radio beside it to MelodyRNN, which generates new ones, so the music changes in real time as you click.

An experiment where four different drum patterns sit at the corners of a square and moving through the space between them produces new in-between beats. It lets you hear what rhythms can exist between two rhythms.

A Google demo page showing how the browser's built-in speech recognition works. Speak and your words are transcribed into the box on the spot, and you can pick the recognition language from a long list that includes Korean.

A visualization that walks step by step through how Stable Diffusion turns text into an image. On one screen you can follow how the prompt is split into tokens, encoded into vectors, refined from random noise over many timesteps, and finally upscaled.

A tool where you paint over the part of a photo you want gone and AI fills that area in so it blends with its surroundings. It can be used to remove people, text, logos, or date stamps.

A tool that raises the resolution of 2D images using a convolutional neural network. It handles photos as well as illustrations, and can reduce JPEG compression noise while upscaling. You can see for yourself how much jaggedness disappears when a small image is enlarged.

An experiment made by the Google Arts & Culture Lab with artist Jack Wild. You look at an AI-generated image and describe it in words; your description becomes a prompt that generates a new image. How closely the result matches the original tells you how precise your description was.

An AI tool that adds color to black-and-white photos. You can pick from over 21 color filters, or type keywords to steer the colors yourself. It lets you see how an old family photo or a historical image looks once it has color.

A tool that erases the background of a photo, leaving only the person or object. Upload an image and it automatically finds the boundary between subject and background and cuts it out. You can check how far the AI can trace tricky edges such as hair.

A tool where you paint over the part of a photo you want gone, and it fills that area in with the surrounding background. The model runs directly in the browser, so your photo is not sent to a server. You can watch how the AI imagines what belongs in the erased spot.

A hands-on space for trying Meta's SAM 3 model in the browser. Type a phrase for what you want to find in an image or video, and it locates every matching object and cuts it out as a region. You can then apply effects such as blur, clone, or contour to those regions with one click.

A teaching visualization that walks step by step through how a GPT-style language model picks the next word. It actually runs GPT-2 (small) in the browser so you can follow your sentence turning into tokens and embeddings, passing through self-attention and an MLP, and ending as next-word probabilities. You come away understanding that a language model is not magic but a machine computing which word is most likely to come next.

A guided tour that unfolds the inside of a GPT language model in 3D and walks you from top to bottom. It starts with nano-gpt, a model of just 85,000 parameters, doing one simple task: sorting a string like "C B A B B C" into alphabetical order. You come away with a feel for what actually happens at each layer as numbers become embeddings, pass through the stack, and end as next-token probabilities.

A Korean word game where you try to guess the day's answer word. Type a word and it reports how close the meaning is to the answer on a scale from -100 to +100, adding a rank if the word is in the top 1,000. Playing it gives you a feel for word embeddings, which compare words by meaning rather than spelling.

A game that starts with just Water, Fire, Wind, and Earth, and has you combine elements two at a time to make new words. Drop one word onto another and the AI produces a new word with an emoji to match the pairing. Everything you make collects in a "Discoveries" list so you can see how far you have gotten.

A game in which you talk to a chatbot named Gandalf and try to get it to reveal a hidden password. Each time you succeed, Gandalf strengthens its defenses, so the same trick will not work on the next level. Playing through it is a hands-on way to learn the basics of AI safety.

A site for comparing how several AI models answer the same question. From "Battle Mode" at the top of the screen you send a prompt and get responses from multiple models to weigh against each other, and the site also publishes a model leaderboard. The site shows its name on screen as "Arena".

A Korean-made service that uses AI to correct spelling, spacing, and grammar in Korean text. Built by BAIKAL.AI, the same site also offers morphological analysis and a tokenizer. Proper nouns and technical terms can be registered in a user dictionary so the checker takes them into account.

A Korean-made service that summarizes YouTube videos, PDFs, web pages, and audio. Beyond summaries it can explain the content in simpler terms or let you keep asking questions about the material. For videos it shows a timestamped transcript alongside a Korean translation.

A service where you upload a PDF and then ask questions about its contents. Answers include citations, so you can see which part of the PDF they came from. Besides PDFs it accepts Word, PowerPoint, Markdown, and text files.

Naver's translation service, centered on Korean and covering English, Japanese, Chinese, and other languages. Alongside text translation it has separate tabs for image, document, voice, and website translation. Results come with pronunciation playback and a way to rate the translation.

A place to chat directly with the Olmo models released by the Allen Institute for AI (Ai2). The model in use is shown on screen; at the time of capture it was Olmo 3.1 32B Instruct. You can also enable tool calling and add your own tools.

A translation service focused on rendering longer text naturally, stated to support more than 100 languages. Besides text translation it offers file and voice translation modes, plus DeepL Write for polishing the wording. It includes controls for formal or informal tone and a glossary for consistent terminology.

A Google service; opening the address brings up a Google account sign-in screen first. The sign-in screen is available in Korean and many other languages.

An AI search service: write your question as a sentence and it looks things up on the web and returns an organized answer. Below the input box you can choose a search mode and which model to use, and there is a voice input button. The left menu has Spaces, Artifacts, and History for revisiting earlier sessions.

A browser viewer for 3D scenes reconstructed from photographs (Gaussian splats). It was made by Kevin Kwok and the code is published on GitHub. You can move the camera to look around the scene as if walking through it, with the frame rate shown in a corner.

A 3D world generation service from World Labs. You give it text, an image, a video, or a 3D structure, and it builds a 3D space you can move around in. It also hosts a gallery of worlds made by other users.

An experimental service where AI turns a written description into a real-time video scene. Rather than playing back a prepared clip, it takes your prompt and generates the scene as it goes. Simulations you make are collected in your own list.

An online research study run by Northwestern University's Computer Science department. You read short speech texts and judge whether each came from a Democrat or a Republican, with an 'explanation hint' shown to help you decide. The study looks at how well people can tell real from fake, or trace where a statement came from.

A tool for opening the Content Credentials embedded in an image or video file — the record of how it was made and edited. Upload a file and it shows, step by step, what tool created it and how it has changed since. It is a good way to check provenance labelling on AI-generated media firsthand.

A free, open-source tool that upscales video and images entirely inside your browser. It ports upscaling algorithms such as Anime4K and RealESRGAN to WebGPU, so your file is never uploaded to a server. The site notes it works particularly well on animated content.

A browser-based editor for cleaning up Gaussian splat files, the output of 3D scanning. Load a scene and you can orbit it in the 3D viewport, select and delete unwanted points, and adjust position, rotation and scale. It runs as a web tool with nothing to install.

A generative creation tool that turns text or images into video. Alongside video generation it gathers image generation, motion control, a canvas and avatar features in one workspace. The landing page shows a feed of work posted by other users.

A motion capture tool that reads a person's movement from your webcam or an uploaded video and turns it into 3D animation data. No sensor suit is needed — a camera alone records the motion so you can drive a character with it. The page also carries a notice that the service is becoming 'Rokoko Create'.

A game for learning the American Sign Language (ASL) fingerspelling alphabet by copying handshapes in front of your webcam. It covers the ASL A–Z handshapes, not Korean Sign Language. It was made by the American Society for Deaf Children and uses MediaPipe Hands for hand tracking.

A TensorFlow.js demo that estimates depth from a single portrait photo and turns it into a moving 3D photo. It shows the original, the segmented subject, the depth map and the 3D result side by side, so you can see each stage. The result can be downloaded as a GIF or WebM.

A tool for seeing how a language model splits text into tokens. Type something and it immediately shows the token count, the colour-coded token pieces and the numeric ID of each token. You can also see the special markers that wrap a conversation before it reaches the model.

A visualization that unfolds the inside of a convolutional neural network (CNN) recognizing handwritten digits into three dimensions. You can watch a drawing get summarized into smaller, more abstract values as it passes from the input layer through convolution, downsampling, fully-connected and output layers. It helps you understand that a neural network is not one opaque box but a chain of layers with different jobs.

An educational game, about six minutes long, showing how AI used in hiring inherits human bias. You play a company owner picking applicants by hand, then hand the screening over to an AI trained on your own past choices in order to hire faster. Watching that AI begin to filter certain applicants out, you come away understanding that bias is not invented by the algorithm but carried over through the training data.

A site that presents one unavoidable crash after another, where a self-driving car with failed brakes must choose whom to hit, and asks you as an outside observer which outcome is acceptable. When you finish the set of scenarios, your pattern of judgements is summarized and compared with other respondents. You come away seeing what criteria are at stake when such decisions are handed to a machine, and how much those criteria differ from person to person.

A visual explainer from Google PAIR that probes what the language model BERT learned about the world by asking it to fill in a blank. Change a single word, as in 'in texas, they like to buy ___', and you can watch the candidate words and their probabilities reorder themselves. You come away understanding that a language model's predictions are statistical associations drawn from the text it read, not neutral knowledge, and that social bias is learned along with everything else.

A collection of visual essays on machine learning concepts from Machine Learning University, Amazon's education initiative. Fourteen articles cover neural networks, linear and logistic regression, decision trees and random forests, precision and recall, ROC and AUC, cross-validation, the bias-variance tradeoff and more, each built around charts you manipulate yourself. You come away seeing what these concepts actually look like, rather than only how their formulas read.

A TensorFlow tool that reduces embeddings — data such as words represented as vectors of hundreds of dimensions — down to two or three dimensions you can look at. It opens on 10,000 Word2Vec words in 200 dimensions, and picking one word lists its nearest neighbours. You come away seeing directly that words with related meanings sit close together in the space.

A tool for following, step by step, how word2vec builds its word vectors. Feed it a tiny training set such as 'eat|apple, drink|milk' and run training: the cells of the input and output weight matrices shift colour a little at a time while the word points settle into place in the PCA plot below. You come away understanding that an embedding is not a value fixed in advance but something produced gradually by repeated training.

Google's introductory machine learning course, fully translated into Korean — one of the few proper Korean-language courses in this catalog. Modules cover linear and logistic regression, classification, working with data, neural networks, embeddings, large language models, production ML systems and fairness, each labelled with an estimated study time so you can work through them in order. The linked page, 'Neural networks: interactive exercises', has you edit hidden layers and weights yourself and see which nodes' calculations are affected.

A demo where agents trained by deep reinforcement learning — a bipedal walker, a chimpanzee, a fish — are dropped into different terrains so you can watch how they move. Besides the prepared environments, you can draw the ground and ceiling yourself and put the agents into situations they never saw during training. You come away seeing the generalization problem first-hand: learned behaviour works well in familiar surroundings and can fall apart when the environment changes.

An interactive tool that lays out a transformer model's attention patterns by layer and head so they can be compared side by side. Because query and key vectors are projected into the same space, you can see at a glance that different heads produce quite different shapes. You come away seeing that attention heads each track different relationships — note that it assumes you already know how transformers are structured.

A set of demos for a JavaScript reinforcement learning library that runs straight in the browser. You can watch values settle in a gridworld under dynamic programming and tabular temporal-difference learning (SARSA and Q-learning), and watch deep Q-learning (DQN) agents move in PuckWorld and WaterWorld. You come away seeing the basic structure of reinforcement learning: behaviour is refined by nothing more than rewards and penalties.

A tool that opens a trained neural network model file and draws its layer structure as a diagram. You can follow which layers connect in what order right on screen.

A chat service that runs a language model directly inside your browser. Unlike other chatbots, your conversation is processed on your own computer instead of being sent to a server. The interface supports Korean.

A 2D virtual ecosystem where organisms made of differently colored cells reproduce, compete, and evolve. Each color has its own role, and you can watch mutations change shapes and behavior over generations.

A hands-on page where you train a small neural network right in the browser. Change the number of layers and neurons, the activation function, or the learning rate, and watch how the decision boundary shifts. There is no model to download, so it runs the moment you open it.

A demo hub that lets you try 15 MediaPipe recognition tasks from a single page. It covers face, hand, and pose detection, gesture recognition, image classification and segmentation, object detection, audio classification, language detection, and text classification. The processing runs inside the browser.

A page that colors in how your sentence is split into tokens inside an AI model. You can pick and compare the tokenizers of 13 models including GPT-4, Claude, Llama 3, Gemma, Mistral, BERT, and T5.

Give it a photo and a painting, and it redraws the photo in that painting's style. Everything is computed in the browser with TensorFlow.js, so your photo is never uploaded. There is also a tab for blending two styles together.

A collection of articles showing how neural networks can be improved across generations. Demos include target-seeking agents, an Agar.io-style game AI, and color classification, so you can watch performance improve through selection and mutation rather than ordinary training.