AI/ML Visualizations: Explore LLMs, Tokens, and Neural Networks
Understanding how AI works can feel daunting. Concepts like Large Language Models (LLMs), tokenization, neural networks, and next-word prediction are often abstract and hard to grasp from text alone. I created this collection because interactive visualizations were the key to my own learning, transforming complex machine learning theory into tangible, explorable concepts.
This page gathers the best interactive tools and videos that helped me. They are designed to let you see and interact with the fundamental building blocks of AI, from how a model processes a single "token" to the high-level architecture of deep learning systems. My hope is that they make your journey into the world of AI/ML as insightful as they made mine.
Interactive Websites
Tiktokenizer
Tiktokenizer is a tool that lets you see firsthand how Large Language Models (LLMs) process text by breaking it down into units called "tokens". This visualization reveals why LLMs can struggle with seemingly simple tasks like counting or precise calculations involving large numbers; the model doesn't see "1234567890" as a single number, but as several distinct tokens like "123", "456", and "7890". The corresponding full vocabulary is here.
Moebio Mind
Moebio Mind is an interactive art project that provides a playful, abstract visualization of a neural network. Instead of analyzing data, this tool lets you directly interact with a simulated mind; each click sends a pulse through a web of nodes, revealing how simple connections can create complex patterns. It serves as a conceptual model, allowing you to build an intuition for how interconnected systems - the foundation of modern AI - process and react to stimuli. (Created by digital artist and data scientist Santiago Ortiz, see also this interview on his work.)
Transformer Explainer
The Transformer Explainer is an interactive visualization that demystifies the architecture behind modern LLMs. It allows you to step inside a live GPT-2 model to see how it processes text, specifically illustrating the "self-attention" mechanism - the technique models use to weigh the importance of different words when determining context and meaning. This tool is particularly valuable for understanding how an AI goes from seeing a sequence of words to making an informed prediction about what comes next.
BertViz
BertViz is an interactive, open-source tool in Google Colab (all cells pre-loaded) for visualizing the inner workings of Transformer-based models like BERT, designed for developers and researchers. The tool specifically illustrates the "self-attention" mechanism by showing how a model weighs the importance of different words when processing text; for example, you can see how it connects a pronoun like "it" to the noun it represents. This makes it an invaluable resource for debugging and interpreting why a model makes certain predictions, turning abstract attention scores into a clear and explorable visualization. (Developed by Jesse Vig, BertViz on GitHub, and introduced in the paper A Multiscale Visualization of Attention in the Transformer Model.)
LLM Visualization
LLM Visualization is an interactive tool that breaks down the step-by-step process of how a Large Language Model generates text. It walks you through converting a prompt into tokens, then into numerical representations called "embeddings," and finally shows how the model predicts the next word in a sequence. It is useful for understanding the core mechanics of an LLM, transforming the complex "black box" of AI into a sequential process and revealing it as a sophisticated act of pattern matching and probability. (Created by Brendan Bycroft.)
LLM Architectures
LLM Architectures refer to the specific design and combination of components that define how a Large Language Model is built. Sebastian Raschka's The Big LLM Architecture Comparison provides an excellent set of visualizations that let you look "under the hood" to see how leading models like Llama, Mistral, and DeepSeek are constructed. The comparison highlights the critical engineering choices - for instance, using a "Mixture-of-Experts" (MoE) approach for greater efficiency, or adopting different "attention" mechanisms to better process context.[1][2][3] Understanding these architectural differences is valuable because it explains why some models are more cost-effective or better suited for specific tasks than others, moving the conversation beyond simple performance benchmarks.
Videos
Large Language Models explained briefly
This video from the excellent YouTube channel 3Blue1Brown offers in just 8 minutes a clear, conceptual overview of how Large Language Models (LLMs) work, making it a perfect introduction for beginners. It explains that at their core, LLMs are sophisticated systems trained on vast amounts of text to do one thing very well: predict the next word in a sequence. The animation visualizes how this simple-sounding task, when repeated, allows the model to generate fluent and coherent text. For those looking to go deeper, the entire channel is a valuable resource known for LLMs and Neural Networks, but also making complex topics in math and computer science equally accessible.
Deep Dive into LLMs like ChatGPT
In this comprehensive 3,5h video lecture, AI expert Andrej Karpathy provides an exceptionally clear explanation of the three-stage process that transforms a raw neural network into a sophisticated assistant like ChatGPT. The talk is valuable for anyone looking to understand why LLMs behave the way they do, from their impressive capabilities to their notorious flaws like hallucinating or being bad at counting. Karpathy breaks down the process into: 1) Pretraining, where the model ingests the internet to create a "Base Model"; 2) Supervised Finetuning (SFT), where it's taught to be a helpful conversationalist using high-quality examples; and 3) Reinforcement Learning (RLHF), the current frontier, where the model is refined based on human feedback to discover better cognitive strategies. The video is a must-watch for a deep, yet accessible, understanding of modern AI. (Andrej's Tweet.)
Fun Challenge
Bouncing Ball Hexagon
The Bouncing Ball Hexagon Challenge is a popular benchmark, first proposed by developer Flavio Adamo in January 2025, used to test the code generation and physics simulation capabilities of different AI models. AIs are prompted to write a program, usually in Python, that creates a realistic simulation of a ball bouncing inside a rotating hexagon, correctly applying principles of gravity, friction, and collision detection. This is a surprisingly difficult task that requires the AI to integrate concepts from geometry, physics, and programming to translate complex visual dynamics into functional code. The results vary widely between models, making it a quick and effective visual test to compare their reasoning and problem-solving abilities, with some models producing perfect simulations while others fail in amusing ways.
Vitruvian Robot
The Vitruvian Robot is a fun, creative test for AI models run by the author of this site, inspired by similar challenges like Simon Willison's prompt to draw a "pelican riding a bicycle". The experiment involves giving various AI models the same simple prompt: "Vitruvian Robot in SVG (embedded in HTML). Don't ask back.". This challenges the model to first access its world model to understand the reference to Leonardo da Vinci's 'Vitruvian Man' and then creatively merge it with the concept of a 'robot'. The resulting collection of images from over 37 models (and counting), ranging from simple stick figures to detailed robotic designs, provides an interesting and often amusing snapshot of how different AI systems interpret the same creative concept and their varying ability to generate functional SVG code. It serves as a playful exploration of the unique quirks and personalities of different models.