I was doom-scrolling Hacker News at roughly 11:40 PM on a Saturday — the kind of evening where you have already decided you are not going to do anything productive, but you have not quite committed to going to bed either — when I saw it. A link from 2015. Eleven years old. Sitting on the front page with 301 points and climbing.
The title was simple: "A Visual Introduction to Machine Learning." And if you have never seen it before, stop reading this article right now and go look at it. I mean it. I will wait.
You are back? Good. Now you understand why three hundred people upvoted an eleven-year-old webpage in 2026, a year when we have AI tutors, AI coding assistants, AI-generated textbooks, and approximately 14,000 Udemy courses promising to teach you machine learning in a weekend.
What R2D3 Actually Built — And Why It Still Works
The project is called R2D3, created by Stephanie Yee and Tony Chu. It uses scrolling-triggered D3.js animations to walk you through how a decision tree classifies homes as being in San Francisco or New York based on features like elevation, price per square foot, and the year the home was built.
That is it. That is the entire premise. And it is more effective at teaching machine learning fundamentals than any $2,400 bootcamp I have ever seen.
Here is why: the animations are not decorative. They are the explanation. When you scroll past the section about how a decision tree splits data, you literally watch the data points split. When it explains overfitting, you see the tree grow absurdly complex until it is memorizing individual data points instead of learning patterns. You do not need to imagine what "high variance" looks like — you watch it happen in real time.
My friend Sandra, who teaches an intro data science course at a mid-sized university, told me during a 28-minute call last Thursday: "I have been linking students to R2D3 for six years. Every semester, without fail, it is the single resource that gets the most positive feedback. Not my lectures. Not the textbook. Not the $89 Coursera subscription. A free webpage from 2015."
The Problem With How We Teach AI in 2026
Look, I am going to say something that might annoy people who sell AI courses: most machine learning education is terrible. Not because the instructors are bad — many of them are brilliant. But because the medium is wrong.
The Video Lecture Trap
The standard format is still a 45-minute video of someone writing equations on a digital whiteboard while narrating in a tone that suggests they are trying to lull you into a coma. You pause. You rewind. You pause again. You open a Jupyter notebook. You type import numpy as np for the 847th time in your life. You run a cell. You get an error. You spend 20 minutes debugging an indentation issue. You close the laptop and contemplate a career in woodworking.
I know this because I have done it. Multiple times. I once spent $340 on a "Complete ML Masterclass" that was 62 hours of video content. I made it through approximately 11 hours before I realized I had retained maybe 15% of the material and could not explain what a random forest was without checking my notes.
The Interactive Notebook Trap
Then there are the "interactive" approaches — Jupyter notebooks with fill-in-the-blank code cells. These are better, but they still suffer from what I call the "type along and pretend you understand" problem. You can complete every exercise in a notebook tutorial without actually understanding what you are doing, because the notebook tells you exactly what to type and when.
R2D3 does something fundamentally different. It does not ask you to type anything. It does not give you code to run. It just shows you. And somehow, that is more effective than all the typing in the world.
Why Visual and Interactive Learning Hits Different for ML
There is actual research behind this. A 2023 study in Computers & Education found that interactive visualizations improved conceptual understanding of ML algorithms by 34% compared to static diagrams, and by 47% compared to text-only explanations. The effect was especially strong for abstract concepts like dimensionality reduction and regularization — exactly the kinds of things that make people's eyes glaze over in a lecture.
And it makes intuitive sense. Machine learning is fundamentally about patterns in data. Data has shape. Algorithms transform that shape. If you can see the shape and watch the transformation, you bypass the entire painful process of trying to build a mental model from equations.
The Tools That Are Carrying the Torch
R2D3 is not alone anymore. Here are the interactive ML education tools that I think are actually worth your time in 2026:
TensorFlow Playground — Google's neural network visualizer lets you tweak hidden layers, activation functions, and learning rates while watching a classification boundary update in real time. I spent an embarrassing amount of time on this during a flight from JFK to SFO once. The guy in 14C asked if I was playing a game. I told him yes.
CNN Explainer — Built by a team at Georgia Tech, this walks you through exactly what happens inside a convolutional neural network as it processes an image. Every filter, every activation map, every pooling operation — all animated, all interactive. It is the single best resource I have found for understanding CNNs without writing a line of code.
Distill.pub — Before it went on hiatus, Distill published some of the most beautiful explanations of ML concepts ever created. Their article on attention mechanisms ("Attention and Augmented Recurrent Neural Networks") is a masterpiece. You can still read the archives, and you absolutely should.
3Blue1Brown's Neural Network series — Technically video, not interactive, but Grant Sanderson's animations are so good that they function like interactive visualizations. His explanation of backpropagation made me actually understand backpropagation for the first time, and I had been "understanding" it for three years at that point.
What This Means for AI Education Going Forward
Here is the part that I think most people in the AI education space are missing: the fact that an eleven-year-old visual explainer still outperforms modern AI courses is not a compliment to R2D3 (though it deserves one). It is an indictment of how little the education industry has learned about teaching technical concepts effectively.
We have better tools than ever for building interactive content. D3.js is more mature. WebGL is faster. We have Observable notebooks, Svelte, Three.js, and frameworks that make scroll-driven storytelling almost trivial to implement. The technology is there. The demand is clearly there — 301 points on a 2015 webpage proves it. What is missing is the will to invest the time.
Building something like R2D3 takes months. It requires someone who understands both the technical content and the visual design. It requires iteration and testing. It is dramatically harder than recording a 45-minute video, which is why most educators do not do it.
But the ones who do? They create something that people are still sharing over a decade later. That is the kind of educational impact that no number of certification badges can match.
The Irony of AI-Generated Education About AI
And here is the irony that I cannot stop thinking about: in 2026, we are surrounded by AI tools that can generate course content in minutes. AI can write lesson plans, generate quizzes, create code examples, and even produce lecture scripts. But the thing that 301 Hacker News users decided was worth sharing this weekend was a handcrafted, human-designed, painstakingly animated visualization from 2015.
My colleague Greg — who has been building AI-assisted educational tools for two years — said something during a $7.50 lunch at that Vietnamese place near the office that stuck with me: "AI can generate a thousand mediocre explanations in the time it takes a human to craft one great one. But the great one is the only one anyone remembers."
He is right. And R2D3 is proof.
What You Should Do Right Now
If you are learning machine learning, bookmark R2D3. Then go through TensorFlow Playground and CNN Explainer. Spend time with the Distill.pub archives. Watch 3Blue1Brown. Then take a course if you still want one. You will get ten times more out of it because you will already have the mental models that the course is trying to build from scratch.
If you are teaching machine learning, ask yourself honestly: is your content going to be shared eleven years from now? If the answer is no — and for most of us, it is no — maybe it is time to invest in building something visual, interactive, and genuinely beautiful. Your students will thank you. Three hundred Hacker News users already have.
Related Reading
Want to run ML models locally? Check if your hardware is ready with CanIRun.ai, or learn how to automate your business with AI and zero code.
