Artificial intelligence often feels like a collection of impenetrable black boxes. When trying to learn how modern neural networks work, students usually face a wall of dense academic jargon or abstract code libraries that hide the actual mechanics.
Today, we are thrilled to launch "AI from Scratch", our new interactive visual curriculum designed to change how you learn machine learning. Comprising 19 live-computed explainers across 5 bite-sized modules, this series builds modern AI from first principles using nothing more than school-level mathematics.
Here is a look behind the curtain at why we built this curriculum and what you will learn.
A Unified Visual Language
We believe learning is faster when the rules don't change. The entire curriculum leans on a singular café sales dataset and a strict colour vocabulary that persists across every diagram and formula:
- Data and keys stay in their designated colour.
- Models and queries take another.
- Error, blame, and loss are always grouped visually.
Because these visual boundaries never shift, your brain doesn't have to re-learn the landscape when moving from a basic linear model to a complex transformer architecture.
Break the Math to Understand It
We skipped the static text and pre-recorded videos. Every single explainer runs live in your browser using our custom STEM explainer toolkit.
We heavily leverage interactive sliders that let you control variables like learning rate (\(\eta \)), weights, biases, and temperature. We actively encourage you to push these sliders past their limits. By deliberately breaking the simulation, you can visually witness gradient descent violently zigzagging out of control or watch a model overfit data in real time. True understanding lives at the boundaries where these concepts break.
The 5-Step Journey to Modern AI
The curriculum is structured to build your intuition step-by-step:
- Module 1: What Learning Is. Start by drawing a simple rule through sales data. Map out a "loss landscape" and physically step down the slope to discover how gradient descent functions without a map.
- Module 2: The Neuron and the Network. Move from a straight cut on a plane to a hidden layer that folds geometric space. Watch a network learn a complex double-spiral live, and visualise backpropagation as a fluid flow of "blame" using the chain rule.
- Module 3: Seeing. Slide a matrix across pixels to grasp convolutions and feature hierarchies. Explore embeddings to see how words become geometric points where semantic arithmetic (like king − man + woman = queen) perfectly aligns in vector space.
- Module 4: Sequences and the Transformer. Deconstruct the architecture powering modern language models. Interact with attention heads, map out the sine-wave fingerprints of positional encoding, and watch how residual streams flow through a transformer block.
- Module 5: Beyond Fitting. Tackle the trade-offs of bias vs. variance, explore reinforcement learning via an agent navigating a grid, and reframe probability distributions as evolving Bayesian beliefs.
Start Exploring Now
"AI from Scratch" succeeds because it trades memorisation for intuition. You don't need a PhD in mathematics to grasp the core concepts of modern machine learning; you just need the right perspective.
The entire series works perfectly on your phone or desktop, requires no prerequisites beyond basic math, and takes just a few minutes per module.
