One neuron
Follow two ticket features through weights, a score, and a probability. See why changing a routing threshold does not change the model.
Open One neuron lesson →Fundamentals of Machine Learning
Understand how a model learns
You can call an LLM API without knowing how it was trained. But what does a falling loss mean? Why does the learning rate matter? How can one deployment process several requests at once?
Start with a tiny ticket classifier whose calculations you can do by hand. Follow its prediction through loss, gradients, and a weight update. See how to turn competing class scores into probabilities.
Use the controls in your phone or desktop browser, then explain what you learned. Sign in for AI feedback; AI requests have usage limits. No code setup is needed. Basic algebra is enough to start; we introduce derivatives when you need them. The optional Python examples assume basic Python.
Six lessons on how models learn
Start at lesson 1 or pick a subject to review. The examples are deliberately small and fictional so you can inspect each calculation; they are not production-ready models.
Loss functions
Measure how much probability reaches the right answer, then apply the same calculation to demonstrated response tokens in SFT.
Open Loss functions lesson →Gradients and backpropagation
Find which direction a weight should move, and check the derivative by making a small change yourself.
Open Gradients and backpropagation lesson →Optimizers and training loops
Compare learning rates, repeat parameter updates, and find out why a training checkpoint needs more than weights.
Open Optimizers and training loops lesson →Batches and shapes
Work out how one set of weights processes many requests, what shape comes out, and why batching can improve throughput.
Open Batches and shapes lesson →Softmax and class probabilities
Convert competing class scores into probabilities without numerical overflow. Understand why the correct class is a training target, not a random draw.
Open Softmax and class probabilities lesson →Next: build a language model
Apply these fundamentals to text in the separate Building Language Models from Scratch track. That is where n-grams, token embeddings, and attention now live.
Optional practice and references
Run the shared NumPy examples locally to see the calculations in code. For a presentation-format review, open the slides and notes or download the PowerPoint.
Choosing your next resource
Compare six AI learning resources for visual explanations, a broader ML curriculum, or deeper implementation work.