The recap
Phase 1: Foundations
1. AI learns patterns from examples instead of following hand-written rules.
2. AI contains machine learning, which contains deep learning, which contains generative AI. Nesting dolls.
3. Learning = guess, measure the error, adjust, repeat.
4. Data = rows of examples, features = clues, labels = answers.
5. Training is slow and done ahead of time; inference is the fast answer every time you use it.
6. Supervised (learn from answers), unsupervised (find patterns), reinforcement (learn from rewards).
7. Accuracy can mislead. Precision = how much to trust alarms; recall = how many real cases are caught.
8. Overfitting = memorizing instead of learning. Always test on data the model never saw.
Phase 2: Neural networks
9. A neuron multiplies inputs by weights, adds them up and decides. Layers of them = a network.
10. Loss measures wrongness, backpropagation assigns blame, gradient descent nudges every weight.
11. Embeddings turn meaning into numbers, so similar things sit close together.
12. Image models build up from edges to shapes to objects.
13. Transfer learning reuses a pretrained model; GPUs do the math in parallel.
The quiz (answers below, no peeking)
Q1. A fraud model never flags anything and is 99.5% accurate. Is it good?
Q2. Your model scores 97% on its training data and 70% on new data. Whats going on?
Q3. You have 100,000 customer records, no labels, and want to find natural customer groups. Which type of learning?
Q4. When you ask a chat assistant a question, is that training or inference?
Q5. Cheap flight and budget airfare sit close together. Close together in what?
Bonus: in Day 9s should I go for a run? neuron, which single number would you change to make it lazier?
Answers
A1. No. Its recall is 0%: it catches nothing. Accuracy hides that because fraud is rare.
A2. Overfitting. It memorized the training data.
A3. Unsupervised learning (clustering).
A4. Inference. The models weights dont change while it answers you.
A5. Embedding space: their embeddings are similar. Bonus: lower the bias, e.g. from −1 to −3.
5/5? Your'e ready for Phase 3.
What's next
Day 15: what generative AI actually is, and how it differs from everything so far.