How to learn a new field from scratch with AI: a machine learning walkthrough

Short answer: write your goal as what you will be able to do, and ask the AI to list the points to master in order. Then study one point at a time, answer questions yourself before reading the explanation, ask about confusing sentences on the spot, test yourself again a few days later, and finish with a small project that uses everything. Here are the six steps, walked through with “machine learning from scratch”.

Step 1: Make the goal concrete

“I want to learn machine learning” is too broad; the AI can only give you a generic overview. Write something like this instead:

Two weeks, 40 minutes a day. By the end I can explain what supervised learning does, train and evaluate a classifier with Python and scikit-learn, and tell when a model is overfitting.

A concrete goal decides what to learn and what to skip, and tells you when you are done.

Step 2: Have the AI list what to master

Ask something like:

I know basic Python and high school math. My goal is (paste step 1). List 10–12 points to master in learning order. For each, write one sentence on what I should be able to do, and flag any prerequisites I need to fill in.

You will get a list along these lines:

1. What supervised learning is: features, labels, training and test sets
2. Linear regression and loss functions
3. The intuition behind gradient descent
4. Logistic regression and classification
5. Overfitting, and what training, validation and test sets are each for
6. Regularization
7. Evaluation metrics: accuracy, precision and recall
8. Decision trees and random forests
9. Cleaning data and preparing features
10. An end-to-end project with scikit-learn

Go through it: cross out what you already know, add what is missing, push the hardest items later. This list is your mastery goal, and every later step checks against it.

Step 3: One point at a time, ending with questions

Cover a single point per session and end with 2–3 questions: one asking you to explain it in your own words, one applying it to a new situation, one asking you to spot a mistake. Write your answers before reading the explanation.

Two findings back this up:

• In Roediger and Karpicke’s 2006 experiment, students who took repeated recall tests after reading remembered 61% a week later; students who only reread remembered 40%.
• Chi and colleagues (1994) asked eighth graders reading a text on the human circulatory system to explain each sentence to themselves. They improved more than students who read the text twice, and all of the strongest self-explainers ended up with the correct model of how circulation works.

Tell the AI not to reveal answers until you have tried.

Step 4: Ask about the sentence that lost you

When a sentence doesn’t make sense, paste that exact sentence and ask: “What does ‘gradient’ mean here? Show me with an example small enough to work out by hand.” That gets to your actual sticking point faster than asking “what is gradient descent?”

Keep a list of what you asked. At the start of the next session, have the AI go over those points again before moving on.

Step 5: Test yourself again days later

Come back to each point after a day, three days and a week. Test yourself rather than rereading: have the AI write new questions with new examples, and answer them first.

Dunlosky and colleagues (2013) rated ten common study techniques. Only practice testing and distributed practice were rated high utility; rereading, highlighting and summarizing were rated low.

Step 6: Finish with a small project

Take a public dataset (the Titanic passenger survival data is a classic), then load it, clean it, train two models, compare their metrics, and write a paragraph explaining which one you chose and why. If you can do that on your own and explain it, you have met the goal from step 1.

Common traps

• Asking the AI for the answer. In Bastani and colleagues’ 2025 study, high school students who practised with a standard chat interface improved their practice scores by 48%, but scored 17% lower on the exam, once the AI was removed, than students who never had it.
• Too much at once. Three concepts in one session means none of them gets practised.
• Reading without writing. Following an explanation is not the same as being able to produce it.
• No finish line. Without a list, you can’t tell when you are done or what is still missing.

Doing it with Bloom

Bloom builds all six steps into the product:

• You write your learning goal; if it is too vague, it asks you one or two questions first.
• It drafts a syllabus of things to master that you can edit before teaching starts, at Simple, Standard or In-depth depth.
• It teaches one lesson at a time. Each ends with 2–3 reflection questions, and the next lesson opens by reviewing your answers.
• Select any sentence you don’t understand to ask about it; the answer appears beside it and shapes the next lesson.
• Each lesson starts by reviewing the last, and earlier ideas come back in new examples.
• When the syllabus is covered you get an assessment and a summary, and Learn next suggests three directions to continue.

Start your first lesson in the field you want to learn.

References

• Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.
• Chi, M. T. H., De Leeuw, N., Chiu, M.-H., & LaVancher, C. (1994). Eliciting self-explanations improves understanding. Cognitive Science, 18(3), 439–477.
• Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students’ learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58.
• Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26).