The Three Learners #

This workbook serves three distinct types of learners. Find yourself. Follow your path.

Tyla: The CS Undergrad #

Background

• 3rd year Computer Science major
• Calculus I-II, Linear Algebra (knows the procedures, not the intuition)
• Python intermediate, some PyTorch from an ML class
• Wants to do AI safety research after graduation

Your Strength You can do the exercises. You have the math. You can code.

Your Risk You'll complete everything mechanically without understanding why. You'll pass tests without building intuition. By Chapter 1, you'll realize you memorized procedures without forming mental models.

Your Path After each section, you must answer:

1. What did this teach me about how transformers work? (Not "how to code")
2. What assumption did I make that I should verify?
3. What paper could I read to go deeper?

You don't get to proceed until you've written these down.

Your Assessment Weights

• Technical correctness: 40%
• Conceptual explanations: 30%
• Research methodology: 20%
• Capstone contribution: 10%

Aaliyah: The Bootcamp Developer #

Background

• 2 years as a JavaScript developer
• Coding bootcamp graduate—no college math
• Strong JS/TS, learning Python
• Career transition to AI, sees interpretability as the path

Your Strength You learn by doing. You can ship working code. You understand systems.

Your Risk Math notation will make you freeze. You'll see softmax(QK^T/√d_k)V and your working memory will fill with "I don't understand this" instead of the actual content. You'll skip math sections and hit walls in Chapter 1.

Your Path Code comes first. Equations come after (or never).

Every math concept has a code translation:

# The equation: softmax(QK^T/√d_k)V
# The code:
def attention(Q, K, V):
    scores = Q @ K.transpose(-2, -1)  # QK^T
    scores = scores / math.sqrt(K.shape[-1])  # divide by √d
    weights = torch.softmax(scores, dim=-1)  # softmax
    output = weights @ V  # multiply by V
    return output

Don't memorize equations. Understand functions.

Your Assessment Weights

• Working implementations: 50%
• Code quality: 20%
• Demonstrated understanding: 20%
• Capstone contribution: 10%

Maneesha: The Instructional Designer #

Background

• 8 years in corporate Learning & Development
• Statistics from a social science degree
• Basic Python, heavy Notion/Airtable user
• Core question: "AI will learn anything—what does education become?"

Your Strength You understand learning. You see the meta-level. You can design experiences.

Your Risk You'll get lost in implementation details. The trees will obscure the forest. Your working memory will fill with "how does this code work" when you should be thinking "what does this teach us about AI?"

Your Path You're allowed—encouraged—to use AI for implementation.

Your exercises are different:

• "Here's working code. Don't implement it. Instead, answer: What is attention trying to accomplish? What analogy would you use to teach it?"
• "What's the 'aha moment' that makes this click?"
• "What prerequisite knowledge does someone need?"

You're not here to become a machine learning engineer. You're here to understand what AI means for how humans learn.

Your Assessment Weights

• Conceptual understanding: 40%
• Learning design analysis: 30%
• Meta-learning documentation: 20%
• Capstone contribution: 10%

Finding Your Path #

Every chapter marks exercises for each path:

• 🎓 Tyla Path: Full exercises with reflection gates
• 💻 Aaliyah Path: Code-first explanations, math translated
• 📚 Maneesha Path: Conceptual focus, implementation optional

You can switch paths or blend them. The goal is learning, not rigid compliance.

But be honest with yourself about which blockers are real for you.