Recently, I’ve been struggling—in the best way possible—with a challenging book: Building a Large Language Model from Scratch. It is a fantastic resource packed with hands-on examples, but it is undeniably dense.
The book dives deep into concepts I either never formally studied or haven’t touched in nearly two decades: PyTorch, complex attention mechanisms, and calculus (like chain rule derivatives). My initial expectation was to power through a chapter every two to three days.
Reality hit me hard right at the beginning.
There is an appendix cheerfully titled “Learning PyTorch in an Hour.” I spent an entire week on it.
The main reason? I kept falling down the rabbit hole of brushing up on foundational math. Math was my absolute strongest subject as a student, but since I originally learned those concepts in Chinese over 18 years ago, bridging that language/symbol gap and dusting off the cobwebs was a much heavier lift than I anticipated.
Why Push Through the Friction?
Despite the steep learning curve, I didn’t give up. Pushing through this friction is worth every hour for two clear reasons:
- Customer Empathy: I want to better serve my strategic AI customers who are building the most advanced models in the industry. To help drive meaningful outcomes for them, I need to understand model architecture from the ground up, not just at an API abstraction layer.
- My Own Book Project: I am currently writing my own book focusing on AI through the lens of what I know best—data. Building a fundamental, code-level understanding of LLM internals is critical to my own writing process and technical depth.
How AI Transformed My Learning Workflow
What keeps my momentum going is how AI has completely transformed the way I learn. Instead of fighting dense material in isolation, I’ve built a 4-step learning workflow:
- Step 1: Macro View via NotebookLM: I feed the entire chapter into NotebookLM and ask it to generate an overview to give me the big picture first before diving into code.
- Step 2: Focused Manual Read-Through: I do a deep, uninterrupted manual reading of the chapter text.
- Step 3: Hands-On Manual Coding: I manually code all the exercises line by line to build muscle memory and get real hands-on experience.
- Step 4: AI Assistant as a Personal Tutor: Whenever I hit a wall with a specific concept or one-off math question, I use an AI assistant as a 1-on-1 tutor, asking it to explain the complex concept to me as if I were a beginner.
The Bottom Line
The learning journey is continuous. The tools and the form of learning have evolved tremendously for the better thanks to AI.
But at the end of the day, the core drivers—the motive, the commitment, and the human curiosity—remain exactly the same.
What dense technical subject have you tackled recently, and how has AI changed your personal learning workflow? Let’s discuss in the comments.
Annie Xu 
Leave a comment