What I’ve learned, and how deeply.

Five-course sequence covering fundamentals through sequence models. Completed all programming assignments in raw numpy first, then re-implemented in PyTorch to feel the difference. The convolution and RNN backprop derivations are the ones that stuck.

Worked through the entire book over two months, typing every line and running experiments at each checkpoint. Fine-tuned a small model on my own writing corpus at the end.

Counter-programming to the usual math-first approach. Built an image classifier that outperformed my earlier Coursera version in a weekend. The philosophy reshaped how I learn.

Four months of evenings. Passed on the first attempt. The ethics section mattered more than expected — it's not rules, it's how to think about conflicts of interest.

Convex analysis through interior-point methods. Final project modeled portfolio construction as a second-order cone program. The convex duality chapter reshaped how I see modeling problems.

Read the first fifteen chapters carefully, skimmed the rest. Worked through binomial tree and Black–Scholes derivations by hand until I could reproduce them from memory. Wrote a small options pricer in Python afterward.

Shipped five production projects across startups and research. Comfortable with advanced generics, conditional types, and designing type-safe APIs. Contributed to two open-source TypeScript libraries.

Read cover to cover over six weeks, with detailed notes. Rewrote the consistency-models chapter as a tutorial for interns. Stream processing reshaped how I think about any real-time system.

Worked through twenty case studies, then designed three of my own for systems I actually use daily. Can now reason about read/write ratios, cache strategies, and partition keys without looking things up.

Redesigned three of my past projects using the book's principles. The section on spacing scales and color systems is my most-referenced resource. I reread parts of it every quarter.

Use Figma daily. Built a design system with 40+ components for a side project. Comfortable prototyping complex interactions and handing off to engineering with clean specs.

Hands-on workshop on the history and mechanics of type. Drew letterforms on paper, then digitized them. I'm nowhere near a type designer, but I now notice kerning, leading, and x-heights everywhere.

Covered event schema design, attribution models, and the difference between GA4 and Universal Analytics. Set up analytics on two personal projects end-to-end. Funnel analysis changed how I design landing pages.

Eight-week cohort with PMs from major product companies. Applied the frameworks to a friend's early-stage startup as a practicum. Habit-loop decomposition was the single most useful tool.

Hamming's Bell Labs lectures, revised for publication. The chapter on choosing research problems should be required reading for anyone early in a technical career. I reread it every six months.

Read in a weekend, re-read three times over the next year. The parts on monopoly, secrets, and the contrarian-question framing rewired how I evaluate my own ideas.

Half memoir, half the psychological framework Frankl built from surviving the camps. The core idea — that the last freedom is the freedom to choose your response — is something I come back to whenever I'm stuck.

The latticework-of-mental-models framing is the single most useful lens I've picked up in two years of reading. Still working through which models I actually use vs. which I just know exist.