Case Study 2: David's System One Year Later

An Honest Check-In at Month 18 of Learning Machine Learning

David is 36 years old. A year ago, he was eight months into his ML learning journey — organized, deliberate, making progress, but also somewhat fragile: his system depended heavily on the 5am alarm, and any disruption to that alarm cascaded into skipped sessions.

This is his honest assessment at month 18.

What He Still Uses

Anki, every day, without question. This has become so habitual that it no longer feels like a choice. It's closer to how he experiences brushing his teeth. He has 2,100 cards. His daily review is 45-70 cards, taking 20-25 minutes. His 30-day retention rate is 91%.

"I can't imagine doing technical learning without it now. The alternative is what I did before — read something, feel like I understood it, try to use it three weeks later, remember essentially nothing. Anki is not optional. It's the whole point of the retention side of learning."

The 5am block, modified. He still maintains the early morning learning block, but he's made it more flexible. It's now "before 7am" rather than "5-6am specifically." On days when his daughter woke at 4am, he used to skip the session because 5am had passed. Now he does 6-7am instead. A small change with a large impact on consistency.

The night-before prep ritual. Every evening, he sets up the next morning's session: opens the right notebook, writes three specific questions he wants to answer, stages any reference materials. This has not changed and is not going away. "Without it, I sit down at 5:30am and spend 10 minutes figuring out what to work on. With it, I sit down and begin."

Projects over exercises. David builds things. He doesn't complete tutorials — he uses tutorials to learn techniques he then applies to real problems he cares about. His current project is an ML system for predicting when his company's infrastructure needs scaling adjustments. This is more motivating and more effective than textbook exercises.

What He Modified

The Feynman technique: more targeted, less comprehensive. David spent his first year applying the Feynman technique to every major concept. He no longer does this for everything — only for concepts that his Anki cards reveal as shallowly understood (consistently getting them right but with a sense that he's pattern-matching rather than understanding).

"I now use it diagnostically. When my Anki card for something feels like a chant — I say the right words but I'm not sure I understand why — that's the signal to do a Feynman session."

Interleaving: formal to organic. In his first year, he formally interleaved subjects — deliberately mixing ML with statistics with calculus. Now, interleaving happens organically through his projects: a project on natural language processing naturally requires revisiting probability theory, attention mechanisms, and data cleaning all in the same week. The deliberate structure has been replaced by the natural interleaving of applied work.

Study group: expanded. David now participates in a weekly ML reading group with four colleagues at work. They take turns presenting papers — each person teaches one paper to the group monthly. The protégé effect is real and he experiences it directly. "When I have to present a paper, I understand it three times better than when I read it for my own use."

What He Dropped

Elaborate note-taking systems. David spent months in his first year building a sophisticated Obsidian vault for his ML notes — interconnected, tagged, hierarchically organized. He no longer actively maintains it.

"I spent more time organizing my notes than reading and practicing. The organization was the most interesting part, so I kept doing it. That's exactly the trap the book warned about. Now I take minimal notes in plain text and put the important things in Anki. The elaborate system was productive procrastination."

Formal weekly review sessions. He still does informal mental reviews — on his commute, he often thinks through what he studied in the past week — but the formal 45-minute weekly review he did in his first year has been replaced by this lighter, more integrated practice.

He notes: "I'm not sure this was the right call. My first year had more explicit metacognition about my learning. Now I'm more on autopilot, which is efficient but probably means I'm missing some gaps. This is something I'm aware of."

Two-hour sessions. David's 5am sessions are now typically 60-90 minutes rather than two hours. He found that his cognitive performance dropped noticeably in the second hour of complex conceptual work. Ninety minutes of genuine focus outperformed two hours of increasingly shallow focus.

What He Wishes He'd Started Earlier

Applying the Feynman technique from week one. "I did a lot of 'learning' in my first few months that was actually just recognizing vocabulary. If I'd tried to explain every concept from week one — in plain English, to an imaginary student — I would have caught the word-without-understanding problem much earlier. I probably wasted three months of learning time on shallow vocabulary acquisition."

Building projects from the beginning. "I did exercises for four months before I started building real projects. Projects are better in every way: more motivating, more efficient at identifying what you actually need to know, more connected to real applications, more memorable. I'd tell anyone learning a technical skill to build a real project from week one — small and broken at first, but real."

Accepting the beginner phase. "The first four months were hard in ways I didn't expect. I felt stupid. The people in the online course forums seemed to understand things faster than I did. I almost quit twice. What I understand now is that the beginner phase is temporary and unavoidable. You have to go through it. The learning science made it slightly less uncomfortable — understanding why the retrieval felt hard (productive struggle, not failure) helped me keep going. But I still had to go through it."

His Current System in Brief

• Anki: 45-70 cards daily, immediately after breakfast, 20-25 minutes
• Primary learning: 60-90 minute morning session, usually 6-7am, on real projects
• Reading group: Weekly, one paper presented per month
• Feynman sessions: As needed, triggered by shallow Anki card patterns
• Night-before prep: Every evening, 5 minutes
• Monthly review: First Sunday, 20 minutes — what have I built, what do I still not understand, what's next?

His Honest Assessment of Progress

At month 18, David is working on ML systems in production at his company. He's not an ML researcher, not a data scientist, and not going to become one. He is a software architect who now understands machine learning well enough to: - Design ML-integrated architectures for production systems - Evaluate proposed ML solutions for technical feasibility and appropriate use - Read and extract value from ML papers relevant to his work - Identify when ML is and isn't the right tool for a problem

"That's exactly what I set out to be able to do. I've hit the goal I set 18 months ago. Which is both satisfying and slightly strange, because now I have to set new goals."

His new 18-month goal: deep competence in reinforcement learning and large language model deployment, specifically for the systems his company is starting to build.

He writes this goal in his manifesto and sets his first three weeks of a new learning plan.

What He Would Tell Someone Starting This Journey

"The two most important things: Anki daily (or whatever spaced repetition system you'll actually maintain), and build real things from much earlier than feels comfortable.

Everything else helps at the margins. The deep work environment, the morning schedule, the Feynman technique — all of those made things better. But if you do those two things and nothing else, you will learn. If you do everything else without those two, you probably won't.

One more thing: the learning is never finished. I thought at some point I would 'know ML.' I now understand that I know some ML, that the field is enormous, and that there will always be more to learn. That's not discouraging — it's the whole point. The goal isn't to finish learning. The goal is to get good at it."