The learning loop — staying sharp

After this 90-day sprint, the habit that matters most isn't 'read a paper a week'. It's:

1. Build something monthly — even tiny. Code is how you internalise.
2. Write monthly — blog post, README, internal doc. Forces you to compress what you learned.
3. Teach quarterly — talk, workshop, mentor a junior. Teaching exposes every fuzzy concept.

Three to five hours a month, sustained. Beats 50 hours of passive reading every time.

Don't read every paper. Read the right one, deeply.

My filter:
1. arxiv-sanity / paperswithcode top-of-week
2. Threads from people I trust on the topic (Twitter / LinkedIn)
3. The original paper for the model I'm using right now

When reading: 3 passes.
- Pass 1 — abstract, intro, figures, conclusion. Decide if relevant.
- Pass 2 — methods, results. Skim math.
- Pass 3 — math, code, ablations. Only on the 1 paper a month you want to deeply understand.

Most papers stop at pass 1. That's fine.

Every week, log: what I learned, what I built, what I tried and failed at.

5 minutes/week. After a year, you have 50 entries. Reviewing them in December tells you exactly what you learned that year.

Beats 'I read a paper a day' as a forcing function — that turns into doomscrolling. The log measures actual outputs.

Information overload isn't a learning hack. It's a procrastination habit.

My diet:
- Newsletter: The Batch (Andrew Ng's) OR Import AI (Jack Clark)
- Podcast: Latent Space, Cognitive Revolution, or DwarkeshPatel
- Twitter list: 30 ML engineers / researchers I respect — checked once a day

Subtract more than you add. The signal-to-noise on AI Twitter is awful — a curated list helps. If a source isn't earning its slot, drop it.

Day 89 done.

- Build / write / teach
- 3-pass paper reading
- Weekly learning log
- Ruthlessly curated inputs

Tomorrow (Day 90): the wrap. Thank you, what's next, where to find me.