Skip to content

English · Español

Lab 04 — Final reflection

🇪🇸 Después del postmortem técnico, una reflexión personal. No para el público; para futuro-Borja. ¿Qué cambió en mi cabeza? ¿Qué hábito quiero conservar? ¿Qué error no quiero repetir?

Objective

Produce learners/borja/phase-40/reflections.md: a personal reflection on the 40-phase journey. Not a postmortem (that was lab 01). Not a reading list (that was lab 02). This is the subjective half — what changed in Borja's head, what habits formed, what surprised, what disappointed.

Setup

  • Every learners/borja/phase-NN/reflections.md from phases 0-39.
  • Every learners/borja/journal/YYYY-MM-DD.md entry.
  • The postmortem from lab 01 (technical perspective) — to contrast with this lab's subjective perspective.

Tasks

Five short sections, each ~150-300 words.

1. What changed in my mental model

Before the project, Borja's mental model of LLMs was X. After 40 phases, it's Y. The delta is the lesson. Be specific:

  • "Before: 'attention is some magic.' After: 'attention is a learned weighted average; the weights are softmax over query-key dot products; complexity is O(T² · d).'"
  • "Before: 'fp16 vs fp32 is a memory trick.' After: 'fp16 has a 5-bit exponent; overflow is real at scale; bf16 has a 7-bit exponent and the same range as fp32; this matters for stable training.'"

Aim for 5-10 such "before / after" pairs.

2. Habits that formed

What did Borja start doing during the journey that's worth keeping?

Candidates: - Writing BLUEPRINT.md before code. - Running manifest.json for every experiment. - Writing the test list before the test bodies. - Bilingual summary writing (English doc + Spanish gloss). - Reading the spec before starting each phase. - Asking subagents (math-reviewer, numerical-stability-checker) for second opinions. - Daily journal entries.

For each kept habit, write: (a) what it is, (b) one concrete time it caught a mistake or saved time, © whether it should be a default in the next project.

3. Habits that didn't stick

Equally important. What did Borja try to keep up but couldn't, and why?

Examples: - "I wanted to write a journal entry every day. I averaged 4/7. Reason: weekends were off; the ritual felt forced when there was nothing to log." - "I wanted to run /quiz weekly. I did it once a phase. Reason: the quiz tool overhead vs. the marginal value didn't pencil out for me."

Be honest. The point is to design the next project to either drop these aspirational habits or make them more tractable.

4. What surprised me

Surprises are the highest-signal lessons. Three to five concrete surprises:

  • "I expected the KV cache to be a small optimization. It's the entire architecture of modern LLM serving."
  • "I expected np.add.at to be a niche function. It turns out to be the linchpin of correct embedding gradients."
  • "I expected continuous batching to improve throughput. It mostly improves latency tail. Different problem."
  • "I expected mypy --strict to slow me down. It probably saved me an entire phase of debugging across the project."

The format: one sentence stating the expectation, one sentence stating what actually happened.

5. The one sentence

Distill the whole journey into one sentence. It will be wrong, but the act of distilling forces choosing what matters.

A few candidates Borja might write:

  • "Building the small thing carefully teaches more about the big thing than building the big thing carelessly."
  • "Every framework is a tradeoff hidden behind an API; doing it by hand once lets you see the trade."
  • "An AI system is a chain of decisions; the chain only holds if every link is examined."

Pick one. Live with it.

Deliverable

learners/borja/phase-40/reflections.md, ~1000 words total, in the five-section structure above.

Acceptance

  • File exists at the right path.
  • Five sections, each ~150-300 words.
  • "What changed" has ≥ 5 before/after pairs.
  • "Habits that formed" has ≥ 3 entries with concrete examples.
  • "Habits that didn't stick" has ≥ 2 entries — honest about what failed.
  • "What surprised me" has ≥ 3 entries.
  • "The one sentence" is a single sentence (no padding).

Pitfalls

  • Making this section public. This is a personal reflection. It lives in learners/borja/, not in docs/. It's not for an audience.
  • Conflating with the postmortem. The postmortem is technical and structural. This is subjective and personal. Different documents, different jobs.
  • "Everything was great." If the whole reflection is positive, you're not being honest. Find the things that didn't work.
  • "Everything was hard." Same in reverse. Acknowledge what worked, including dumb luck.
  • Skipping section 3. It's the most uncomfortable section, and the most useful.

Stretch

  • Read this file at 3-month, 6-month, and 12-month intervals after Phase 40. Add a footer entry each time with what's still true and what isn't.

End of Phase 40 labs. After this:

  1. Write PHASE_40_REPORT.md.
  2. Flip the ROADMAP row 40 to 🟢.
  3. Commit.
  4. Stop.