How I Think: Learning

Question

How do you learn new tech?

Surface Answer (Layer 1)

I don't "learn" in the traditional sense - I build. 12,500 AI conversations over 2.5 years = systematic pattern recognition through solving real problems, not completing tutorials.

---

The Full Answer (Layer 2)

The Pattern

Most people approach learning like this: 1. Read documentation 2. Watch tutorials 3. Complete exercises 4. Then maybe build something

I do this: 1. Identify real problem I need to solve 2. Build minimum viable solution with AI assistance 3. Encounter gaps in understanding 4. Research only what's needed to solve those gaps 5. Iterate until production-ready

Difference: I learn just enough, just in time vs everything upfront, maybe useful later

Why It Works

Traditional learning:

Front-loads information you might not need
Optimizes for breadth (know many things)
Low retention (if you don't use it, you lose it)

Building-first learning:

Loads information when you need it (high motivation)
Optimizes for depth (understand what you actually use)
High retention (learned through application, not memorization)

Example: I didn't "learn PostgreSQL" - I needed vector search for 253K messages, discovered pgvector, implemented it, encountered performance issues, learned keyset pagination, optimized. Now I understand pgvector at production level, not tutorial level.

The AI Multiplier

12,500 conversations with AI tools taught me:

How to ask questions that get useful answers
How to recognize when AI is wrong (and why)
How to translate vague requirements → specific prompts
How to iterate from working → production-ready

Key insight: AI doesn't replace understanding. It accelerates the feedback loop.

Old: Read docs → write code → wait hours/days for results → debug → repeat New: Ask AI → get code → test immediately → refine → repeat (minutes not days)

The skill isn't "knowing everything" - it's efficient iteration toward understanding.

---

Concrete Examples (Layer 3)

Example 1: Learning RAG

Didn't do: Complete "RAG from scratch" course Did do: 1. Had problem: search 253K messages semantically 2. Built basic vector search (FAISS) - worked but slow 3. Discovered pgvector - migrated, faster 4. Noticed irrelevant results - added classification metadata 5. Still getting wrong results - built query routing 6. Users needed context - added conversation expansion 7. Now: production RAG system with 19 metadata fields Result: I understand RAG at implementation depth, not tutorial depth

Example 2: Learning PostGIS

Didn't do: PostGIS certification course Did do: 1. Needed: verify if address is in Empowerment Zone 2. Found: PostGIS can do spatial queries 3. Implemented: ST_Contains(geometry, ST_Point(lng, lat)) 4. Slow: added spatial indexes 5. Fast: <50ms per address lookup 6. Production: verifying hundreds of addresses Result: I know exactly what I need for this use case, nothing more

Example 3: Learning TypeScript

Didn't do: TypeScript handbook cover-to-cover Did do: 1. Project needed: 1,220-line mapping service (nested → flat) 2. Started in Python - hit type safety issues 3. Migrated to TypeScript - caught errors at compile time 4. Learned generics when needed for type-safe mapping 5. Learned conditional types when needed for flexible interfaces 6. Now: comfortable with advanced TypeScript patterns Result: Learned through 1,200 lines of real code, not toy examples

---