The "how do I learn AI" debate, settled

Build first. Teach it back. Race only yourself.

Andrej Karpathy's three rules compress a decade of pedagogy research into one strategy. The simulator below shows why project-first learning compounds faster than bottom-up study.

1
Build real projects

Learn on demand, not bottom-up. A project creates a "knowledge pull" — every concept you learn has an immediate use, so retention roughly doubles versus passive reading.

2
Teach every concept back

The Feynman technique: explaining in your own words exposes gaps instantly. Studies on retrieval practice show explaining boosts recall by 25–50% over re-reading.

3
Compare only to past-you

Social comparison drains motivation, which is the real bottleneck. Consistency beats intensity: 1% better per day is 37× better in a year (1.01³⁶⁵ ≈ 37.8).

Simulator: 12 months of learning, two strategies

Each column is one month of skill. Toggle strategies and adjust hours per week — drag the scene to rotate. Project-first compounds because knowledge is used immediately; bottom-up decays before it's ever applied.

Skill after 12 mo:

Why the model behaves this way

The math inside the simulator

Project-first: each month's gain = hours × 1.9 × (1 + 0.06 × months of prior applied knowledge). Using knowledge compounds it.

Bottom-up: gain = hours × 2.2 up front (courses feel productive!) but 18% of unapplied knowledge decays monthly — the forgetting curve Ebbinghaus measured in 1885: without retrieval, ~60% of new material fades within days.

A concrete 30-day plan

  • Week 1: pick one project (e.g., a spam classifier for your own inbox). Ship something ugly.
  • Week 2: hit a wall (say, overfitting). NOW learn regularization — you'll never forget it.
  • Week 3: write a 300-word explanation of what you built, as if teaching a friend.
  • Week 4: compare v1 to v4 of your project. That delta is the only leaderboard.
Enjoy this tool? Build your own with Super