1. Many repeated attempts: Thousands of games, shots, or problems, each with a clear outcome.
2. A valid environment: The domain must have learnable, predictable patterns. Chess and physics qualify; roulette and lottery numbers don’t.
3. Timely feedback: Fast enough that you can connect action to result.
4. Deliberate practice: Thousands of hours at the edge of your ability, attacking weaknesses rather than repeating what’s comfortable.

With all four, performance looks like magic. Without them, you get “experts” who do no better than chance or novices, like many political forecasters and active fund managers.

The Expertise Loop

A simulated learner's skill (R²). Weaken any element and skill climbs slower or levels off lower.

0.85

Attempts, log scale  

1Attempts

2Valid environment

3Timely feedback

4Deliberate practice

How the model works

32 situation types, each with a true answer s_k. Each attempt picks k with p_k ∝ (error_k)^α, predicts w_k, sees y = s_k + noise, and applies w ← w + μ(y − w). With probability 1 − m the delay makes it credit the wrong situation, where m = 1/(1 + delay/1 week). V is the learnable share of each outcome. For even practice:

R²(n) ≈ V · [1 − (1 − m(1 − (1 − μ/K)ⁿ))²]

• 1 n = 0: nothing learned yet.
• 2 V = 0: the ceiling is zero, and fitting noise lands slightly below chance.
• 3 m = 0: every update hits the wrong situation, so w → m·s = 0.
• 4 comfort practice (α < 0) starves the high-error situations, so they never improve.

attempts	skill

References

• The Expert Myth / The 4 things it takes to be an expert (Veritasium)