AI Letters #03 · Engineers of AI

Two Paradigms, One Winner

Drag the compute slider to see where hand-crafted knowledge engineering loses to general learning at scale — and why it always does.

Available Compute Budget 10× baseline
1× (1990s) 10× (2010s) 100× (2016) 1,000× (2020) 10,000× (2024)
← Crossover: Scale surpasses knowledge
🧠
Knowledge Engineering
Paradigm 1
Performance curve Logarithmic
Scaling behaviour Ceiling
Early compute Efficient ✓
High compute Saturates ✗
Discovery Limited to encoded knowledge
Scale + Learning
Paradigm 2
Performance curve Power law
Scaling behaviour Unbounded
Early compute Inefficient ✗
High compute Keeps improving ✓
Discovery Discovers unknown patterns
What this means for engineers building today
1
Know where you are on the compute curve. If your compute budget is low and your domain knowledge is high, Paradigm 1 may still win. But know the ceiling exists.
2
Infrastructure investment compounds; clever features don't. The team that can run 50 experiments per week beats the team with the most elegant hand-crafted pipeline.
3
The crossover has already happened in most domains. If you're still investing heavily in feature engineering for text, images, or structured prediction, check whether you are past the crossover point.
4
The Bitter Lesson doesn't mean humans are irrelevant. It means the value moves: from encoding knowledge to specifying objectives, building evaluation, and interpreting emergent behaviour.
Interactive visualization · www.engineersofai.com