The dominant pattern
Across many domains, the history of technology looks like repeated cycles of:
- Functional breakthrough (it works at all)
- Miniaturization (smaller, lighter, cheaper)
- Optimization (faster, more reliable, more efficient)
- Commoditization (ubiquity → new use cases)
This happens because:
- Physics rewards shorter distances (less energy, less noise, faster signals)
- Economics rewards lower marginal cost
- Reliability improves as systems shrink and simplify
Progress often comes from better representations, not smaller parts:
- Assembly → C → Python → ML models
- Hand-designed circuits → HDL → synthesis → auto-layout
- Manual biology → standardized parts → automation → ML-guided design
Nothing “shrinks,” but leverage explodes.
A more accurate model
Technology progresses by optimizing the binding between constraints:
- Physics constraints (energy, noise, thermodynamics)
- Information constraints (bandwidth, latency, representation)
- Human constraints (attention, labor, coordination)
- Economic constraints (fixed vs marginal cost)
When miniaturization stalls, the next gains usually come from:
- New substrates (biological, quantum, photonic)
- New abstractions (models that collapse complexity)
- New feedback loops (shorter sense → decide → act cycles)
Better representations
“Better representations” = better internal state descriptions of a system that make useful operations cheaper.
Better representations often:
- Reduce dimensionality
- Factor variables
- Expose invariants
- Collapse irrelevant degrees of freedom
Examples:
- Classical mechanics → phase space (position + momentum)
- Quantum mechanics → state vectors + operators
- Neural nets → latent embeddings
- Biology → pathways instead of molecules
More variables ≠ better
Right variables ≫ many variables
Towards some objective function
What is being represented?
Always one of these:
- State – what the system is right now
- Dynamics – how the state changes
- Constraints – what cannot happen
- Objectives – what matters
A representation is better if it makes at least one of these cheaper:
- Prediction
- Control
- Optimization
- Generalization
- Composition (combining subsystems)
So, towards some economic objective.