The Thesis

Defining the Question

Key Definitions

Before asking whether evolution produces intelligence, we must define our terms precisely.

Emergence

Self-replication in BFF is emergent: nothing in the instruction set "wants" to copy itself. Replication arises from the dynamics of program interaction.

Complexity

Where \(H_{max}\) is maximum possible entropy and \(H_{observed}\) is actual entropy. Complexity is constrained randomness—structure that deviates from maximum disorder.

Abstraction

Example: Seeing the pattern "all items move right" across different input grids requires abstracting over:

• Object identity (color, shape don't matter)
• Grid size (works on 3×3 or 10×10)
• Starting position (pattern holds regardless)

Intelligence

This is why GPT-4 scoring high on bar exams doesn't prove intelligence: it was trained on essentially the entire internet. The question is: can it solve novel problems with minimal examples?

Chollet's Thesis

The ARC Benchmark

"On the Measure of Intelligence" (2019) introduced ARC to test:

• Fluid intelligence: Solving novel problems, not recalling trained patterns
• Few-shot learning: 2-5 examples per task, not millions
• Core knowledge priors: Only innate human concepts (object permanence, basic geometry, counting)

Why LLMs Fail

Despite trillions of parameters, LLMs perform poorly on ARC because:

1. Memorization: They've seen similar patterns in training data
2. Interpolation: They blend known patterns, don't extrapolate
3. No search: They generate tokens sequentially, can't backtrack

The 2025 ARC winner was a 7-million parameter recursive model that beat 671-billion parameter systems. Architecture matters more than scale.

Kauffman's Thesis

The Origins of Order

Stuart Kauffman's work suggests that:

• Random Boolean networks spontaneously develop order at certain connectivity levels
• Autocatalytic sets emerge with high probability once chemical diversity crosses a threshold
• Evolution doesn't create order from scratch—it channels pre-existing self-organization

Edge of Chaos

Ordered ←——————— Edge of Chaos ———————→ Chaotic
(frozen)          (complex)            (random)
     |                 |                   |
crystals         life, minds           gas, noise
  

Systems at the edge of chaos exhibit maximal computational capacity. Too ordered: can't adapt. Too chaotic: can't maintain information. Life operates at the boundary.

The Gap

Here's the problem: Kauffman explains why complexity emerges. Chollet measures intelligence. Are these the same thing?

Possible answers:

1. Continuity: Intelligence is just complexity at scale. Given enough time and selection pressure, abstraction emerges from replication.
2. Discontinuity: Intelligence requires specific architectural innovations (cortex, attention, memory) that evolution produces but doesn't guarantee.
3. Duality: Complexity creates the substrate for intelligence, but intelligence requires additional environmental coupling.

Why This Matters

For AGI

If intelligence emerges from complexity:

• Scale might eventually work (just need more parameters, more training)
• We should study emergence, not engineer architectures

If intelligence requires specific structure:

• Scale alone won't work—we need new architectures
• Study human cognition, not just optimization

For Origin of Life

If abstraction is continuous with replication:

• Life → complexity → minds is one trajectory
• Other universes with physics allowing replication eventually produce intelligence

If abstraction is discontinuous:

• Life ≠ minds; many possible paths
• Intelligence might be rare even where life is common

For Philosophy

The question touches on:

• Reductionism: Can mind be reduced to physics?
• Emergence: Are there genuinely new causal powers?
• Understanding: What does it mean to "understand" vs "merely compute"?

Next Steps