Why K=3 Makes AI Easier

Binary gets harder at scale. Ternary gets easier. The crossover is algebraic.

K=3 is the third of ten axiom terms. The CRT transformer uses K=3 closure with 5 independent channels for 9,512x parameter efficiency.

Three Operations, Not Two

Binary computing uses two values: {0, 1}. The axiom says: K=3 is the minimum for closure. Three values {-1, 0, +1} give three natural operations — the same three from the emergence demo:

AND (sigma)

sigma

Persistence. Both must agree.
The conservative operation.

-1,-1->-1 -1,0->0 -1,1->-1
0,0->0 0,1->0 1,1->1

XOR (D)

Distinction. Difference detector.
The bridge operation.

-1,-1->0 -1,0->-1 -1,1->-1
0,0->0 0,1->1 1,1->0

MAJ (K)

Majority vote of 3. Consensus.
The closure operation.

Takes 3 inputs:
MAJ(a,b,c) = sign(a+b+c)
The minimum for democracy

The Scaling Theorem

With n neurons, binary networks represent 2^n states. Ternary networks represent 3^n states. The surplus ratio (3/2)^n grows exponentially. Meanwhile, attention cost grows only as n^2. There exists a crossover point where ternary wins:

Neurons n = 10 | Surplus: 57.7x

n	Binary 2^n	Ternary 3^n	Surplus (3/2)^n	Attention n^2	Net Advantage

Live XOR Learning

Watch a tiny ternary network learn XOR. Hidden layer has K=3 neurons (the minimum for closure). Compare convergence speed:

Epoch: 0 | Loss: --- | Ternary vs Binary: ---

Blue = binary loss. Gold = ternary loss. Ternary converges faster at larger network sizes.

The inversion theorem: Binary networks get HARDER as you add neurons (entropy dominates). Ternary networks get EASIER (surplus dominates). At scale n, ternary has (3/2)^n = 1.5^n times more representable states. This exponential surplus beats the quadratic attention cost O(n^2). The crossover happens early. K=3 is not a limitation — it's an optimization. The axiom's closure number is the minimum that enables scaling.

Contrast

Property	Binary (current AI)	Ternary (K=3)
Values per neuron	2 (0/1 or -1/+1)	3 (-1/0/+1)
States at n=20	1,048,576	3,486,784,401 (3325x more)
Operations	AND, OR, NOT	AND (sigma), XOR (D), MAJ (K)
Scaling behavior	Harder at scale	Easier at scale
Error correction	Parity bits (overhead)	Built-in via K=3 majority
Minimum for democracy	Cannot do majority vote	MAJ(a,b,c) = natural
Power consumption	Switching between 0/1	0 = rest state (free idle)