Three scaling routes

capabilitycost
Complete comparison: the black current state, a purple continuation, a green pointwise transformation of the black frontier, and the blue outer envelope generated from multiple black-frontier seeds.
Overview

Three ways to scale intelligence

One frontier. Three routes.

Purple continues the current state. Green transforms that frontier point by point. Blue is the outer envelope induced by a new scaling axis.

01 Continue02 Improve03 Change
Route 01

Continue scaling up.

Raise the ceiling.

Spend more parameters, data, and compute to continue along the same logarithmic frontier—each linear capability gain demands an exponentially larger cost.

Parameters+Data+Compute
Route 02

Better trade-off.

Reach the same ceiling sooner.

Transform the current-state frontier point by point. Every intermediate operating point improves, while the green curve remains attached to the same final endpoint.

R2RMoAFrameFusion
Route 03

New scaling axis.

Change the curve.

Each blue branch is another instance of the logarithmic scaling form, initialized from a model on the black frontier. Their upper envelope keeps the best instance at every budget.

C(p,s) = C₀(p−s) + C₁(s)C₀(x) = α log(1+x/τ) · C₁(s) = α log(1+s/τ)shared α and τ · max over 0 ≤ s ≤ p
TaHC2C