The science
Each node has a scalar capability c. Self-improvement is modelled as a power law (Good 1965; Yudkowsky 2013 discusses the idea):
dc/dt = r c^α
For α = 1 this is exponential growth, c = c0 e^(rt), with doubling time ln 2 / r. For α > 1 the exact solution is
c(t) = [ c0^(1-α) - (α-1) r t ]^(-1/(α-1))
which diverges at the finite time t* = c0^(1-α) / ((α-1) r): this blow-up is a property of the equation, not a forecast. Real systems run into limits, which the model adds as a compute wall K = 10^k · size · hardware multiplier (capability saturates at K), knowledge sharing (a node moves toward stronger neighbours at rate s·w, w = exp(-d²/2σ²), where d is the distance between the nodes in the picture), and oversight damping c → c e^(-o dt).
Numerical checks (below the controls) compare the page's exact time step with independent RK4 integration of the same equation, the blow-up and wall times with adaptive RK4, the doubling time with ln 2 / r, and the damping law.
Honest simplifications. Capability is one abstract number with no units; r, α, K and o are free sliders, not measured values; there is no data, energy or algorithm bottleneck other than the wall; the network is a toy graph. Nothing here predicts how real AI systems develop, and the title idea of an all-powerful AI is not modelled: the page only studies how feedback strength, hardware limits, sharing and oversight change a growth curve. Power draw assumes one capability unit = one 700 W GPU-equivalent.