A researcher at the University of North Carolina at Greensboro has published what might be the most rigorous mathematical treatment of AI dominance dynamics we've seen. Nathan Langley's "Singleton Attractor" project combines three heavyweight frameworks—Yudkowsky's recursive self-improvement equation, Omohundro's instrumental resource acquisition thesis, and Lotka-Volterra competitive exclusion—into a single coupled ODE system that formally derives when a dominant agent emerges from competition.
The Core Model
The paper doesn't make the claim that one AI will inevitably win. Instead, it works backward from assumptions to theorems: if certain conditions hold (particularly Assumption A4—that the scaling parameter β can flip negative for some reachable capability level), then a singleton emerges in finite time under adversarial resource dynamics. Theorem 3.4 shows that once a leading agent crosses the β-threshold, its capability ratio over any competitor diverges in finite time. Theorem 7.1 extends this to stochastic settings using Dufresne perpetuities.
Eleven Simulations, Reproducible Results
Langley didn't just publish theorems—he shipped code. The repository contains eleven Python scripts generating all figures from the paper, with fixed random seeds ensuring bit-for-bit reproducibility across runs. Key simulation findings: in 200/200 trials with N=10 agents, the initial leader won every single time. Elimination order follows a strictly weakest-first pattern. Post-threshold moats grow from 3× to over 10^6× capability advantage within three time units.
The Coalition Problem
Perhaps most interesting for those hoping competition keeps any single actor in check: Theorem 6.2 derives an exact transcendental condition governing coalition suppression, with numerical solution α* ≈ 0.64. Finding F20 confirms that even when a coalition has sufficient combined power (N=2 suffices at α ≥ 0.64), rational defection rate is zero and the singleton still wins—the math makes coordination seem futile once threshold dynamics kick in.
The Empirical Check
Here's where it gets interesting for the doomsayers: Section 8 calibrates the model's central premise against six capability proxies (frontier compute plus five public benchmarks including GPQA, FrontierMath, and ARC-AGI). The curvature γ of log-frontier scaling is statistically nonpositive across these measures. On two benchmarks it was significantly negative—but "significantly negative" in this context means no β-flip detected. The hypothesis β(S) < 0 is not supported by the data through 2026.
Honest Limitations
The paper doesn't hide its gaps. Langley acknowledges three structural limitations: the model uses scalar capability (real optimization is multi-dimensional), the N-agent simultaneous dynamics correction lacks analytical bounds, and A4—the load-bearing assumption—is empirically unsupported so far. This is exactly the kind of epistemic honesty we need more of in AI safety research.
The Bottom Line
This is serious mathematical work on a genuinely scary question: what happens when recursive self-improvement kicks in? The good news is that Langley's empirical calibration suggests we haven't crossed that threshold yet—and may not for some time if the data holds. But the formal structure is sound, the simulations are reproducible, and the theorems don't lie. When (if?) A4 becomes true, this model tells us exactly what to expect.