Recursive Criticality of AI Self-Improvement
Mikhail Burtsev
cs.AI
2026-09-01
Burtsev sets self-amplifying AI R&D at R_AI=χa/σ>1. No-recursion AGI is 24 years, AGI-to-ASI 72; strong recursion compresses the gap to 0.45 years. Scenario outputs, not forecasts.
AI systems already write code, run experiments, and reproduce papers, and that work sits inside the R&D process that builds the next system. Is that recursive self-improvement? Most arguments pin the critical point to a capability level, as if AGI arrival would detonate a takeoff. Burtsev moves the critical point from a score to a feedback structure: whether one increment of capability produces enough extra R&D productivity in later cycles to outrun the fact that further progress gets harder.
He borrows a reproduction number RAI from epidemiology. Above 1, increments amplify across cycles; below 1, they damp. Speed can come from spending and compute. Self-amplification is a different property. A system can cross the threshold before acceleration is visible, and it can grow quickly without ever crossing.
The state x is a resource-cost coordinate on a fixed task panel, not an IQ scalar. The dynamics split into four pieces:
Realized recursive gain is g=χa. Delay τ is the end-to-end cycle before an improvement returns in a successor. The minimal model is ẋ(t)=r(t) f[x(t)] exp(Φ(x[t-τ],t)), with ∂Φ/∂x = g. Hardening is σ = −d ln f / dx.
Linearizing around a reference trajectory yields the local threshold RAI = g/σ = χa/σ. All characteristic roots have negative real part when RAI<1; a positive real root appears when RAI>1. Baseline throughput v scales the damping term and the feedback term together, so raising r steepens the curve in the minimal model without by itself pushing the system across the threshold. Near criticality the dominant root is λ ≈ vσ(RAI−1)/(1+vσ RAI τ). At high throughput, λ approaches ln(RAI)/τ, so cycle length caps how fast amplification can run.
A finite frontier f(x)=(1−x/X)^β gives σ=β/(X−x). As x approaches X, σ diverges and RAI falls to 0, so a supercritical spell can end on its own. With multiple actors, each row of K has Kij=gij/σi, and the network reproduction number is the spectral radius of K. Every actor can be subcritical and the ecosystem still supercritical if cross-actor transfer is strong enough.
Every number is a scenario demonstration. The author says they are not forecasts. Current capability is set to 0, the effective frontier to X=1, and illustrative AGI/ASI thresholds to 0.50/0.80. The no-recursion baseline is calibrated to a median expert AGI year of 2050, so from 2026, rref=1/24≈0.0417 per year: AGI in 24 years, ASI in 96. β=2, χ is logistic with k=10 so χ(0)=0.5, delay τ=0.5 years, and only recursive gain a is swept.
| Scenario | a | AGI (yr) | ASI (yr) | AGI→ASI |
| No recursion | 0 | 24.00 | 96.00 | 72.00 |
| Smooth scaling | 0.5 | 21.35 | 74.13 | 52.79 |
| Weak supercritical | 3 | 12.90 | 24.73 | 11.83 |
| Transient takeoff | 6 | 8.40 | 11.08 | 2.69 |
| Rapid transition | 15 | 4.50 | 4.95 | 0.45 |
Subcritical smooth scaling still pulls AGI forward by about 2.7 years and ASI by about 22. Rapid transition pulls AGI by about 19.5 years and ASI by about 91, and compresses AGI-to-ASI from 72 years to 0.45. Weak supercriticality only briefly exceeds 1 before AGI, then hardening pulls it back; the capability lead accumulated in that window is not undone.
Three organizational scenarios start from the weak-supercritical reference:
| Structure | network ρ(K) | AGI | ASI | gap |
| Closed labs | 1.01 | 10.40 | 16.53 | 6.13 |
| Open ecosystem | 1.55 | 9.75 | 12.58 | 2.82 |
| Global competition | 1.15 | 8.97 | 13.42 | 4.45 |
In the open ecosystem every actor is subcritical (largest diagonal 0.65); transfer lifts the network to 1.55 and produces the shortest gap. Global competition uses effort multiplier 1.2 and delay 1 year, reaches AGI first (about 9.0 years), and still takes 4.4 years AGI-to-ASI, longer than the open case. On hardware, assume compute demand rises 100× from AGI to ASI and physical compute grows 25% per year: physical headroom F=8 at the unconstrained AGI crossing does not bind; F=2 binds between AGI and ASI; F=0.75 binds before AGI.
The usable distinction for practitioners is that a steep benchmark slope is not self-amplification. Diagnosing recursive criticality means estimating χ, a, σ, τ, and how improvements move between organizations. Extra compute and headcount mainly change speed in the minimal model; they do not by themselves change whether the system is over the threshold. Shortening the development cycle and raising the share of AI output that actually lands in the next model are what move RAI. Open diffusion can push an ecosystem over 1 even when no single actor is.
This is a conceptual frame plus a parameter scan. a, β, and k are not directly measured, and the years in the tables are not a calendar.
The author treats rref and τ as illustrative calibrations, and a, β, k have no empirical anchors. The capability coordinate depends on a fixed task panel and fixed resource prices; change the mix and x is a different quantity. The 0.50/0.80 AGI/ASI marks are there to separate two dynamical regimes on one axis, not operational definitions. The multi-actor result is stated in a zero-delay approximation. Physical constraints are deliberately split off from the recursive dynamics, so the pathway where compute shortens cycles or raises closure is not in the main equation. The conclusion typesets RAI=χ/aσ, which is a typo; the boxed formula in the body is χa/σ.
The sharper gap is causal identification. MLE-bench, RE-Bench, and PaperBench measure skill on research tasks. They do not measure the causal effect of this generation's research capability on the next generation's R&D productivity. Without that measurement, RAI remains a symbol.