Stochastic Explosion Timing and Detection Limits

Sources#

• Humans Are Not Chinchillas: Revisiting "How Quick and Big Would A Software Intelligence Explosion Be?"
• The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion

Summary#

Chan et al. write the software intelligence explosion as one deterministic equation, dA/dt = c·A^(1+ε) with ε = λ − β = β(r − 1). With r > 1 it reaches infinity at a single date. Li, Gao, Yang & Sha (arXiv 2610.18445, 2026-10-13, UNSW / CityU Hong Kong / UT Austin) keep that drift but replace continuous progress with discrete, randomly timed improvements, and they prove what follows. The paper is mathematics with proofs and one calibration borrowed from Chan et al. It has no data. The authors say three times that its numbers "are not empirical forecasts of AI development".

There are five results, and each one bears on a different page:

1. The threshold is unchanged. The process explodes almost surely if r > 1 and never if r ≤ 1, whatever the distribution of idea sizes. In the paper's words, "randomness changes when the explosion happens, but not whether it does."
2. The date is a distribution, and idea size sets its width. Many small ideas give a tight distribution around the deterministic date. A few large ones give a wide, right-skewed distribution with more mass early.
3. Watching does not narrow the forecast. With every parameter known, the remaining time scaled by the current idea rate has the same distribution at every moment. Relative uncertainty about the time left never shrinks as the explosion approaches.
4. Detection has a cube-root law. Even an ideal observer needs about (12 z_α²)^(1/3)·ℓ^(−2/3) ideas to tell acceleration from steady exponential growth. For lumpy progress, that point arrives near the end.
5. A fixed deployment lag removes the singularity. In the delayed-feedback model, growth becomes double-exponential. Diverting effort away from capabilities moves the threshold only if the diverted share grows with capability.

The model: a "Good process"#

The model is named after I. J. Good's 1965 ultraintelligent machine. In state A, the next improvement (an "idea") arrives after an exponential waiting time with rate κA^ε, and it multiplies A by (1 + H). The sizes H are independent and strictly positive. When every idea has the same size h, the idea count is a pure birth process whose rates rise geometrically, q_n = q₀ρⁿ with ρ = (1+h)^ε. The explosion time is

ζ = Σ_{n≥0} e_n / (q₀ρⁿ), e_n i.i.d. Exp(1).

The key quantity is ℓ = ε·log(1+h), the increase in log-rate per idea. It measures how lumpy progress is.

What the model leaves out, in the authors' own list: compute and data (which Chan et al. treat as separate frictions), negative shocks, dependence between ideas, and a time-varying ε. The idea size h "is a free parameter without a direct empirical counterpart". The authors also point out what a real observer actually sees. An ideal observer here sees every effective idea and its arrival time. A real one sees manuscripts, experiments and model releases with incomplete verification. Their example is OpenAI's math repository, which listed 722 manuscripts on release and 719 a day later after a sign error. The count is of manuscripts, "not of independent, externally certified breakthroughs".

Chan et al.'s 17 months is a mean, and the blow-up is close behind it#

"Doubling time" means three different things for a jump process, and the paper keeps them apart. All its numbers use the local log drift g = log 2 / 4.5 months = 0.1540 per month, with ε = 0.39 and r = 1.386 (Chan et al.'s central values, "for illustration only"). Under that calibration the deterministic equation blows up at T* = 1/(εg) = 16.65 months. If the deterministic curve's actual first doubling is instead fixed at 4.5 months, it blows up at 19.00 months.

Remark 3.6 reads Chan et al.'s arithmetic as the mean of a random process. Chan et al. assume each doubling takes 2^(−0.39) ≈ 0.76 of the previous one, start from 4.5 months, and sum nine terms. For a Good process in which every idea doubles A (h = 1), that sum is exactly the expected time to the ninth idea, 𝔼T₉ = 17.3 months. The full series, 19.00 months, is the expected explosion time 𝔼ζ. So under their own arithmetic, the tenfold-pace point comes about 1.7 months before the model's singularity in expectation. Table 3's h = 1 row gives the time remaining after the ninth idea: a mean of 1.67 months, with a 5–95% range of 0.85–2.80 months.

The full distribution (Table 2, ε = 0.39; the rows match pdftotext -layout on the PDF and the prose):

The deterministic 16.65 months is the small-idea limit (Proposition 2.4). Its coefficient of variation is √((1−q)/(1+q)), with q = e^(−ℓ). It goes to 0 like √(ℓ/2) as ideas shrink and to 1 for a single giant idea. Equal-sized ideas are the most predictable of all size distributions with the same mean (Corollary 3.8). Lumpy progress also brings early arrival: at equal drift, a 14–20% chance of explosion within a year for h = 1 and h = 9. The authors note that this comparison, unlike the CV, depends on the calibration convention. One distribution has a closed form: Lomax-distributed idea sizes make the standardized explosion time exactly Gamma (Corollary 3.9).

Watching the explosion does not sharpen the forecast#

Theorem 4.1: at any stopping time, the remaining time multiplied by the current idea rate has the same distribution, independent of everything observed so far. Absolute uncertainty in calendar time does shrink, in proportion to the current expected waiting time. Relative uncertainty stays fixed. The authors draw the planning consequence: a plan keyed to this model "is naturally indexed to the current expected doubling time rather than to a calendar date".

A worked lead-time rule (Remark 4.3), for h = 1 at ε = 0.39. The expected remaining time is 4.22 × m_t, where m_t is the current expected doubling time. Its 5% and 95% quantiles are 2.148 × m_t and 7.098 × m_t. A response that needs twelve months and must be in place with 95% probability has to start while m_t ≥ 12 / 2.148 = 5.59 months. All of this assumes the model and its parameters are known.

Log-periodic precursors are real but invisible. Exact discrete scale invariance produces Sornette-style log-periodic oscillations in idea density near the singularity. Proposition 4.5 gives their amplitude exactly. It is 3.4 × 10⁻¹⁵ for doubling-sized ideas, 3.1 × 10⁻³ at ℓ = 1.2, and reaches 11% only when each idea multiplies the rate of progress tenfold. It is also an ensemble statement centred on an explosion time that no observer knows in advance, so it is not a detection procedure.

Uncertainty about r makes the expected wait infinite#

Proposition 4.6: if the growth exponent ε is uncertain and its prior has positive density at 0 (the critical value r = 1), then conditional on explosion, P(ζ > t) decays only like c/t. The conditional mean waiting time is infinite, even though explosion is certain given ε > 0. Intrinsic randomness alone has an exponential tail. Parameter uncertainty produces the heavy tail.

The calibration fits a log-normal prior on r to Ho & Whitfill's 90% interval of 0.727–2.094, the one Chan et al. quote. With β = 1.01 this gives P(r > 1) = 0.743. Conditional on r > 1, the median deterministic waiting time is 17.3 months, P(wait > 5 years) = 0.150, P(wait > 10 years) = 0.074, and the tail constant is 8.67 months. The authors add that these figures "validate the calculation under the selected prior; they do not validate the prior or the underlying interval."

A second prior on r is more pessimistic, with its median at the critical point. Tillman's revision of Davidson & Houlden's size model (Forethought, 2026-10-05, practitioner-opinion; full treatment on Software-Driven Intelligence Explosion) draws r log-uniform on 0.33–3. Its median is 1.0 and P(r > 1) ≈ 0.50, against 0.743 for the log-normal above (wiki arithmetic). Its density at r = 1 is positive, so Proposition 4.6's infinite conditional mean applies to it too. A heavier tail does not follow. Per unit of log r, the log-uniform puts about 0.45 of density at r = 1 and the fitted log-normal about 1.0, so the log-uniform is flatter there even though its median sits on it (wiki arithmetic; the paper computes no tail constant for this prior). The two priors differ in r's definition as well as its spread: Davidson & Houlden's r is an adjusted start-of-explosion return, and Ho & Whitfill's is a labor-only historical fit.

Detection: the cube-root law#

Chan et al.'s first policy priority is visibility, and part of it is detecting onset (Frontier Pause Verification). The paper studies the most favourable case. The monitor knows the baseline idea rate and the size of the acceleration ℓ. It sees the first n waiting times exactly. It makes one decision between H₀, steady exponential progress (r = 1), and H₁, rates rising by e^ℓ per idea from the start.

Theorem 5.1: the minimum sample size at error level α satisfies n* ~ (12 z_α²)^(1/3) · ℓ^(−2/3). The constant is ≈ 3.19 at α = 0.05, and there are non-asymptotic bounds on both sides. The authors' one-line explanation: the log-rate drifts by ℓ per idea, so after n ideas the KL divergence is of order ℓ²n³, and discrimination needs that to reach order one.

What it means for lead time (Table 3, α = 5%, reconciled against the PDF):

W* is the expected remaining duration at the decision divided by the initial expected explosion time. Granularity sets the warning. Small-idea explosions become distinguishable while most of the expected duration is still ahead, though W* approaches 1 only slowly, at rate ℓ^(1/3). For doubling-sized ideas, the earliest reliable decision coincides with Chan et al.'s tenfold point, and about 1.7 months remain.

The caveats are part of the result. The test is not an anytime alarm, and its error bounds "must not be reused for repeated testing after every idea". It does not handle an unknown onset. If ℓ is unknown and only ℓ > 0 is specified, no fixed-sample test has a nontrivial uniform error bound (Remark 5.4). An unknown baseline rate plausibly raises the constant to (48 z_α²)^(1/3), about 1.59× as many ideas. The authors have not proved that (Remark 5.5). By data processing, any coarsened record, such as counts of manuscripts or releases, can only do worse. A real monitor with information outside the model might do better.

Model extensions: delay, diversion, and an oversight race#

These are three separate models, and the authors warn they are "not properties of one unified controlled stochastic system".

A deployment delay removes the singularity (Theorem 6.1). Suppose effective labour depends on capability deployed D months ago, dA/dt = c·A(t)^(1−β)·(k·A(t−D))^λ. Then any fixed D > 0 rules out finite-time blow-up for any r. If r > 1, growth becomes double-exponential: log log A(t) / t → log r / D. With D = 3 months and r = 1.386, log A doubles about every 6.37 months. The stochastic version is proved only not to explode. This formalizes Chan et al.'s remark that some research steps are inherently sequential. On Domestic Frontier Pacing, the AI Futures Project's 9-month capability lag has the same structure. Plugging D = 9 months into the same formula gives a log-A doubling time of about 19.1 months (wiki arithmetic, not the paper's).

Diversion moves the threshold only if it rises with capability (Proposition 6.2). If the share kept for capabilities research falls as (A/A₀)^(−γ), explosion happens if and only if γ < 1 − 1/r, which is 0.279 at the central values. Diverting 47% of effort at ten times today's capability and 85% at a thousand times turns the singularity into exponential growth. A constant diverted share θ only postpones the blow-up, since T* scales as (1−θ)^(−λ).

Oversight keeping pace depends on exponents, not on budget (Theorem 6.3). Let oversight capability S be produced by the diverted labour under its own idea production function. As A blows up, the ratio S/A goes to 0, to a finite limit, or to infinity according to whether oversight's own exponent ε_S is below, equal to, or above ε. A fixed allocation share changes constants and dates but not the regime, except at exact equality. Example 6.4 shows that removing the singularity is not the same as oversight keeping pace: the diverted share goes to one and S/A still goes to 0. The authors' policy reading is that oversight research's returns to labour λ_S and its difficulty growth β_S "deserve measurement alongside its budget". They also caution that nothing here shows fixed safety spending is useless, and that treating S/A as oversight adequacy is a modelling assumption.

Where it meets the rest of the corpus#

• It refines Software-Driven Intelligence Explosion without contradicting it. It uses the same equation and the same r = 1 threshold, read as a random process. Chan et al.'s ~17 months to tenfold is the mean time to the ninth doubling for doubling-sized ideas. Their own geometric series implies a singularity about two months later, and in the stochastic version its 5–95% range runs from 9.7 to 31.9 months.
• It answers part of the forecasting question on Intelligence Explosion Dynamics, mostly in the negative. Three limits apply even inside the friendliest model. Relative uncertainty about time remaining never shrinks. Telling acceleration from steady growth needs a number of ideas that scales as ℓ^(−2/3). Uncertainty about whether r exceeds 1 makes the expected wait infinite. Fitting early-onset data cannot beat these limits. What can help is information the model lacks.
• It agrees with Brown's serial-experiment brake and Chan et al.'s time-intensive friction, in theory. A fixed lag in the feedback loop turns the hyperbola into a double exponential. That is very fast growth with no finite-time singularity. It is not an S-curve, which still needs saturation (Remark 3.3: any saturating rate function prevents explosion).
• It gives Domestic Frontier Pacing a theorem for two of its options. Option 3's lag works as a delay D and removes the blow-up in this model. Option 2's capabilities share is a fixed allocation. In the paper's terms, a fixed allocation stretches T* but leaves the regime unchanged. The mapping from the proposal's compute shares to the model's labour shares is the wiki's, not the paper's.
• It puts a number on the detection half of AGI-to-ASI Pathways' compounding question, for a single loop. It says nothing about several pathways compounding together.

Connections#

• Software-Driven Intelligence Explosion: the deterministic r > 1 model this page randomizes. Its 17 months to tenfold is 𝔼T₉ for doubling-sized ideas, and the singularity comes about two months later
• Intelligence Explosion Dynamics: the curve-shape debate. The hyperbolic regime gets a timing distribution, and early detection gets a sample-size law
• AGI-to-ASI Pathways: pathway 3's early-detection question, answered for one loop under ideal observation
• Domestic Frontier Pacing: the capability lag as a delay that removes the singularity, and the compute-allocation floors as a fixed share that only postpones it
• Frontier Pause Verification: the visibility priority. Detecting onset statistically needs many ideas when progress is fine-grained, and comes late when progress is lumpy
• Recursive Self-Improvement: the mechanism whose growth law is modelled here
• Noam Brown: the serial-experiment brake, which the delay theorem formalizes

Open Questions#

• Does the cube-root law survive an unknown baseline rate and an unknown onset? The authors conjecture (48 z_α²)^(1/3) for the first case and leave anytime alarms with false-alarm control open. A real monitor faces both, plus an unknown ℓ, for which no uniform fixed-sample test exists.
• What is the empirical idea size h, the lumpiness of AI-R&D progress? It decides both the width of the timing distribution and the warning time, and the paper calls it "a free parameter without a direct empirical counterpart". A lab's internal series of capability gains per improvement would bound it.

Sources#

• The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion: Xiaoyu Li, Zheng Gao (UNSW), Chiwun Yang (CityU Hong Kong) & Zhizhou Sha (UT Austin), The Singularity Has a Distribution: Exact Timing Laws and Detection Limits for a Stochastic Intelligence Explosion, arXiv 2610.18445, 2026-10-13, 25pp (practitioner-opinion, tier corrected from prediction at compile). Theorems 3.1, 3.4, 4.1, 5.1, 6.1, 6.3, Propositions 4.5, 4.6, 6.2, Remarks 3.6, 4.3, 5.4, 5.5, and Tables 2–3. Calibration borrowed from Chan et al.; illustrations "not empirical forecasts". Code and every number are in a public repository (github.com/xiaoyulics/singularity-distribution). The authors disclose AI assistance in drafting and in the numerical code. Parse note: PDF-derived (docling 2.126.0 + mlx). Display equations are sound. Inline math lost its superscripts and subscripts (e.g. "( 12 𝑧 2 𝛼 ) 1 / 3 ℓ - 2 / 3"), so formulas here come from the abstract and the display equations. Tables 2 and 3 were reconciled cell by cell against pdftotext -layout and against the prose (the h = 1 row: 17.3 + 1.67 ≈ 19.0 = 𝔼ζ). The parse-time table-collapse flag on "1. 2 × 10 - 13" is a false positive. The theorems in §6.3 and Example 6.4 were garbled by the parse; only their prose statements are used