📈 economics

The Dynamics of Intelligence Explosions

This paper mathematically analyzes the dynamics of AI-driven intelligence explosions, demonstrating that singular growth is more difficult to achieve than previously thought and highlighting generation time as a critical, often neglected factor that determines whether feedback loops result in a vertical asymptote or faster-than-exponential but finite growth.

Original authors: Toby Ord

Published 2026-08-17

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Toby Ord

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a car race where the cars are getting better at building themselves. In the world of artificial intelligence, there is a fascinating idea called an "intelligence explosion." It's the concept that if a computer gets smart enough to help design the next, even smarter computer, that new computer could help design an even better one, and so on. This creates a feedback loop, like a microphone getting too close to a speaker and creating a screeching feedback noise, but instead of noise, it's a rapid explosion of intelligence. Scientists have been trying to figure out exactly how fast this could happen. Would it be a steady climb, a rapid sprint, or a sudden, impossible jump to infinity in a blink of an eye? This is the question that mathematician Toby Ord tackles in his paper. He looks at the math behind these self-improving machines to see if they can truly reach a "singularity"—a point where intelligence grows so fast it hits a vertical wall in a finite amount of time, or if there are other, slightly slower, but still incredibly fast, ways this explosion could play out.

The Race to Infinity: Why the "Jump" Might Be Harder Than We Think

Imagine you are baking a cake, but instead of using a timer, you have a magical oven. Every time you bake a cake, the oven gets smarter and builds a better oven for the next round. The first oven takes 10 minutes. The second, being smarter, builds a third one in 5 minutes. The third builds a fourth in 2.5 minutes. If this keeps happening, and the time it takes to build the next oven keeps getting cut in half, you might think you could bake an infinite number of cakes in just a few hours. This is the dream of an "intelligence explosion": a machine that keeps making itself smarter, faster and faster, until it surpasses all human understanding in a flash.

For a long time, many experts thought this explosion would look like a "vertical asymptote." In math terms, imagine a graph where the line shoots straight up, like a rocket hitting the sky, reaching infinite height in a finite amount of time. It's the ultimate "boom." But Toby Ord's paper suggests that while this vertical rocket is possible, it's actually much harder to build than we thought. In fact, there's a whole other category of explosions that are incredibly fast—faster than exponential—but they don't shoot straight up to infinity. They just zoom really, really fast without ever hitting that vertical wall.

The Secret Ingredient: The "Generation Time"

To understand why the vertical rocket is so hard to build, we need to look at a hidden variable that most people forget: Generation Time.

Think of the feedback loop like a relay race. The first runner (AI 1) passes the baton to the second runner (AI 2), who passes it to the third (AI 3). The "generation time" is how long it takes for one runner to finish their leg and get the next runner ready.

In the old, simple math models, people assumed this relay happened instantly, or at least at a constant speed. They thought if the runners got faster and faster (more "intelligence"), the race would just speed up to infinity. But Ord points out a crucial flaw: You can't run an infinite number of laps in a finite amount of time unless the time it takes to run each lap shrinks to zero.

Imagine you are running a race where you have to complete an infinite number of laps.

Scenario A: You run each lap in 10 seconds, then 9, then 8. You get faster, but you still spend 10 seconds, then 9 seconds. No matter how fast you get, you will never finish an infinite number of laps in just 10 minutes. You just run out of time.
Scenario B: You run the first lap in 10 seconds, the second in 5, the third in 2.5, the fourth in 1.25. Here, the time for each lap is shrinking rapidly. If you keep cutting the time in half, you can actually finish an infinite number of laps in a finite amount of time (like 20 seconds total).

Ord's paper shows that for an intelligence explosion to hit that "vertical asymptote" (the singularity), the generation time (the time to build the next AI) must shrink toward zero incredibly fast. It's not enough for the AI to just get smarter; the process of making the next AI must get impossibly fast.

The "Middle Ground" Explosion

Here is the twist: The paper suggests that while the "vertical asymptote" is possible, it's actually quite rare. There is a huge, neglected middle ground.

Imagine a growth curve that is super-exponential. This means it's growing faster than a standard exponential curve (like compound interest). It's zooming up the chart! But, because the generation time doesn't shrink fast enough to hit zero, the line never goes vertical. It just keeps climbing, faster and faster, forever, but it never hits that "infinity in finite time" wall.

Ord calls this "sub-singular" growth. It's still terrifyingly fast. It's still an explosion. But it's not the "magic vertical jump" that some models predict. It's more like a rocket that accelerates so hard it feels like it's going to break the universe, but it never quite reaches that impossible vertical cliff.

Why the Math Matters (and Why It's Tricky)

The paper uses some fancy math to prove this, but the core idea is simple: You cannot bypass the constraints of time.

In the old models, people used equations that assumed the AI could improve its own speed instantly. But in the real world, building a new AI takes time. You have to design it, train it, and test it. Even if the AI is a genius, it still takes time to write the code or run the experiments.

Ord shows that if you take this "time" seriously (using what he calls "time-embedded difference equations"), the conditions for a singularity become very strict. You need two things to happen at the same time:

The AI must keep getting better and better without ever stopping (unbounded growth).
The time it takes to make the next version must shrink toward zero very quickly (the "Zeno condition").

If the time to make the next AI hits a floor—say, it takes at least 1 hour to train a new model no matter how smart the AI is—then the vertical explosion is impossible. The growth will just become a very steep, very fast exponential curve, but it won't be a singularity.

The Trap of Measurement

The paper also warns us about how we measure "intelligence." Imagine you are measuring how good a chess player is.

Measure A: How many games they win in a row.
Measure B: How many seconds they take to make a move.

If an AI gets so good that it never makes a mistake, Measure A might go to infinity (infinite wins). But Measure B might just hit a limit (it can't move faster than light). The paper argues that some popular ways of measuring AI progress might look like they are hitting a "singularity" (going to infinity) just because of how we measure them, not because the AI is actually becoming infinitely smart. It's like a coordinate singularity in physics: a glitch in the map, not a real cliff in the terrain.

The Bottom Line

So, what does this all mean for the future?

Toby Ord isn't saying the explosion won't happen. He's saying that the "vertical asymptote" (the instant jump to infinite intelligence) is much harder to achieve than the simple math models suggested. It requires the AI to not only get smarter but to make the process of getting smarter impossibly fast, shrinking the time between generations toward zero.

Instead, we are more likely to see a "super-exponential" explosion. This is still a period of incredibly rapid change where AI capabilities double and triple in speed, potentially outpacing our ability to control them. It's a "fast lane" explosion rather than a "vertical jump."

The paper suggests that if we want to understand the future of AI, we shouldn't just look at how much smarter the AI gets. We need to watch the generation time. How long does it take to build the next version? If that time stops shrinking, the vertical explosion stops. If it keeps shrinking, we might be in for a ride that is even wilder than we thought.

In short: The explosion is real, but the "vertical cliff" might be a mirage. The real danger is a speed that is so fast it feels like a cliff, even if it's technically just a very steep hill. And that's still enough to change the world forever.

Technical Summary: The Dynamics of Intelligence Explosions

Problem Statement
The paper addresses the mathematical dynamics of AI R&D Automation (AIRDA), a feedback loop where AI systems assist in designing and building subsequent, more capable AI systems. While the concept of an "intelligence explosion" (coined by I.J. Good) suggests that such feedback could lead to rapid, potentially unbounded growth in capabilities, existing economic-inspired models often conflate different types of explosive growth. Specifically, standard models frequently assume that any super-exponential growth necessarily leads to a "finite time singularity" (a vertical asymptote where capability $A$ reaches infinity in finite time $t^*$ ). The author argues that this assumption overlooks critical physical constraints, particularly the discrete nature of feedback loops and the time required to complete them (generation time), leading to an overestimation of the likelihood of singular growth.

Methodology
The paper employs mathematical modeling to analyze the dynamics of AIRDA, progressing through three distinct modeling frameworks:

Differential Equations: The author first reviews standard continuous models of the form $\dot{A} = f(A)$ . It analyzes the conditions under which $A(t)$ exhibits super-exponential growth versus singular growth (finite time blow-up). The analysis distinguishes between the "blow-up condition" (convergence of the integral of the reciprocal of $f(A)$ ) and simple super-linearity.
Difference Equations: Recognizing that AI R&D occurs in discrete steps rather than continuous flows, the paper shifts to difference equations ( $\Delta A_n = f(A_n)$ ). It demonstrates that in pure discrete time, singular growth is mathematically impossible regardless of the growth rate of $f(A)$ , as a finite number of steps cannot reach infinity.
Time-Embedded Difference Equations: To bridge the gap between discrete steps and continuous time, the author introduces a model where the generation time ( $T_n$ ), or the duration of the $n$ -th feedback loop, is a variable that can shrink over time. This framework allows for the analysis of how the rate of loop completion interacts with the magnitude of capability gains.

Key Contributions and Results

Decoupling Super-Exponential and Singular Growth:
The paper demonstrates that super-exponential growth (growth faster than any exponential function) does not necessarily imply a finite time singularity.
- In continuous models ( $\dot{A} = f(A)$ ), functions like $f(A) = A \log(A)$ produce doubly exponential growth but fail the "blow-up condition," meaning $A(t)$ never reaches infinity in finite time.
- In discrete models, singular growth is impossible unless the generation time $T_n$ shrinks sufficiently fast.
The Critical Role of Generation Time:
The central finding is that singular growth in a realistic, discrete feedback loop requires the generation time ( $T_n$ ) to approach zero rapidly enough to satisfy a "Zeno condition" (the sum of all generation times must converge to a finite value).
- Singularity Theorem: For a time-embedded difference equation, $A(t)$ $A (t)$ has a singularity if and only if:
  1. The generation time $T_n$ meets the Zeno condition (converges fast enough to allow infinite loops in finite time).
  2. The capability $A_n$ meets the boundlessness condition (grows without bound).
- The paper argues that achieving the Zeno condition is physically difficult. If generation times hit a lower bound (e.g., due to physical limits of chip design or training time), singular growth becomes impossible, even if capability gains are massive.
Re-evaluating "Explosive" Scenarios:
The analysis suggests that the most likely outcome of AIRDA is super-exponential but sub-singular growth. If generation times decrease but hit a floor (e.g., saturating at machine speeds), the system may transition from a super-exponential phase to a high-rate exponential phase, and finally to a logistic saturation, rather than a vertical asymptote.
Measure Dependence:
The paper analyzes how the choice of capability metric ( $A$ ) affects the observation of singularities.
- For "similar" measures (where one goes to infinity iff the other does), the threshold for singular growth remains consistent.
- However, for "dissimilar" measures (e.g., Mean Time Between Failures vs. general intelligence), a singularity in one metric (e.g., infinite reliability on a narrow task) does not imply a singularity in the other. The paper warns that metrics like "time horizon" may diverge to infinity without representing a true explosion in general intelligence.

Significance and Claims
The paper claims to refine the theoretical understanding of intelligence explosions by correcting a conflation in previous literature between super-exponential growth and singular growth.

Modesty of Claims: The author explicitly states that the mathematical singularities discussed are likely idealized limits that real-world physical processes will not reach due to upper bounds on intelligence, resources, or generation time. The goal is not to prove that an intelligence explosion will happen, but to clarify the conditions required for the most extreme forms of growth.
Policy and Measurement Implications: The paper argues that because singular growth is harder to achieve than previously thought (requiring generation times to approach zero), observing super-exponential growth rates (e.g., increasing monthly growth percentages) is not sufficient evidence of an impending singularity.
Focus on Generation Time: The author concludes that tracking "generation time" is more pivotal for predicting explosive dynamics than tracking the magnitude of capability gains per cycle. The paper suggests that if generation times cannot be reduced arbitrarily, the "explosion" will likely manifest as a rapid acceleration followed by a plateau, rather than a vertical asymptote.

Ultimately, the paper posits that while AIRDA poses significant risks by accelerating progress, the specific dynamic of a "singularity" (infinite capability in finite time) is a fragile mathematical possibility that relies on conditions (vanishing generation time) that may be physically unattainable.