The Singularity Is Harder Than You Think: Toby Ord Re-examines Intelligence Explosions with Math

> 33 pages, one theorem — dissecting Silicon Valley's favorite doomsday narrative down to its skeleton.

An Awkward Silence in a Conference Room

Imagine this scene: at a weekly meeting at a frontier AI lab, a researcher draws a curve on the whiteboard. The curve climbs slowly at first, then suddenly bends upward, and at some time point \(t^*\) shoots vertically toward the ceiling as if ignited.

"Based on our internal benchmarks," she taps the whiteboard, "the singularity of the intelligence explosion is projected at—"

Someone raises a hand and interrupts: "Wait. Will this curve actually go to infinity?"

The room goes quiet for two seconds. Everyone is thinking the same question, but no one dares to say it out loud: Is the "singularity" we keep talking about a mathematically real singularity, or just an exponential running too fast?

This sounds like nitpicking over words. But Toby Ord — Oxford philosopher and author of *The Precipice* — in a 33-page paper posted to arXiv in August 2026, proves with pure mathematics: this distinction is not semantics, it's destiny.

The Core Problem: Super-exponential ≠ Singularity

First, two concepts that are often conflated.

Super-exponential growth: the growth rate itself accelerates. For example, \(A(t) = e^{e^t}\), or month-over-month growth of 10%, then 20%, then 30% — that's super-exponential. It's terrifyingly fast, but it still takes finite values at every finite point in time.

Singular growth: at some finite time \(t^*\), \(A(t)\) actually goes to infinity. Not "very large" — infinite.

In the Silicon Valley narrative, these two are often treated as the same thing: as if AI self-improvement, once accelerating, necessarily leads to a singularity. Ord's first contribution is to cleanly separate the two mathematically.

Continuous Models: A Higher Bar Than You Think

The simplest intelligence-explosion model is a differential equation:

Here \(A\) is some measure of AI capability, \(\dot{A}\) its rate of change, and \(f\) the feedback mechanism of "stronger capability → faster improvement."

Intuitively, as long as \(f\) grows faster than \(A\) — i.e., superlinearly — that should suffice for a singularity. Ord proves this is wrong.

Superlinearity only gives you super-exponential growth, not a singularity.

For a singularity, \(f\) must cross a higher threshold, which Ord calls the blow-up condition:

Note the ellipsis — it is an infinite product. Raising the last term to any power greater than 1 is enough to produce a singularity; without that, it isn't.

There is an extremely narrow transition zone. For instance, \(f(A) = A \cdot \log(A)\) yields double-exponential growth \(A(t) = e^{e^t}\) — absurdly fast, but no singularity. \(f(A) = A \cdot \log(A) \cdot \log(\log(A))\) is also insufficient. You must raise the last term to some power greater than 1.

In other words, there exists a large class of super-exponential growth that never diverges in finite time. The Silicon Valley narrative lumps all of these together as "the singularity," which is mathematically imprecise.

Discrete Models: The Real Disruption

That was just the warm-up. Ord's real contribution is at the next level.

Differential equations assume time is continuous — feedback happens at every instant. But AI R&D feedback loops don't work that way. Training a new model takes months; deployment takes weeks; gathering feedback takes more time. Feedback is discrete, not continuous.

Ord introduces what he calls a time-embedded difference equation. Let \(A_n\) be the capability after the \(n\)-th feedback cycle, and \(T_n\) the time the \(n\)-th cycle takes (the generation time). Then:

This model tracks two things: how much capability grows per cycle, and how long each cycle takes.

Here's the key finding. In a pure difference equation (fixed \(T_n\)), no singularity is possible no matter how fast \(f\) grows. Because by any finite time, only finitely many cycles have run, each adding a finite amount — the sum stays finite.

For a singularity, the generation time \(T_n\) must tend to zero, fast enough that infinitely many cycles fit in finite time. Ord calls this the Zeno condition — it is exactly the mathematical structure of Zeno's paradox of Achilles and the tortoise: infinitely many steps completed in finite time.

The other requirement is the boundlessness condition: \(A_n\) must grow without bound.

Singularity Theorem: In a time-embedded difference equation, \(A(t)\) has a singularity if and only if \(T_n\) satisfies the Zeno condition and \(A_n\) satisfies the boundlessness condition.

Neither condition alone suffices. Doing well on one cannot rescue failure on the other.

Generation Time Is the Real Driver

The theorem itself is clean. But Ord's next observation is the paper's deepest point.

Compare two scenarios:

Scenario 1: Each cycle doubles \(A\) (\(f(A) = A\), exponential growth), with generation time \(T_n = 1/n\). \(A_n\) grows extremely fast (double-exponential), but \(T_n\) fails the Zeno condition — no singularity.

Scenario 2: Each cycle adds only \(1/n\) (\(A_n\) barely grows), but generation time is \(T_n = 1/n^{1.01}\). Both conditions hold — a singularity occurs.

In the second scenario, each cycle adds almost nothing, but because cycles get faster and faster, the system still reaches infinity in finite time.

This reveals something: in discrete feedback models, singularities are driven almost entirely by generation time, not by per-cycle increments.

Ord offers a beautiful geometric intuition. A singularity requires the curve's gradient to blow up in finite time. Gradient = rise / horizontal distance. For an infinite gradient, either the "rise" must be infinite, or the "horizontal distance" must shrink to zero. Continuous models can't distinguish these; in discrete models, the "horizontal distance" is the generation time — only if generation time shrinks to zero is a singularity possible.

Coordinate Singularities vs. Essential Singularities

The math is already disruptive enough, but Ord borrows a distinction from physics to go deeper.

In general relativity, a black hole's event horizon looks like a mathematical singularity in Schwarzschild coordinates, but it's merely an artifact of the coordinate choice — change coordinates and it disappears. The singularity at the black hole's center is a true "essential singularity," removable by no coordinate change.

Ord says: AI capability metrics have the same problem.

Consider the METR time horizon — the length of tasks AI can complete autonomously. This metric has grown exponentially in recent years, and some worry it will diverge. But Ord points out: an infinite time horizon only means the AI reaches 100% reliability on that class of tasks. It does not mean infinite intelligence overall.

This is a "coordinate singularity" — the metric diverges, but there is no essential singularity behind it.

More generally, Ord proves: if two metrics \(A\) and \(B\) satisfy "one is unbounded iff the other is" (he calls them *similar measures*), then they either both have singularities or neither does. But if one metric's full range maps to only a finite interval of the other — e.g., \(A \to \infty\) while \(B\) only reaches 100% — they are not similar, and \(A\)'s singularity may be merely a coordinate singularity.

This is a warning for AI evaluation: the metric you choose determines whether the "singularity" you see is real or an illusion.

The Most Likely Scenario: A Three-Phase Logistic

What would a real intelligence explosion look like? Ord offers a most-likely scenario.

A singularity requires generation time tending to zero. In reality, generation times have a floor — training new models takes time, designing new chips takes time, physical experiments take time. Even if AI can compress some loops (e.g., scaffold optimization) to seconds, these are only a small fraction of the full AI R&D pipeline. Most stages will hit a physical floor.

So the most likely trajectory is:

Phase 0: Human-paced exponential growth. AI isn't yet participating in R&D; growth is driven by human researchers.

Phase 1: Super-exponential growth. AI starts participating in R&D; generation time compresses from human pace (months) toward machine pace (hours). Because generation time keeps shrinking, the growth rate keeps rising — the hallmark of super-exponential growth.

Phase 2: Machine-paced exponential growth. Generation time hits the floor, and the growth rate stabilizes at a high level. Growth returns to exponential, just much faster than Phase 0.

Phase 3: Logistic saturation. Capability approaches some ceiling (physical limits of algorithms, data, compute); growth decelerates and eventually saturates at \(A^*\).

The whole trajectory is a logistic curve — exponential at the start, a stretch of super-exponential in the middle, back to exponential, then saturating. No singularity, but still terrifyingly fast.

Ord emphasizes at the end: a singularity being hard to achieve doesn't mean AI won't be dangerous. Even a mere linear speedup — compressing a decade of human progress into one year — would be enough to upend society. "Not singular" doesn't mean "not dangerous."

Engineering Implication: Track Generation Time

The most direct engineering implication for practitioners comes from Ord's policy recommendation in the conclusion:

Frontier labs should report their current generation times — especially for pretraining and RLVR post-training.

Why? Because the core theorem tells us the singularity criterion lies not in \(f(A)\) but in \(T_n\). Watching benchmark scores rise can't tell you whether you're heading toward a singularity; you have to watch how much the feedback loop itself is accelerating.

This is the exact opposite of current AI evaluation practice. The whole industry scrambles to measure capability scores (MMLU, GPQA, SWE-bench), but almost no one measures "how long does a full train-evaluate-deploy cycle take." Capability scores tell you "how strong it is now"; generation time tells you "how fast it will get next" — the latter is the key variable for predicting an explosion.

An analogy: meteorologists forecasting hurricanes don't just look at current wind speed; they look at the rate of pressure drop, because pressure-drop rate is the driver of hurricane intensification. Capability scores are wind speed; generation time is pressure. The whole AI evaluation industry is staring at the anemometer, but few are reading the barometer.

Personal Reflection: Demystifying the Singularity Narrative

My strongest takeaway from this paper is a sense of demystification.

Silicon Valley's singularity narrative has a fatalistic flavor — as if once AI self-improvement starts, it must march toward some uncontrollable endpoint. That narrative is both exciting and frightening, but it doesn't hold up mathematically.

Ord proves in 33 pages: a singularity requires two independent conditions to hold simultaneously. Generation time must satisfy the Zeno condition (tending to zero fast enough); capability must satisfy the boundlessness condition (no ceiling). In the real world, both are hard to satisfy — generation times have physical floors, and capabilities face algorithmic and data ceilings.

But demystifying isn't the same as relaxing. The paper's closing passage stuck with me most:

> "AIRDA could raise the pace of AI R&D to dangerous levels even with only linear acceleration. If humanity's original trajectory was \(A(t)\), AIRDA turns it into \(A(10t)\) — we cover a decade of human progress every year. This introduces most of the danger even though the shape of the curve hasn't changed at all."

The danger lies not in the singularity, but in the speed.

This is an adjustment for the AI safety community. If you've been preparing for a "singularity moment" — some \(t^*\) where capability suddenly explodes — you may have missed the real risk: a world with no singularity, but with continuously compressing generation times, can still push AI beyond human-comprehensible speeds within a few years.

Infinity is not required for things to spiral out of control.

Paper Links

• arXiv: The Dynamics of Intelligence Explosions (Toby Ord, 33 pages, 3 figures)
• Toby Ord's website: tobyord.com
• Related research: Forethought - Preparing for the Intelligence Explosion

One-line summary: A singularity is not the inevitable destiny of super-exponential growth — it requires generation times satisfying the Zeno condition, and real-world generation times have a physical floor. But "not singular" doesn't mean "not dangerous": a singularity-free intelligence explosion can still push AI beyond human-comprehensible speeds within a few years.