The Singularity Is Harder Than You Think: Toby Ord Revisits Intelligence Explosions with Math
> 33 pages, one theorem—and the most popular Silicon Valley doomsday narrative is stripped to its skeleton.
An Awkward Silence in a Meeting Room
Imagine a weekly meeting at a frontier AI lab. A researcher draws a curve on the whiteboard: it climbs slowly, then bends sharply upward, and at some time \(t^*\) it shoots vertically toward the ceiling like it has been ignited.
"Based on our internal benchmarks," she taps the board, "the singularity of the intelligence explosion is projected to arrive at—"
Someone raises a hand: "Wait. Will this curve actually reach infinity?"
Two seconds of silence. Everyone is thinking the same question but nobody dares ask it out loud: is the "singularity" we keep talking about a real mathematical singularity, or just an exponential that's moving too fast?
It sounds like pedantry. But Toby Ord—Oxford philosopher and author of *The Precipice*—in a 33-page paper posted to arXiv in August 2026, proved with pure mathematics that this distinction is not wording; it is destiny.
Core Problem: Super-Exponential ≠ Singularity
First, two concepts that are constantly conflated:
Super-exponential growth: growth whose rate itself accelerates. Example: \(A(t) = e^{e^t}\) — growing 10% one month, 20% the next, 30% the next. Terrifyingly fast, but still finite at every finite time.
Singular growth: at some finite time \(t^*\), \(A(t)\) genuinely goes to infinity. Not "very large"—infinite.
Silicon Valley narratives treat the two as the same: if AI can improve itself and growth accelerates, a singularity must follow. Ord's first contribution is to cleanly separate them mathematically.
Continuous Models: A Higher Threshold Than You Think
The simplest intelligence-explosion model is a differential equation:
Intuition says that as long as \(f\) grows faster than \(A\)—i.e., super-linearly—a singularity follows. Ord proves this is wrong.
Super-linearity gives you super-exponential growth, not a singularity.
For a singularity, \(f\) must cross a higher threshold that Ord calls the blow-up condition:
Note the ellipsis—this is an infinite product. Raising the last term to any power greater than 1 suffices for a singularity; without it, it doesn't. There is an extremely narrow transition zone: \(f(A) = A \cdot \log(A)\) yields double-exponential growth \(A(t) = e^{e^t}\)—absurdly fast, but no singularity. Even \(f(A) = A \cdot \log(A) \cdot \log(\log(A))\) falls short.
In other words, there exists a large class of extremely fast super-exponential growths that never reach infinity in finite time. Calling them all "the singularity" is mathematically sloppy.
Discrete Models: The Real Disruption
Differential equations assume continuous time—feedback happening at every instant. But AI R&D feedback loops don't work that way: training a new model takes months, deployment weeks, feedback collection longer. Feedback is discrete, not continuous.
Ord introduces a time-embedded difference equation. Let \(A_n\) be capability after the \(n\)-th feedback cycle and \(T_n\) the time that cycle takes (the generation time):
The key finding: in a pure difference equation (fixed \(T_n\)), no singularity is possible regardless of how fast \(f\) grows—finitely many cycles in finite time, each adding a finite value.
For a singularity, generation times \(T_n\) must shrink toward zero fast enough that infinitely many cycles fit into finite time. Ord calls this the Zeno condition—exactly the mathematical structure of Zeno's paradox of Achilles and the tortoise. A second requirement is the boundlessness condition: \(A_n\) must grow without bound.
Singularity Theorem: In time-embedded difference equations, \(A(t)\) has a singularity if and only if \(T_n\) satisfies the Zeno condition AND \(A_n\) satisfies the boundlessness condition.
Neither can compensate for the other.
Generation Time Is the True Driver
Compare two scenarios:
- Scenario 1: capability doubles every cycle (\(f(A) = A\), exponential), generation time \(T_n = 1/n\). \(A_n\) grows double-exponentially, but \(T_n\) fails the Zeno condition—no singularity.
- Scenario 2: each cycle adds only \(1/n\) (nearly flat growth), but \(T_n = 1/n^{1.01}\). Both conditions hold—a singularity appears.
- Phase 0: human-speed exponential growth, driven by human researchers.
- Phase 1: super-exponential growth. AI joins R&D; generation times compress from months to hours, so growth rates keep rising.
- Phase 2: machine-speed exponential growth. Generation times hit the floor; growth returns to exponential, just much faster.
- Phase 3: logistic saturation. Capabilities approach limits imposed by algorithms, data, and compute.
- 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
In the second scenario, each cycle barely improves anything, yet the explosion happens because cycles get ever faster.
In discrete feedback models, singularities are driven almost entirely by generation time, not per-cycle gains.
Ord offers a geometric intuition: a singularity requires the curve's gradient to blow up. Gradient = rise / horizontal distance. In discrete models, "horizontal distance" is generation time—only shrinking generation time toward zero makes a singularity possible.
Coordinate Singularities vs. Essential Singularities
Ord borrows a distinction from physics. In general relativity, a black hole's event horizon looks singular in Schwarzschild coordinates but disappears with a coordinate change; the central singularity is essential—no coordinate change removes it.
AI capability metrics have the same issue. Consider the METR time horizon—how long a task AI can complete autonomously. It grows exponentially and some fear it goes to infinity. But an infinite time horizon only means AI reaches 100% reliability on that class of tasks—not infinite intelligence overall. That's a coordinate singularity: the metric goes to infinity, but there is no essential singularity underneath.
More generally, Ord proves that "similar measures" (each unbounded iff the other is) share singularity status. But if one metric's entire range maps to a finite interval of another—like \(A \to \infty\) while \(B \to 100\%\)—they are dissimilar, and \(A\)'s singularity may be merely a coordinate one.
This is a warning for AI evaluation: the metric you choose determines whether the "singularity" you see is real or an artifact.
The Most Likely Scenario: Three-Phase Logistic
Generation times have physical floors—training models, designing chips, running experiments take time. Even if AI compresses some loops (e.g., scaffold optimization) to seconds, those are a small fraction of the R&D pipeline. The likely trajectory:
The whole trajectory is a logistic curve—no singularity, but still frighteningly fast. Ord stresses: a singularity being hard doesn't mean AI isn't dangerous. Even linear acceleration—compressing a decade of human progress into one year—could upend society. "Not singular" ≠ "not dangerous."
Engineering Implication: Track Generation Times
Ord's most direct recommendation for practitioners: frontier labs should report current generation times—especially for pretraining and RLVR post-training.
The theorem says the singularity criterion lives in \(T_n\), not \(f(A)\). Benchmark scores can't tell you whether you're heading toward a singularity; you need to know how fast the feedback loop itself is accelerating. Current practice is the opposite: the industry obsessively measures capability scores (MMLU, GPQA, SWE-bench) but almost never measures "how long does one full training-evaluation-deployment cycle take."
An analogy: hurricane forecasters watch not just wind speed but the rate of pressure drop, because pressure drop drives acceleration. Capability scores are wind speed; generation time is pressure. The whole evaluation industry is staring at anemometers while few read barometers.
Personal Reflection: Demystifying the Singularity Narrative
Silicon Valley's singularity narrative has a fatalistic flavor—as if self-improving AI must inevitably reach an uncontrollable endpoint. Mathematically, it doesn't hold. Ord proves a singularity needs two independent conditions, both of which reality makes hard: generation times have physical floors; capabilities have algorithmic and data ceilings.
But demystification isn't reassurance. As the paper concludes:
> "AIRDA could raise the pace of AI R&D to dangerous levels even with only linear acceleration. If the human trajectory would have been \(A(t)\), AIRDA turns it into \(A(10t)\)—covering a decade of human progress per year. This introduces most of the danger, even if the shape of the curve never changes."
The danger lies not in the singularity, but in the speed.
If you've been preparing for a "singularity moment," you may be missing the real risk: a world with no singularity but continuously compressing generation times can still push AI beyond human-supervisable speeds within years.
You don't need infinity to lose control.
Links
One-line takeaway: A singularity is not the inevitable destiny of super-exponential growth—it requires generation times satisfying a Zeno condition, and real-world generation times have physical floors. But "not singular" is not "not dangerous": a singularity-free intelligence explosion can still outrun human oversight within years.