Intelligence Is Not the Bottleneck

What's supported about the singularity — the intelligence explosion, the scaling it forces, and the limits underneath

Synthesis · the best-supported account of the takeoff, as of mid-2026

Andraž Đurič

Version 1 · June 2026

What this is. A synthesis of the best-supported account of the singularity as of mid-2026: what the intelligence explosion is, why it must scale into the physical world, and what sets the pace of that. It surveys the serious models and the current evidence, marks what is settled and what is contested, leads with a stated lean on the open speed question, and adds one small distinction of its own.
What this is not. Not a forecast with a date. Not a safety analysis — the question of what such systems would do to us is a large separate literature, bracketed here. Not a defence of any single scenario.

1. What the singularity actually names

The word carries a great deal of drama, but the load-bearing idea under it is precise and old. In 1965 I. J. Good observed that a machine able to design better machines — including better designers — would set off a chain, and called the result an intelligence explosion: the first such machine, he said, is the last invention humanity need ever make. Everything serious about the singularity is some version of this. The phenomenon is recursive self-improvement in capability.

The real debate is not whether the idea is coherent but how it behaves, and it turns on three questions that have to be kept apart. Does the loop close — can a system actually improve its own capability without a human in the step. How fast does it run in clock-time. And how continuous is it — a smooth steep ramp, or a sudden discontinuous jump. Christiano's main contribution was to insist these come apart: there can be an enormous, world-transforming acceleration that is nonetheless continuous, with AI that substantially speeds AI arriving before AI that radically speeds it, which counts as “slow” only for lacking a discontinuity while in absolute terms it is the Industrial Revolution at ten to a hundred times the pace. Hold speed, continuity, and locality apart and most of the confusion in this area clears.

2. The cognitive phase, and where the evidence sits

The first half of the story is cognitive: software improving software. Whether this runs away on its own — a software intelligence explosion, before robotics or new hardware enter — is the most contested question in the area. Eth and Davidson argue that software progress alone could compound into accelerating gains. Erdil and Barnett argue the reverse: that AI's economic weight will come from broad deployment across the whole economy rather than from a software singularity, because algorithmic insight plausibly gets harder to find as the easy ideas are spent.

The data, as of mid-2026. METR's task-horizon series — the length of task a frontier agent completes autonomously — has grown from about nine seconds in 2020 to roughly fourteen hours, doubling about every seven months across the period and faster, every three to four months, through 2024–2025. Real and still accelerating. It is also mostly a software-engineering measure; the people who track it expect it to slow as reinforcement learning consumes a larger share of compute; and the automation of research observed so far is real but partial, with one widely-cited projection putting the algorithmic speedup from automated R&D at roughly fifty percent by early 2026 — an acceleration, not a runaway. Forecasters asked the odds of compressing 2018–2024's progress into two years answered around twenty percent for AI specialists and eight for generalist superforecasters.

My own read — and this is a lean, with wide error bars, not a reading of consensus — is that the cognitive phase is real and probably steep, but the clean months-to-godlike runaway is the optimistic tail rather than the centre. The mechanics push against the extreme version: a “generation” minimally means training a successor and verifying it holds at scale, both bound to wall-clock time, and the diminishing-returns worry about algorithmic insight is not easily dismissed. I land closer to continuous-and-fast than to discontinuous runaway, while granting a non-trivial probability to the fast tail. Someone who weights the recent acceleration and the hardware overhang more heavily than I do could land faster without being unreasonable.

3. Why it cannot stay in software

Grant the cognitive phase, fast or slow. It does not stay there, and this is the part of the story the drama tends to rush. Cognitive capability is valuable, but the things that make a singularity a singularity — abundant energy, new materials, manufacturing, anything that visibly remakes the world — are physical. They are cashed in action on matter, not in better reasoning alone. And the cognitive phase runs into a physical wall from the inside as well: to keep scaling, a system needs more compute, which means more chips, more fabs, more power, all of it physical capital that has to be built. So the arc turns, of necessity, from cognition to the physical world. This is why the most careful recent framing, Cotra's, measures takeoff not by raw intelligence but by the rate of progress in physical technology — that is where “everything changed” actually lands.

How the two phases relate is the fast-or-slow question restated. On the fast view, the software explosion runs nearly to completion first, with physical technology barely moving, and only then does a now-superhuman system turn to remaking the physical world quickly. On the slow view, the two are coupled: sustaining the software loop requires physically scaling the whole AI supply chain, so physical buildout paces the cognitive runaway rather than following it. The evidence does not yet settle which. But nothing serious has the physical phase being skipped — the disagreement is over its timing relative to the cognitive phase, not its necessity.

4. The physical clock

The physical phase has been measured, which is the most useful and least-known part of the whole picture. Davidson and Hadshar separate the intelligence explosion from what they call the industrial explosion and give the second a slower clock, for two concrete reasons. The underlying technologies improve at different rates — AI compute and algorithms double on the order of one to two years, robotics more slowly, perhaps every one to four. And the feedback loops differ in kind: once you train a capable model you can run millions of copies at once, but a robot has to be physically built before it can build the next. Bits copy for free; atoms do not.

The numbers follow from that asymmetry. The world's robot count last doubled over about six years. With current physical technology but abundant AI cognitive labour, independent estimates — Davidson and Hadshar, Epoch's supply-chain analysis, and a recent input–output model of the entire economy — converge on physical doubling times of roughly a year: not a month, not a decade. The technological-feasibility ceiling, anchored to biological replicators, is days to weeks, with the standing caveat that a single large facility still takes months to years to build. So the physical phase is fast in historical terms and slow in the drama's terms. The part that actually remakes the world runs on a clock of years compressing toward something faster — not milliseconds.

5. The deeper reason the world sets the pace

Beneath the engineering numbers there is a more general reason the physical phase resists compression, and here I will state the framing this corpus uses, kept deliberately to one side of the survey because it is a reframing rather than a new finding. The point is about what cognition can and cannot substitute for. Scaling capability that already works means copying a model already validated against reality — and copying is free, which is the whole content of “bits copy, atoms are built.” But acquiring genuinely new physical capability — the materials and devices that do not yet exist — means testing predictions that have not been validated, and an untested prediction is settled only by running it: synthesising the thing, building it, measuring what the world does. No amount of internal modelling discharges that step, because a model's verdict on an untested configuration is a guess — free to hold, and charged only when reality is allowed to answer.

That suggests a clean way to split the physical phase, which is the one small thing worth adding here: scaling versus discovery. The industrial-explosion numbers mostly measure scaling — replicating designs that work, where the answer is about a year. Discovery — finding new physics, new materials — is bounded at the point where prediction meets reality, where the numbers are softer and the literature thinner. Scaling copies a validated map; discovery has to query the territory.

The same discipline that earns this framing its keep also limits the claim, so it should be stated plainly: this is a reframing, not a tighter bound. The genuine thermodynamic floors — the energy to erase a bit, the relaxation times of matter — sit far below the constraints that actually bite, which are synthesis, fabrication, and construction. The limit that binds discovery is engineering, not the second law. The thermodynamic framing explains the shape of the asymmetry — why cognition cannot buy its way past the physical clock — without setting a number the engineering estimates do not already set. Even its natural physical anchor, a thermodynamic bound on self-replication, is contested rather than settled. The framing is a lens, offered as a lens.

6. The supported picture

Put together, the best-supported account runs like this. There is a real mechanism, recursive self-improvement, and a cognitive phase that is probably steep — though whether it runs away in months or unfolds over years is genuinely open, and my lean is toward the slower, continuous end with a live fast tail. That phase does not stay in software: the arc turns, necessarily, toward the physical world, both because that is where transformation is cashed and because scaling cognition itself demands physical capital. The physical phase has been measured — roughly yearly doublings with current technology, a ceiling of days to weeks, single facilities lagging by months to years — and it is fast in historical terms while remaining slow against the drama. And the constraint that binds, beneath the engineering, is not intelligence. It is the physical world, which answers new questions only by being run, on its own clock, at its own cost.

That is the inversion the dramatic version tends to miss. The intelligence explosion is the tractable half. What sets the pace of a singularity is the part no quantity of intelligence shortcuts: building the world out, and querying it for what is not yet known. Intelligence is not the bottleneck. The territory is.

What this corpus adds to that is modest and clearly marked — a single distinction, scaling versus discovery, drawn from the principle that a map costs nothing until it is cashed against the territory. The survey is the substance; the distinction is a refinement of one joint in it.

Standing of this document. A synthesis note, dated to mid-2026. The empirical figures about AI progress will move; the structural claims — that the arc turns physical, that the physical phase is measured and bounded, that the binding constraint is the territory rather than intelligence — are meant to outlast them. The speed lean is a judgment, marked as such, and is the most revisable claim here. Corrections are welcome and expected; they are the program operating.

References

1. Good, I. J. (1965). Speculations Concerning the First Ultraintelligent Machine. Advances in Computers 6, 31–88. The original intelligence-explosion argument. Standard
2. Christiano, P. (2018). Takeoff Speeds. The Sideways View. Slow takeoff as continuity rather than small impact; speed, continuity, and locality held apart. Standard
3. Cotra, A. (2026). Takeoff speeds rule everything around me. Planned Obsolescence. Operationalising takeoff by progress in physical technology. Verified · June 2026
4. Eth, D., & Davidson, T. (2025). Will AI R&D Automation Cause a Software Intelligence Explosion? Forethought. The case for a software-only runaway. Verified · June 2026
5. Erdil, E., & Barnett, M. (2025). Most AI Value Will Come From Broad Automation, Not From R&D. Epoch AI. The case against a software singularity. Verified · June 2026
6. Davidson, T. (2021; 2023). Report on Whether AI Could Drive Explosive Economic Growth; and What a Compute-Centric Framework Says About AI Takeoff Speeds. Open Philanthropy. Interactive model at takeoffspeeds.com. Takeoff as the time from automating roughly 20% to near-total of cognitive tasks. Verified · June 2026
7. Davidson, T., & Hadshar, R. (2025). The Industrial Explosion. Forethought. The physical phase as later and slower; robot doubling times; copy-versus-build. Verified · June 2026
8. Kwa, T., et al. (METR) (2025). Measuring AI Ability to Complete Long Tasks. arXiv:2503.14499; with subsequent METR time-horizon updates through 2026. The task-horizon trend. Verified · June 2026
9. International AI Safety Report (2026). Forecaster estimates on AI-accelerated AI R&D and the mixed state of current evidence. Verified · June 2026
10. Still, S., Sivak, D. A., Bell, A. J., & Crooks, G. E. (2012). Thermodynamics of Prediction. Physical Review Letters 109, 120604. Nonpredictive information and dissipation; the cost interpretation of map–territory divergence. Standard
11. Landauer, R. (1961). Irreversibility and Heat Generation in the Computing Process. IBM Journal of Research and Development 5, 183–191. The physical cost of logically irreversible information processing. Standard
12. England, J. L. (2013). Statistical Physics of Self-Replication. The Journal of Chemical Physics 139, 121923; and (2015) Dissipative Adaptation in Driven Self-Assembly, Nature Nanotechnology 10, 919–923. A thermodynamic bound on self-replication. Standard
13. Kolchinsky, A. (2024). Thermodynamic Dissipation Does Not Bound Replicator Growth and Decay Rates. arXiv:2404.01130. A challenge to the standard interpretation of England's bound. Verified · June 2026
14. Omohundro, S. M. (2008). The Basic AI Drives. Proceedings of the First AGI Conference; and Bostrom, N. (2012). The Superintelligent Will. Minds and Machines 22(1), 71–85. Convergent instrumental sub-goals and orthogonality; the bracketed safety literature. Standard
15. The Fidelity Program. The program statement, including the map–territory principle and the retirement of the “consistency tax” label. Internal / existing on this blog
16. Name the Relata. The diagnostic for relational operators whose frame has been stripped. Internal / existing on this blog

Author: Andraž Đurič, Slovenia. The shape of this synthesis first emerged in conversation with Gemini (Google); the survey, the analysis, and the checking of every source against primary references were developed in collaboration with Claude (Anthropic). Contributions are judged on content rather than origin; doing otherwise would be the genetic fallacy. Text licensed CC BY 4.0.