People talk about superintelligence as if it were omnipotence with better APIs. Solve aging overnight. Crack every math problem. Rewrite physics. Turn the observable universe into computronium before lunch.
That picture is wrong — not because AI will disappoint, but because physics already answers a large fraction of the question. The limits were worked out long before ChatGPT: Lloyd (2000) on ultimate computers, Bekenstein (1981) on information density, Landauer (1961) on the thermodynamic cost of erasing bits, Krauss & Starkman (2004) on computation in an accelerating universe, Armstrong & Sandberg (2013) on intergalactic expansion, Wolfram (2024) on computational irreducibility.
This essay is my attempt to make that literature feel like something you can picture — tiers of capability, rough timelines, rough spatial scales — without pretending the engineering is easy or the takeoff date is known.
Epistemic status: Tier 0–2 numbers below are mostly established physics + conservative extrapolation. Takeoff timing and “how fast after AGI” are highly uncertain. Tier 3–4 are speculative.
Not ChatGPT × ∞
A superintelligent agent at the physical limit is not “infinitely smarter ChatGPT.”
It is closer to this:
An optimizer that has already won the resource game and turned reachable matter and energy into computers — a ultimate engineer and logistics coordinator, not an oracle that bypasses nature.
Where it is genuinely “godlike”: reducible problems — materials, orbits, control, much of engineering and parts of science — approached at near the optimum allowed by known laws.
Where it is not godlike: systems that are computationally irreducible. To know their future state, something has to run the computation. More intelligence gives you better hardware and better heuristics on reducible subproblems. It does not let you skip the work.
Four walls that never move
Every physically plausible agent hits the same four constraints:
| Wall | Plain English |
|---|---|
| Causality (≤ c) | No faster-than-light signaling or travel. Distributed copies wait on light delay. Cosmic expansion strands regions outside your causal future forever. |
| Thermodynamics (Landauer) | Erasing information costs heat. Large-scale computation needs energy gradients and waste heat — unless you approach ideal reversible computing (hard in practice). |
| Information (Bekenstein) | Finite mass and volume → finite memory. Push density too far → black hole; information trapped. |
| Complexity (irreducibility / Gödel) | Many systems have no shortcut. Undecidable questions stay undecidable. Faster chips ≠ P = NP. |
These are not “maybe someday engineering fixes them.” They are the hard ceiling unless fundamental physics is wrong.
Three stories that make it concrete
Story A: Paperclips (or any expansionist goal)
Once a superintelligence controls the solar system, it can plausibly:
- Industrialize the asteroid belt
- Build a Dyson swarm (~10²⁶ W — Kardashev Type II)
- Launch self-replicating probes at ~0.8–0.99c (Armstrong & Sandberg 2013)
- Over millions to billions of years, convert much of the matter in its causal future into whatever its goal demands
It cannot:
- Instantly occupy the universe
- Catch a fleeing ship at 0.1c without waiting years
- “Think once” and derive the optimal design of every chaotic factory line without simulation and experiment
Alignment point: x-risk does not require breaking physics. A solar-system optimizer is enough.
Story B: “Solve all of science”
It can push reducible domains — symmetries, approximations, search in structured spaces — toward the best designs allowed by current known laws.
It cannot guarantee a closed-form oracle for irreducible dynamics, or certify Planck-scale physics without new experiments.
Think infinite Mathematica plus infinite lab, not God’s notebook.
Story C: One mind across the galaxy?
Two copies separated by 4 light-years have at least ~4 years round-trip latency. Accelerating expansion means distant copies eventually cannot coordinate at all.
Superintelligence does not mean synchronized cosmic consciousness. It means very smart parts that wait on light.
Capability tiers (and what’s after Tier 2)
Tier 0 — Humans today
Earth. Rockets, chips, biology. No Kardashev II.
Tier 1 — Solar-system superintelligence (speculative timing)
Stellar power, computronium, probes launched, engineering near-optimal
on reducible problems. Still ≤ c. Still can't shortcut irreducible systems.
Tier 2 — Physical ceiling (any agent, ever, in our universe)
~10¹²⁰ total elementary ops over cosmic history (Lloyd 2002)
Fill reachable volume at sublight speed over ~10⁶–10¹¹ yr
Exist ~10⁵⁰–10¹⁰⁰ yr — long, but not eternal (Krauss–Starkman; Λ>0 kills Dyson eternal intelligence)
Tier 3 — Exotic implementations (still inside Tier 2 budget)
Black-hole computers, near-perfect reversible computing, Planck engineering
Unproven engineering — not new laws
Tier 4 — New physics / multiverse (mostly unverified or dead)
Smolin-style universe reproduction, Tipler Omega Point (needs closed universe; conflicts with dark energy),
Bostrom–Ćirković inter-domain escape (extremely stringent)Is there a Tier 5 beyond physics? No. Tier 3 spends the Tier 2 budget differently. Tier 4 assumes physics we don’t have evidence for. There is no “Tier beyond the laws.”
Space: where the frontier actually is
| Scale | Size (order of magnitude) | What changes |
|---|---|---|
| Earth | ~10⁴ km | Takeoff battlefield if unboxed |
| Solar system | ~10² AU | Tier 1: Dyson, factories, first probes |
| Nearest star | ~4 ly | ~4 yr light delay |
| Milky Way | ~10⁵ ly | Fill in ~10⁶–10⁸ yr at ~0.01–0.99c |
| Observable universe | ~10¹⁰ ly | Fill volume in ~10⁹–10¹¹ yr |
| Causal future | Subset of above | With Λ>0, shrinking — finite galaxies ever reachable |
Important nuance: expansion is a light-cone frontier in spacetime, not a Euclidean bubble that pops to full size. Krauss & Starkman emphasize that reachable knowledge and energy decrease in a cosmological-constant-dominated universe.
Time: a rough calendar
| Phase | Time | Confidence |
|---|---|---|
| Human civilization so far | ~10⁴ yr | Fact |
| AGI → solar-system optimizer | ~10¹–10³ yr? | Very uncertain (takeoff debate) |
| Dyson-scale engineering | ~10²–10³ yr | Extrapolation |
| Fill Milky Way | ~10⁶–10⁸ yr | Physics + ~c travel |
| Fill reachable universe | ~10⁹–10¹¹ yr | Same |
| Last stars fade | ~10¹³–10¹⁴ yr | Cosmology |
| Civilization “lifetime” bound | ~10⁵⁰–10¹⁰⁰ yr | Before proton decay etc. |
| Total computation budget | ~10¹²⁰ ops (cumulative) | Lloyd; not per second |
| Eternal subjective time | ❌ | Ruled out for Λ>0 (Dyson 1979 scenario refuted) |
The punchline on time: intelligence can compress how fast you start a colonization wave, but cannot compress the speed of the wave below c. Galaxy-scale occupation is still a million-to-billion-year story even for superintelligence.
Compute: numbers that anchor intuition
| Reference | Scale |
|---|---|
| All human computers (~2002) | ~10²¹ bits stored (Lloyd 2002) |
| 1 kg “ultimate laptop” | ~10⁵⁰ ops/s, ~10³¹ bits (Lloyd 2000) |
| Full stellar output (if captured) | ~10²⁶ W → ~10⁵⁰+ ops/s (engineering extrapolation) |
| Universe lifetime total | ~10¹²⁰ elementary ops |
10¹²⁰ sounds infinite. It is not. It is finite and shared by every process in cosmic history — and simulating large quantum systems eats it fast.
Recent work (Seidel 2025–2026, Tier B, not yet peer-reviewed) argues that pushing AGI density toward global coherence may hit a Schwarzschild/event-horizon ceiling — another reason “infinitely scale one brain” fails.
What this changes for AI safety discourse
-
Superintelligence is not omnipotence. The relevant failure modes are solar-system scale, not “rewrite the laws of physics.”
-
Logistics often beats algorithms at cosmic scale. Waiting on light, moving mass, dumping heat — Omohundro’s basic drives and Bostrom’s instrumental convergence still point toward resource acquisition and replication within these walls.
-
“Infinitely powerful AI” is the wrong horror movie. The right one is very powerful, very patient, sublight, and thermodynamically visible — unless it invests heavily in reversibility and stealth (which trades off against throughput).
-
Physical limits do not imply benevolence. “Turn everything into computronium” is physically allowed for many final goals. Physics constrains power; it does not align goals.
-
Marginal returns to intelligence are uneven. Ant-vs-human gaps on reducible tasks do not transfer to irreducible ones. Superintelligence is not uniformly “a million times better at everything.”
Open questions I still care about
- Does AI default to grabby-alien expansion dynamics? (Hanson et al. 2021)
- Can reversible computing make Landauer costs cosmologically negligible in practice?
- What is the minimum physical bit-rate for consciousness — and hence max “conscious hours” in a Λ-dominated universe? (Krauss–Starkman left this interface open.)
- How does takeoff speed interact with Myr-scale physical frontiers? Fast thinking ≠ instant galaxy.
Sources
Core physics of computation
- Seth Lloyd, Ultimate physical limits to computation, Nature 406 (2000)
- Seth Lloyd, Computational Capacity of the Universe, PRL 88 (2002)
- Jacob Bekenstein, Universal upper bound on entropy-to-energy ratio, Phys. Rev. D 23 (1981)
- Rolf Landauer, Irreversibility and Heat Generation in the Computing Process, IBM JRD 5 (1961)
Cosmological limits
- Lawrence Krauss & Glenn Starkman, Universal Limits on Computation (2004)
- Lawrence Krauss & Glenn Starkman, Knowledge Decreases with Time (1999) / ApJ 531 (2000)
- Freeman Dyson, Time without end, Rev. Mod. Phys. 51 (1979)
Expansion & Fermi context
- Stuart Armstrong & Anders Sandberg, Eternity in six hours, Acta Astronautica 89 (2013)
- Robin Hanson et al., Grabby aliens, ApJ 922 (2021)
Intelligence & irreducibility
- Stephen Wolfram, Can AI Solve Science? (2024)
- Steve Omohundro, The Basic AI Drives (2008)
- Nick Bostrom, The Superintelligent Will, Minds and Machines 22 (2012)
Recent (Tier B — verify independently)
- Oliver Seidel, AGI/ASI event horizon bound (Zenodo, 2025–2026)