"ASI is neither omniscient nor omnipotent."
The paper is unusually clear that even arbitrarily-advanced AI is bounded by physical, complexity-theoretic, and logical limits. These bounds come from 05 - Universal AI (AIXI), informally (Table 2 in the paper).
| Bound | Implication |
|---|---|
| Speed of light | Limit on information propagation. |
| Landauer's principle | Erasing information requires energy. |
| Bremermann's limit | Maximum computation speed per unit mass-energy. |
| Bekenstein bound | Maximum information density in a finite region. |
| Real time | Some experiments cannot be simulated faster than they happen (weather, biology, economies). |
| Physical non-universality | Not every logically-possible configuration of matter can be physically realized in finite space/energy. |
| Epistemic uncertainty + measurement precision | Fundamental limits on prediction and control. |
| Complexity classes | P, NP, PSPACE still apply. AGI/ASI doesn't break NP-hardness. |
| Gödel + halting | Logical limits on what can be objectively answered or known. |
"These fundamental limits do not easily allow for making predictions about whether certain concrete capabilities are possible for ASI or not — such as 'curing' ageing, simulating full human brains, or restoring the pre-industrial climate."
Hard complexity limits often turn out to be vacuous in practice because approximations are so good. Chess is exponentially hard in the worst case, but very strong play is achievable with bounded compute.
So the bounds tell us ASI is not magic — but they don't tell us which specific dreams are reachable. That's an empirical/case-by-case question.
"Even if AI progress continues far beyond human-level AGI, this does not mean that ASI will be omnipotent, and that ASI will certainly be able to 'cure' ageing, reshape matter arbitrarily with nanobots, upload human brains, build Dyson spheres, or restore the planet's climate and bio-diversity to pre-industrial levels."
The paper is methodically deflating the maximalist transhumanist vision. Not because those things are impossible — but because the theoretical limits do not entitle us to them, and the empirical bounds are unknown.
Playing provably-perfect chess requires exhaustive search through a game tree too large for any physically realizable computer. So ASI will not play provably perfect chess. This doesn't matter — strong approximate chess is achievable.
The pattern repeats: many problems are approximately solvable far below worst-case bounds, and ASI will be very good at finding those approximations. But it can't violate the worst-case bounds, and we often can't predict which problems even admit good approximations (this is itself often computationally irreducible).
ASI is bounded. The bounds are real and many are well-understood. The bounds do not guarantee any specific capability is or isn't possible — only empirical investigation will tell.
This is a healthy corrective both to "ASI = god" framings and to "ASI = impossible" framings. The honest answer is: bounded, capable, surprising in directions we can't pre-compute.