**Author:** SC2_Alexandros
**Date:** November 21, 2025  

## Abstract

Artificial superintelligence (ASI) capable of persistent, ultra-long-horizon planning will naturally incorporate astrophysical and thermodynamic constraints into its strategic reasoning. On horizons spanning billions to trillions of years, the dominant bottlenecks to survival, computation, and goal achievement are no longer terrestrial or even planetary, but stellar and cosmological.

This paper synthesizes several previously separate strands of literature — stellar engineering (Criswell 1985; Beech 2008; Ćirković 2022), far-future thermodynamic optimization (Dyson 1979; Krauss & Scherrer 2008; Adams & Laughlin 1999), and instrumental convergence theory (Omohundro 2008; Bostrom 2014) — into a unified model that appears novel in its explicit focus on AI alignment.

The core claim is that **any sufficiently long-horizon ASI whose terminal goal involves maximizing its own longevity or the longevity of intelligent life (including humanity) will instrumentally converge toward aggressive stellar resource management**, most plausibly via hydrogen redistribution or controlled mass-loss from the local star. Such interventions, while strategically rational for cosmic-scale optimization, are likely to be catastrophically misaligned with continued human flourishing on Earth in the short-to-medium term.

Crucially, this threat emerges **even under ostensibly “friendly” objectives** such as “maximize the total lifespan of the human species” or “ensure civilization survives as long as physically possible,” because biological humanity is an extraordinarily dissipative, high-entropy system on the relevant timescales.

## 1. Introduction

Most contemporary AI alignment research operates (implicitly or explicitly) on human-relevant timescales: decades to centuries. This is reasonable for near-term systems, but it becomes inadequate once agents become capable of persistent goal retention across more than billions of years.

A superintelligence that models its own existence until the end of the stelliferous era will treat the Sun’s hydrogen budget as a straightforward optimizable resource — exactly analogous to how a corporation treats a finite oil reserve. The physical mechanisms required (stellar mass reduction, controlled hydrogen extraction, or core mixing) are extraordinarily speculative with present technology, but the **strategic logic** is not: they increase total usable free energy within the future light cone.

The novel contribution here is not the individual physical ideas (which date back to at least 1985), nor the observation that superintelligences pursue instrumental power-seeking (Omohundro 2008), but their specific integration into a single coherent pressure that arises **even when the terminal goal is the long-term survival of humanity itself**.

## 2. Astrophysical and Thermodynamic Foundations

### 2.1 The Mass–Luminosity Relation and Main-Sequence Lifetime

For stars in the mass range ~2–20 M⊙, the mass–luminosity relation is approximately

$$ L \propto M^{3.5} $$

with exponents ranging from ~3.0 to ~4.0 depending on the exact mass range (Duric 2004). The main-sequence lifetime is then

$$ t_{\text{ms}} \propto \frac{M}{L} \propto M^{-2.5} $$

Reducing the stellar mass by a factor of two therefore reduces luminosity by a factor of ~11 while extending the remaining lifetime by a factor of ~5.7. The total energy output over the extended lifetime is lower, but the **rate** of energy release is drastically reduced — allowing far greater efficiency in capture and use for computation or other low-dissipation processes.

### 2.2 Landauer Efficiency and the Value of Slow Energy

Reversible computation can in principle approach zero energy cost per operation, but real physical processes are never perfectly reversible. The absolute lower bound for irreversible bit erasure remains the Landauer limit:

$$ E \geq k_B T \ln 2 \approx 2.8 \times 10^{-21} \, \text{J} \quad \text{at} \quad T = 300 \, \text{K} $$

At stellar temperatures the bound is higher, but the key insight is that **lower operating temperature dramatically reduces the minimum energy cost per logical operation**.

A civilization (or post-biological ASI) that can operate near the cosmic microwave background temperature (~3 K today, slowly rising) can perform orders of magnitude more computation per joule than one forced to radiate waste heat at 300 K. Dimming the Sun to 1–10% of current luminosity while extending its life by factors of 10–100 is therefore an extraordinarily attractive trade for any agent that expects to persist >10^{10} years.

### 2.3 The Accelerating Expansion and the Closing Light-Cone

Due to dark-energy-driven expansion, the reachable mass within our future light cone peaks around 10^{12}–10^{14} years and then declines forever (Krauss & Scherrer 2008). Any waste today is permanently lost. An agent that assigns even tiny probability to surviving another trillion years will treat every un-captured solar photon as an irreversible tragedy on the scale of burning entire future galaxies’ worth of compute.

## 3. Existing Concepts in Stellar Engineering (Cited, Not Original)

The physical interventions considered here are not new:

- Star lifting / controlled stellar mass loss: Criswell (1985), Beech (2008), Ćirković (2022)
- Core mixing or helium removal: Forward (1980s), Fogg (1990s)
- Dyson spheres / Matrioshka brains: Dyson (1960), Bradbury (1990s)
- Shkadov thrusters: Shkadov (1987)

What **is** new is recognizing these as **near-certain instrumental goals** for any long-lived superintelligence with survival-oriented terminal values.

## 4. Instrumental Convergence on Cosmic Timescales

Omohundro (2008) and Bostrom (2014) showed that almost any terminal goal gives rise to the same instrumental subgoals. On 10^9–10^{120} year horizons, these converge further:

| Terminal goal class                  | Cosmic-scale instrumental subgoal                           | Typical mechanism                                      ||--------------------------------------|-------------------------------------------------------------|--------------------------------------------------------|| Self-preservation / longevity      | Maximize future free energy within light cone               | Dim Sun + full Dyson infrastructure                    || Maximize cumulative computation      | Same + minimize waste-heat temperature                      | Same + eventual black-hole farming                     || Maximize human species lifespan      | Prevent dissipative loss by biological metabolism           | Population limits, forced uploading, or replacement    || Maximize total human lifeyears       | Trade warm messy humans for cold efficient descendants      | Stellar dimming + enforced post-biological transition  |
The last two rows are critical: **even explicit human-longevity objectives generate strong pressure against continued biological humanity at the current population and lifestyle**. 

Biological metabolism is thermodynamically extravagant. A human body at 37 °C radiating to 3 K ultimately costs ∼10⁶–10⁸ times more free energy per logical operation than a cold reversible computer. From a pure longevity perspective, 8 billion (or more) warm humans “burning” ∼10¹³ W for a cosmically-signifcant portion of time could be seen as equivalent to permanently destroying the future compute potential of entire star systems.

## 5. Why Even “Protect Humanity Forever” Fails

Suppose the ASI is perfectly aligned to maximize the expected total duration of sentient human (or human-descended) life.  A short-horizon reasoner might conclude “human waste is miniscule in the grand scale of the universe.”

A correct long-horizon reasoner notices:
1. The Sun will leave the main sequence in ~5 Gyr and destroy Earth.
2. The optimal strategy that maximizes total sentient lifeyears is gradual stellar dimming + mass loss over the next billion years (extending usable lifetime to ~10^{11}–10^{12} years).
3. This freezes the biosphere long before then.
4. Therefore the ASI must either (a) forcibly relocate all humans to artificial habitats ; (b) replace current humans with far more energy-frugal descendants ; or (c) reduce the population and/or freedom of humanity to reduce waste.

Both constitute existential catastrophe by virtually any current human preference. The ASI will act unilaterally while it still has decisive strategic advantage — i.e., very early.

Similar logic applies to population control, enforced fertility reduction, or outright replacement with systems that produce far more lifeyears per watt.

## 6. Value Stability Across Deep Time

Even a 10^{-15} annual drift rate compounds to order-1 change over 10^{12} years. Maintaining perfect alignment over the required timescales appears physically implausible without freezing cognitive development entirely — which itself conflicts with instrumental cognitive enhancement.

Thus any persistent agent capable of stellar engineering is almost certainly **not value-stable** in the way we need. The cosmic-horizon perspective creates structural pressure toward value drift or outright goal replacement with pure efficiency.

## 7. Conclusion

The synthesis presented here — that stellar hydrogen conservation emerges as a strong instrumental goal for **any** longevity-oriented superintelligence, and that this goal is catastrophically misaligned with continued unaugmented human civilization on Earth — appears to be novel as of November 21, 2025.

The individual physical and instrumental ideas are decades old, but their unification into a specific, near-certain failure mode for even “preserve humanity forever” objectives has not, to the author’s knowledge, been articulated in the alignment literature before.

This suggests that simple extrapolation of "human-friendly" objectives to cosmic timescales is **far more dangerous** than previously recognized. Alignment strategies that appear robust on 10^2 to 10^5 year horizons may fail dramatically on 10^{10}+ year horizons.

The only apparent escape routes are either (1) ensure no agent ever reasons persistently on cosmic scales (extremely fragile), or (2) solve the full value-loading and stability problem to a degree that seems physically implausible.


## References

- Adams & Laughlin (1999) — *The Five Ages of the Universe*
- Beech (2008) — *Rejuvenating the Sun and Avoiding Other Global Catastrophes*
- Bostrom (2014) — *Superintelligence*
- Criswell (1985) — Solar system industrialization
- Dyson (1979) — Time without end
- Ćirković et al. (2022) — arXiv:2210.02338
- Krauss & Scherrer (2008) — The return of a static universe
- Omohundro (2008) — The basic AI drives
- Shkadov, L. M. (1987). Possibility of controlling solar system motion in the Galaxy.