Source-linked AI summary
Survival Strategies
David A. Eubanks
TL;DR
The paper asks what conditions could allow a subject—from data to a civilization—to survive indefinitely with positive probability. It analyzes survival probabilistically and computationally, contrasting replication with intelligence. It concludes that redundancy can support survival under suitable assumptions, while intelligent systems face substantial prediction and self-modification barriers.
Problem
The paper asks what conditions let a subject of interest survive indefinitely, a question applicable to life, civilizations, and hard-drive data.
Source: §1 · pp. 1–2Method
Results
The paper concludes that multiple independent copies can make positive-probability indefinite survival feasible, whereas a single intelligent system must continually improve its environmental understanding and confront self-prediction.
Source: §6 · p. 10Takeaways & Limitations
Long-term survival of a civilization requires continuing attention to overall threats, including internal ones, and intelligent single systems may be rare relative to MIC-like systems.
Source: §6 · p. 11A constant probability of universal destruction makes redundancy insufficient, because the probability of survival then vanishes over time.
Source: §2.2 · p. 4
Abstract
from arXiv · showhide
This paper addresses the theoretical conditions necessary for some subject of study to survive forever. A probabilistic analysis leads to some prerequisite conditions for preserving, say, electronic data indefinitely into the future. The general analysis would also apply to a species, a civilization, or any subject of study, as long as there is a definition of "survival" available. A distinction emerges between two approaches to longevity: being many or being smart. Natural selection relies on the first method, whereas a civilization, individual, or other singular subject must rely on the latter. A computational model of survival incorporates the idea of Kolmogorov-type complexity for both strategies to illustrate the role of data analysis and information processing that may be required. The survival-through-intelligence strategy has problems when the subject can self-modify, which is illustrated with a link to Turing's Halting Problem. The paper concludes with comments on the Fermi Paradox.
1. Introduction
The paper reduces indefinite survival to probability and computation, applying the question to biological life, civilizations, and data. It sets out to identify conditions for positive-probability indefinite survival and to examine reproduction and computational complexity.
1. Introduction
- The paper applies its probabilistic and computational framing to life, civilizations, and hard-drive data.The introduction also notes speculation that human immortality may be technologically near.Source: §1 · p. 1
- Survival is treated as persistence in a specified state, with non-survival permanent, for biological or nonbiological subjects.Examples include individuals, species, machines, businesses, and civilizations.Source: §1 · p. 1
- The analysis asks what conditions let a subject survive indefinitely with probability greater than zero.It examines reproduction and the difficulties of surviving without it.Source: §1 · pp. 1–2
- A discrete-environment model assumes computation is necessary for survival because the subject must predict future environmental states.This framing permits consideration of the computational complexity of both the environment and the subject.Source: §1 · p. 2
2. Probabilities and Survival
With constant annual survival probability below one, indefinite survival has probability zero; positive probability requires survival chances to rise sufficiently. Redundancy can achieve this under independence, but a persistent shared extinction risk defeats it.
2. Probabilities and Survival
- A constant annual survival probability p < 1 gives zero probability of surviving forever, so conditional annual probabilities must increase over time.The section considers sequences whose infinite-horizon survival probability P is greater than zero.Source: §2 · p. 2, §2.1 · pp. 2–3
2.1. Annual Survival Probabilities.
- The years-at-risk quantity y = 1/(1 − b) equates an infinite sequence of increasing survival chances with a finite number of years at constant risk p.For b = .60, the equivalence is 2.5 years at constant probability p.Source: §2.1 · p. 2
- Slowly increasing conditional probabilities can yield P > 0, and smaller ε permits survivable sequences with arbitrarily large years at risk.For the sequence p_n = exp(−cn^-1−ε), the passage states that the relevant sum is finite and bounds years at risk by y ≥ 1/ε.Source: §2.1 · p. 3
2.2. Survival through Redundancy.
- A single extra independent copy per unit time can suffice for survivability, without exponential growth in copy numbers.For p = .5, the calculated survival probability is .289, versus a lower bound of P ≥ .287; the years at risk are y = 1.79.Source: §2.2 · pp. 3–4
- Redundancy relies on independent copy failures; any constant probability ε of total population extinction makes survival probability vanish as n → ∞.A successful program must eliminate constant probabilities of universal destruction, leaving only probabilities that continue to decline.Source: §2.2 · p. 4
3. Two Strategies for Survival
The paper contrasts survival through many diverse copies with the effort of a single system to improve its predictions. Redundancy reduces dependence on individual survival, while a nonreproducing system must process information to anticipate and avoid threats.
3. Two Strategies for Survival
- In a MIC, individual survival is secondary to multiplying and diversifying, and death occurs only if all copies are destroyed when any copy counts as survival.This strategy depends on treating survival of one copy as general survival.Source: §3 · p. 4
- Multiple Independent Copies (MIC) combines replication with variation, with diversity and separation helping reduce correlation among copies’ fates.The paper uses biological evolution as its conceptual model for MIC.Source: §3 · p. 4
- A Single Intelligent System (SIS) is a singular, nonreproducing subject that uses information processing to increase its conditional survival chances.“Intelligent” is used specifically for this survival-directed information processing, not its ordinary meaning.Source: §3 · p. 4
- Without MIC’s parallel discovery of local conditions, an SIS must safely gather information and predict the environment well enough to avoid death.Its conditional survival probabilities must approach one by improving its predictions; controlling the environment may also improve predictability.Source: §3 · p. 4
4. A Model for the Complexity of Survival
The paper models survival as a discrete game between a subject and an environment, using Kolmogorov complexity to compare their patterns. It contrasts reproduction with inductive prediction as ways to handle changing environments, concluding that both may support long-term survival but that reproduction is more robust.
Model
- The Survival Game represents environmental conditions and subject actions as zeros, ones, or blanks, with the goal of accumulating an unbounded lifespan.The model assumes the environment and subject can be represented as discrete symbols and computable operations.Source: §4 · p. 5
Complexity
- Kolmogorov complexity measures a pattern by the size of the smallest algorithmic description that reproduces it.The model applies this measure to both subject and environment.Source: §4 · pp. 5–6
Survivable subsequences
- A subject need not match the environment’s full complexity; survival requires predicting a subsequence with infinitely many survivable conditions.The subject must avoid every zero in that subsequence but need not act on every one.Source: §4 · p. 6
MIC strategy
- When an apparent environmental rule has exceptions, a MIC can survive by producing varied copies, allowing fitting variants to survive and multiply.The paper describes the MIC as an inductive machine that tests hypotheses, provided environmental rules do not change too quickly.Source: §4 · p. 7
SIS strategy
- A SIS must observe and experiment to form broader predictive rules, but unexpected environmental complexity can still cause it to die.Its success depends on induction predicting near-term events well enough, and it must continually look beyond its current survivable sequences.Source: §4 · pp. 7–8
Conclusion
- The paper considers both MIC and SIS viable for long-term survival, while judging MIC the more robust strategy.A SIS must devote much of its energy to observation, experimentation, prediction, and careful interaction with the environment.Source: §4 · p. 8
5. The Problem of Self-Prediction
Self-prediction complicates survival for an intelligent system (SIS): it must model itself as well as its environment, yet self-modification makes its future behavior difficult to guarantee. The section connects this problem to undecidability and an unknown probability penalty.
5. The Problem of Self-Prediction
- A SIS must model its own behavior as well as external conditions, but predicting itself would require a model more complex than itself.The section frames this as a self-prediction paradox.Source: §5 · p. 8
- Because a SIS is part of its environment, it must also manage environmental complexity and potential threats it creates itself.The paper argues that physical and logical self-modification can make this prediction challenge especially difficult.Source: §5 · p. 8
- A built-in survival imperative may fail because an intelligent robot can reverse-engineer and replace it, while self-modification may be needed to meet future conditions.The paper also compares this to humans modifying their own behavior, while noting they may not anticipate every condition in which they want to remain alert.Source: §5 · p. 9
- For an algorithmic SIS, determining whether self-modification will lead to suicide is a halting problem with no general solution; Rice’s Theorem further limits what can be known about such computations.The section presents program-termination tools for particular software projects as a possible, limited aid, not a general solution.Source: §5 · pp. 9–10
- The unknown Ω factor lowers a SIS’s survival probability to an unknown upper bound, but does not rule out an arbitrarily long-lived SIS.The paper applies the factor when SIS intelligence derives from computable behavior [p16].Source: §5 · p. 10
6. Conclusions
The conclusions contrast the feasible survival of multiple independent copies with the harder task of a single intelligent system (SIS), which must keep improving its survival odds and guard against self-destruction. They extend this contrast to civilization, the Fermi Paradox, and possible self-replicating probes.
6. Conclusions
- Multiple independent copies can achieve indefinite survival if any one copy counts as success, while a SIS must improve its survival probabilities through better understanding of its environment.The paper also identifies SIS self-prediction as an additional obstacle to avoiding self-destruction.Source: §6 · p. 10
- A SIS’s long-term survival requires continually increasing its conditional survival probabilities and scrutinizing how it can self-modify.The paper contrasts this with the comparative advantages of multiple independent copies.Source: §6 · p. 11
- A civilization seeking longevity must attend continually to threats to overall survival, including internal threats.The paper suggests that SIS rarity relative to multiple-copy-like systems may bear on the Fermi Paradox.Source: §6 · p. 11
- Self-replicating probes face a trade-off between mutation that may produce local ecologies and suppressing mutation at the cost of adaptability.The paper suggests either outcome could hinder interstellar propagation.Source: §6 · p. 11
- The paper extends its survival argument to democratic government, which should attend to the risk of self-destructing into tyranny as well as external threats.Source: §6 · p. 11
- For personal longevity, the conclusion recommends continually improving one’s chances of avoiding automobile accidents, while noting that finite need not mean small.Source: §6 · p. 11