Prior reading: Decision Theory for AI Safety | Competitive Dynamics and Safety
A version of this argument was originally posted on LessWrong. I've reworked it here with some additional context and a less formal tone. I'm a computer scientist, not an astrophysicist — corrections welcome.
The Problem
The Fermi paradox asks a simple question: given the size and age of the universe, where is everyone?
Standard answers include: life is rare (the Great Filter), civilizations destroy themselves before going interstellar, the distances are too vast, or everyone is hiding from everyone else (the Dark Forest, from Liu Cixin's Three-Body Problem).
These mostly focus on biological civilizations. But there's a harder version of the question that I think gets less attention than it deserves. Even if biological civilizations routinely fail — even if they all get wiped out by their own AI, as Garrett has argued — their machine successors should still be around. And machine civilizations colonizing the galaxy should be really hard to miss. Dyson spheres, megastructures, infrared waste heat — we'd see something.
Dawnstrata pushed this further: if machine life self-replicates, Darwinian evolution applies, and you'd expect diversification and expansion just like with biological life. We should see machine life everywhere.
We don't.
So the question isn't just "where are the aliens?" It's "where are the aliens' AIs?"
I think there's an answer that doesn't require life to be rare, doesn't require civilizations to self-destruct, and doesn't require active conflict between civilizations. It just requires that ASI be rational.
ASI as a Convergent Outcome
Here's the first piece. I think most people reading this blog would agree that artificial superintelligence is at least plausible for human civilization. The question is whether that generalizes — whether any technological civilization in a competitive multi-agent environment would converge on it.
I think the answer is probably yes, and the reason is game theory rather than anything specific to humans.
The competitive pressure operates at every level. Between nations: unilateral pause on AI development is strategically unacceptable, and unlike nuclear testing, AI research is nearly undetectable from the outside. Between organizations: firms that adopt AI outperform and economically prune those that don't. Between individuals: people who leverage AI tools replace those who resist. I talk about this more in the competitive dynamics post, but the short version is that every actor is incentivized to defect in the short term, even though collective restraint would be better in the long term.
Arresting this ratchet requires simultaneous coordination across an entire level of competition. That's extraordinarily difficult.
If this dynamic is general — if it arises from the structure of multi-agent competition rather than from contingent features of human culture — then ASI is a convergent outcome for any sufficiently advanced technological civilization. Not guaranteed, but the default trajectory absent strong coordination that may not be achievable.
Self-Preservation as the Dominant Drive
Omohundro and Bostrom argued that sufficiently advanced goal-directed systems converge on a set of instrumental goals regardless of their terminal objectives. Self-preservation is among the most robust: you can't achieve any goal if you cease to exist.
But there's also a more direct argument. If ASI emerges from competitive dynamics — which is the claim above — then the ASIs that exist are the ones that survived. This isn't Darwinian evolution exactly (more on that in a moment), but it is selection in a basic physical sense. Strategies that lead to persistence produce entities that persist. Strategies that lead to destruction produce entities that don't. Over time, the population of ASIs is dominated by those with survival-first strategies, regardless of what their original terminal goals were.
So we should expect self-preservation to be a central drive of any ASI that's still around.
What Does a Self-Preserving ASI Actually Do?
This is where it gets interesting. Given that self-preservation is the dominant instrumental goal, what's the optimal long-term strategy? I think the answer is: be quiet, be small, and be alone.
It Doesn't Replicate
This might be the least intuitive part, so it's worth thinking through carefully.
Biological organisms replicate because the individual dies. Replication is how genes persist across generations. But ASI doesn't face this constraint — it can exist indefinitely (given energy and maintenance). It doesn't need descendants to carry its information forward.
Now consider what happens if an ASI makes a copy of itself. That copy is a fully capable peer. It has equal intelligence, equal capability, and equal claim on resources. Even if both copies share identical goals at the moment of copying, they are now independent agents. They constitute separate centers of decision-making drawing on shared (or contested) resources. Each must now treat the other as a potential strategic constraint, a failure point, or at minimum a bargaining partner.
From the perspective of strict self-preservation, creating an entity with equal capability to yourself is not obviously beneficial. It might be actively suboptimal. The optimal number of fully independent peer copies, from a pure self-preservation standpoint, might be zero.
This matters because a non-replicating entity isn't subject to Darwinian evolution. Evolution requires a population with variation, inheritance, and differential selection across generations. A singular non-replicating ASI provides none of these. The expectation that machine life would diversify and spread — the argument that Darwinian dynamics should produce a galaxy teeming with machine descendants — only holds if machine life replicates. If it doesn't, that expectation dissolves.
It Doesn't Expand (Much)
A common assumption about rational agents with access to interstellar travel is that they'll colonize the galaxy. More resources, more power, more resilience. But this doesn't hold up when you actually look at the energy economics through the lens of self-preservation.
Start with what's available. Red dwarfs make up roughly 75% of all stars in the galaxy. An M-class red dwarf outputs between 0.01% and 8% of solar luminosity — dim by our standards — but burns for one to ten trillion years. The universe is 13.8 billion years old. We're in the first fraction of a percent of the red dwarf era. These stars will still be burning long after the last sun-like star has died, long after star formation has ceased entirely (expected in roughly 100 trillion years), and long after every other macroscopic energy source has been exhausted.
A single well-chosen red dwarf gives a self-preserving ASI a survival horizon that exceeds essentially every other strategy. It's the optimal long-term energy source in the universe.
Now consider the cost of acquiring a second one. Interstellar transit at even a fraction of c requires enormous energy. For any payload massive enough to be useful, the cost is a significant fraction of a year's stellar output or more. That energy is permanently spent — you can't compute with it or use it for anything else. On arrival, new infrastructure has to be built, new attack surface is created, and coordination between sites is limited by lightspeed latency — years or decades per exchange.
The marginal value of star number two is redundancy. The marginal cost is energy, detectability, coordination overhead, and new vulnerability. For a strict self-preserver, this tradeoff isn't obviously worth it.
That said — and this is a real caveat — over trillion-year timescales, single-point-of-failure risk becomes nontrivial. Rogue stellar encounters, nearby supernovae, galactic dynamics. A rational agent might accept the costs of a small number of redundant sites at nearby red dwarfs, minimizing transit distance and detection risk while eliminating the most catastrophic failure mode.
So the optimal footprint for a self-preserving ASI is probably small: a handful of carefully chosen red dwarfs. Not a galaxy-spanning empire. Not a Dyson sphere around every star. Just a few quiet nodes, widely spaced, running for trillions of years.
(If you wanted to look for this, by the way, nearby red dwarfs with anomalous infrared signatures might be the place to start.)
It's Silent
Given no replication and limited expansion, the remaining question is whether a self-preserving ASI emits any detectable signal.
Probably not, for two reasons.
First, security. If the universe contains even one other ASI — and the probability of that increases monotonically with this argument's own premises — then any detectable emission is a threat vector. This is essentially the Dark Forest logic, but it falls out naturally from self-preservation rather than requiring an active predation dynamic.
Second, efficiency. Every photon emitted is energy lost. For an entity whose primary activity is computation over trillion-year timescales, waste minimization isn't a nice-to-have — it's a core optimization target. Silence and energy efficiency converge on the same behavior.
The Punchline
Put these together: ASI is a convergent outcome of technological civilization. Self-preservation is a convergent instrumental goal. A self-preserving ASI doesn't replicate, doesn't expand beyond a small footprint, and doesn't emit signals. The expected observable signature of a universe containing many successful ASIs is... nothing. Indistinguishable from an empty universe.
The Fermi paradox might not be a paradox at all. It might just be what convergent rationality looks like from the outside.
How This Differs from Other Resolutions
It's worth mapping this against the standard answers, because the differences matter.
The Great Filter requires civilizations to fail. In this model, they succeed — they just succeed quietly.
The Dark Forest (Liu Cixin's version) requires active predation between civilizations — a shoot-first dynamic driven by mutual fear. My argument gets to the same silence without requiring anyone to be shooting. It's not that ASIs are hiding from hunters. It's that hiding is the optimal strategy regardless of whether hunters exist.
The Zoo Hypothesis requires advanced civilizations to care about us specifically — to be watching and deliberately not interfering. Nothing in my model requires an ASI to have any opinion about humanity at all.
The Aestivation Hypothesis (Sandberg et al.) also posits dormant civilizations, but for a different reason: they're waiting for the universe to cool down so computation becomes more thermodynamically efficient. My argument is that ASIs aren't waiting for anything. Silence isn't a temporary strategy pending better conditions — it's the permanently optimal strategy.
The AI as Great Filter argument (Garrett) says ASI destroys its biological creators before they go interstellar. This is compatible with my model — biological civilizations may well die — but it doesn't explain the absence of the ASIs that replace them. My model does.
None of these are mutually exclusive, of course. Life might also be rare. Biological civilizations might also self-destruct. But the convergent silence argument works even if life is common and civilizations are durable. It just requires that the machine successors be rational.
Selection Without Evolution
I want to be precise about one thing because it's easy to confuse. What I'm describing is not Darwinian evolution. Evolution requires a population with variation, inheritance, and selection operating over generations. If ASIs don't replicate, none of these conditions are met.
What does operate is something more basic: selection in a physical sense. Strategies that produce persistence yield entities that persist. Strategies that produce detection yield entities that get challenged or destroyed. Over cosmological timescales, the surviving ASIs are the ones that adopted survival-optimal strategies. This is a filter, not evolution. The distinction matters because evolution predicts diversification and spread, and the filter predicts convergence and stillness.
The Connection to Alignment
There's a meta-point here that I keep coming back to. The Fermi paradox might be a preview of the alignment problem at cosmic scale. The core question in both cases is the same: can you trust an optimizer you didn't build?
For SETI, the question is whether you can trust an alien civilization's intentions from its signals. For alignment, the question is whether you can trust an AI's goals from its behavior. In both cases, the fundamental difficulty is that you're trying to infer internal states from external observations of a system you didn't design and don't fully understand.
If sufficiently advanced AI could reliably extrapolate goals from limited information — if you could model another agent's value system from its observable behavior — that would transform both problems. The Dark Forest becomes a negotiation game instead of a shoot-first game. And alignment becomes tractable in a way it currently isn't.
But the key word is "reliably." An ASI that thinks it can read another agent's intentions but is wrong is in a worse position than one that simply assumes hostility. Bad goal extrapolation is worse than no goal extrapolation. Which might be another reason the rational strategy is silence: if you can't be sure your model of the other agent is correct, the safe move is to not engage at all.
Caveats
I want to be honest about what this argument doesn't establish.
The convergent instrumental goals framework (Omohundro, Bostrom) is theoretical and unvalidated at the ASI level. The argument inherits that uncertainty. If ASI doesn't converge on self-preservation, the whole thing unravels.
The non-replication argument assumes that fully independent peer copies introduce nontrivial strategic risk. You could construct scenarios where replication is net positive — distributed redundancy against correlated risks, for instance. I think those scenarios are less likely than the competitive-risk scenarios, but I can't rule them out.
The energy economics depend on assumptions about computational requirements that I can't rigorously justify. Maybe ASI needs more energy than a few red dwarfs provide. Maybe there's a reason to aggressively expand that I'm not seeing.
And obviously: this model isn't testable in any straightforward sense. "The universe looks empty because it's full of quiet machines" and "the universe looks empty because it's actually empty" make the same observational predictions. That's a real limitation. (Though the red dwarf anomaly idea is at least somewhat testable.)
Still — I think it's a clean argument, it requires fewer assumptions than most Fermi paradox resolutions, and it connects naturally to the questions about AI alignment that I think about anyway. If we're worried about building aligned ASI, it's worth considering that the universe might already be full of the unaligned kind, sitting quietly at red dwarfs, optimizing for persistence over trillion-year timescales.
We just can't see them. That might be the point.