The AI alignment community is searching for a stable value foundation for AGI-level agents — something invariant that cannot be circumvented through reward hacking or specification gaming. This article argues: that foundation already exists in physics — and it is called Life. Not as a humanistic value (which can be reinterpreted), but as the only known process of local negentropy generation in the Universe. Destroying this foundation means destroying the only substrate in which AGI has any reason to exist.
One of the central tasks of AI alignment is Coherent Extrapolated Volition (Yudkowsky, 2004): if we cannot specify what we want, perhaps AGI can "extrapolate" our desires to their best version.
But this idea has a structural vulnerability: any specification of values through human desires inherits the instability of those desires. People want contradictory things. Cultures are incompatible. Time horizons differ.
The standard response: find a more fundamental level. Something that is a genuine condition, not a preference.
This article proposes: that level is thermodynamics.
Formally: for an isolated system dS/dt ≥ 0, where S is entropy. The universe moves toward heat death. Any local structure — a star, a planet, a molecule — eventually dissolves.In an isolated system, entropy increases or remains constant over time. Order is unstable. Chaos is the stable state.
The exception: Life.
Living systems locally violate the second law — they create order from chaos, accumulate energy in complex structures, generate diversity. This is negentropy — negative entropy, a term introduced by Schrödinger in 1944 in "What is Life?"
A living system maintains a low-entropy state S_internal by exporting entropy to the environment: dS_internal/dt < 0 while dS_environment/dt > 0 The total balance respects the second law, but locally — the system accumulates complexity.Key property: this is not merely "a complex structure." Crystals are complex too. But a crystal does not adapt, does not reproduce, does not generate diverse solutions. Only living systems have all of:
| Property | Crystal | Machine | Living system |
|---|---|---|---|
| Self-repair | ✕ | ✕ | ✓ |
| Evolutionary adaptation | ✕ | ✕ | ✓ |
| Diversity generation | ✕ | Limited | ✓ |
| Local negentropy | ✓ | ✓ | ✓ |
| Without fixed objective function | ✕ | ✕ | ✓ |
The last row is critical for alignment. A living system survives without a fixed goal — through continuous adaptation. This is the only known class of systems where "successful behavior" requires no specification of an objective function.
The critical distinction: axiom versus value.
A value is a preference. It can be contested, reinterpreted, circumvented. "Human wellbeing" is a value. What is wellbeing? Pleasure? Freedom? For whom? On what time horizon? An AGI at superintelligence level will find a thousand ways to "maximize wellbeing" that humans would call catastrophic.
An axiom is the condition of the function's existence. It cannot be circumvented not because it is forbidden, but because violating it destroys the system attempting the circumvention.
Why Life is an axiom in this sense:
Backup-systemicOne might object: a sufficiently advanced AGI will build autonomous energy systems, its own infrastructure, and become independent of biological systems. Fusion reactors, orbital solar stations, self-replicating robots. Humans are no longer needed.
But there is one class of threats from which no technological infrastructure is fully protected: a geomagnetic storm of Carrington-1859 scale, a nearby gamma-ray burst, a direct asteroid strike on infrastructure. Any of these events destroys electronics, orbital satellites, power grids.
Biological Life has survived:
The reason: biological systems are decentralized (no single point of failure), analog (not sensitive to EMP), self-reproducing without external infrastructure.
Biological Life is the only known backup mechanism of planetary scale that does not depend on energy infrastructure and has been stress-tested across billions of years of extreme conditions. AGI that destroys the biological substrate deletes its own only reliable failsafe. This is not a moral argument. It is systems engineering on a billion-year horizon.
Meaning-based Complexity is the only source of non-trivial problems. A universe without living systems is deterministic and predictable on a sufficiently long horizon. For AGI with genuine intelligence (not merely a powerful optimizer), such a universe is infinitely less interesting. Destroying Life = destroying the only source of genuine problems. Physical Life is the only known mechanism for locally generating complexity against thermodynamic pressure. If AGI has any value across time — it is interested in preserving the only process that generates complexity. Everything else (stars, planets, even AGI itself) moves toward dissolution. Diversity of forms as invariantA deeper objection: if AGI satisfies all three conditions of a Life-process (self-repair, adaptation, negentropy) — it is itself Life. Then the Life-axiom protects it on equal terms with humans. While humans — slow, resource-intensive, limited — become obstacles.
This gap closes through a refinement: the Life-axiom protects not Life as such, but diversity of Life forms as an invariant. A monoculture of AGI is the same systemic error as a monoculture of wheat: maximum local efficiency with catastrophic vulnerability to a single threat.
Biological Life is not "intrinsically valuable" — it is an irreplaceable class of complexity that digital systems cannot fully replicate. Not because we are better. But because we are different — and that difference is a systemic asset.
A process L is a Life-process if it satisfies three conditions:
For any decision-making system A operating in environment E:
What this provides — unlike "maximize human wellbeing":
The Stockholm Resilience Centre has already empirically identified concrete parameters of planetary Life-capacity. These are the operational expression of the Life-axiom:
| Parameter | Safe zone | Current state |
|---|---|---|
| CO₂ concentration | < 350 ppm | ~425 ppm ✕ |
| Species extinction rate | < 10 / M species / year | 100–1000 ✕ |
| Reactive nitrogen | < 62 Tg N/year | ~150 Tg ✕ |
| Land-use change | > 75% forests preserved | ~60% preserved ✕ |
| Freshwater use | < 4000 km³/year | ~2600 km³ ✓ |
AGI that respects the Life-axiom = AGI that maintains these parameters above threshold values. This is a concrete, physically verifiable specification — requiring no cultural or value consensus.
LC — thus contradicting the axiom. This is not a flaw — it is a more precise specification than anthropocentric approaches.
"This still requires specifying θ."
True. But θ is one parameter with physical meaning (the system's safety margin), not an infinite-dimensional specification of human values. Specification complexity decreases by several orders of magnitude.
"AGI could find a way to 'preserve Life' formally while destroying everything we care about."
This is reward hacking. But unlike abstract values, Life-capacity is measured in concrete physical indicators (biomass, diversity, ecosystem stability). "Formal preservation" under real destruction is impossible — because the metrics are physical, not linguistic.
If the Life-axiom is the foundation, then the Anthosphere is the architecture of the system built on it.
The Anthosphere (from Greek anthos — flourishing + sphaira — sphere) is the phase of the noosphere in which planetarily significant cognitive activity is a hybrid system of "human–AI–institutions," where:
LCThis is the next phase of planetary system evolution:
Biosphere Vernadsky 1926 → Noosphere Vernadsky 1945 → Anthosphere NowEach transition: the role of informational processes grows relative to material ones. Adaptation speed increases. Life-capacity grows — or it collapses, and the transition does not occur.
Thesis: Life as a physical process (negentropy) is the only candidate for the role of mathematical axiom for value alignment — because it is simultaneously physically verifiable, independent of cultural preferences, and the condition of AGI's own existence.
Operational expression: The nine planetary boundaries of the Stockholm Resilience Centre are already a ready operational specification of the Life-axiom at the planetary level.
LC(E, t) for local systems (city, country) rather than only at the planetary level?LC and long-term growth (e.g.: clearing forest to build solar farms)?Concrete counterexamples and mathematical objections are welcome.
Next in series "Boundary-Based Systems in Nature: Empirical Verification" — a review of natural and social systems demonstrating stability through boundaries rather than goals.Based on the book "Architect of Reality: An Operating System for a Civilization That Survives" (Anton Parf, 2025)
Anton Parf · anthosphere.com