From Dust to Desire: The Physical Origins of Agency and the Future of AI
- Aki Kakko
- 6 hours ago
- 7 min read
One of the most profound questions in science and philosophy is the "Problem of Agency." The universe involves fundamental laws of physics that describe how matter interacts—atoms collide, gravity pulls, energy disperses. In this mechanical universe, things simply happen. Yet, living beings are different.
Living things don’t just happen; they act.
A rock rolls down a hill because gravity forces it to. A wolf runs up a hill because it wants to catch a rabbit. Where does this "wanting" come from? It is not a mystical force injected into the universe; it is a structural property of matter, built layer by layer. This is the story of how dead matter organized itself into systems that care about their own future, and why this specific hierarchy is the missing key to Artificial General Intelligence (AGI).
The Thermodynamic Origin (The "Why")
Agency as a Rebellion Against Decay:
To understand agency, we must first understand its enemy: Entropy. The Second Law of Thermodynamics states that closed systems tend toward disorder and equilibrium. If you leave a machine alone, it rusts. If you leave a hot drink, it cools. The inevitable fate of all matter is to stop moving and blend into the background.
Dissipative Structures:
In the mid-20th century, physicists like Erwin Schrödinger and Ilya Prigogine realized that life is a "dissipative structure." It is a system that maintains its shape by constantly consuming energy and dissipating entropy (waste) into the environment.
Analogy: Think of a whirlpool in a river. The water is constantly flowing through it, but the structure of the whirlpool remains stable.
The Spark: Agency begins here. It is not "free will" yet; it is a physical necessity. To exist, the system must channel energy. If the flow stops, the structure vanishes.
Agency is the mandatory behavior required to delay the Second Law of Thermodynamics.
The Chemical Origin (The "How")
Closing the Loop and Building the Wall:
Energy flow alone creates tornadoes, not turtles. For true biological agency, the energy needs to be trapped in a cycle.
Constraint Closure (The Engine):
Stuart Kauffman and Terrence Deacon describe the origin of life as the moment chemical reactions achieve "Constraint Closure."
Imagine Reaction A produces Chemical B.
Reaction B produces Chemical C.
Reaction C produces a catalyst for Reaction A.
The Result: The system becomes a self-sustaining loop. It is no longer passive; it has a "job"—to keep the cycle spinning. This is the birth of teleology (purpose). The "purpose" of C is to make A, so that the whole system survives.
Autopoiesis (The Boundary):
Biologists Humberto Maturana and Francisco Varela added the final piece: The Membrane. A chemical loop needs protection from the chaotic outside world. It builds a semi-permeable boundary (a cell membrane).
The Genetic Origin (The "Memory")
Freezing Agency in Deep Time:
If a single cell had to figure out how to process sunlight or digest sugar from scratch every time it was born, life would have ended billions of years ago. The system needed a way to record its victories.
There is a debate in biology: Are genes the "masters" (The Selfish Gene) or just a "library"?In the context of agency, genes act as temporal bridges. They are not agents themselves (DNA doesn't "want" anything in the moment); they are the record of past agency.
When an ancestor organism successfully mutated its metabolic loop to survive a heatwave, that success was encoded in nucleic acid.
The Heuristic Shortcut: Genes provide the organism with priors (pre-loaded knowledge). You breathe automatically not because you decided to, but because billions of years of ancestors "voted" that breathing is a good idea for survival.
The Hardware/Software Split:
The Cell (The Phenotype) is the agent acting in real-time.
The Gene (The Genotype) is the blueprint that constrains what the agent can do.
Synthesis: Genes do not remove agency; they enable it by providing the sophisticated hardware (eyes, nerves, muscles) that allows the agent to interact with the world.
The Cognitive Origin (The "Action")
How does a complex organism use this machinery to navigate the world? The leading theory today is the Free Energy Principle (championed by Karl Friston).
The Prediction Machine:
A biological agent is a prediction machine. It holds an internal model of what the world should be like (e.g., "I should be at 37°C body temperature").
When the senses report data that conflicts with the model (e.g., "It is 0°C outside"), the system registers Surprise (informational entropy). To resolve this surprise, the agent has two choices:
Change the Mind: Accept that freezing to death is the new normal (bad for survival).
Change the World: Act. Put on a coat. Build a fire. Shiver.
Human agency is the process of acting on the world to force it to align with our internal requirements for survival.
Why This Matters for AI Researchers
The "Ghost" in the Machine is Thermodynamic:
We are currently witnessing an explosion in Artificial Intelligence. However, LLMs (Large Language Models) like GPT-series models possess vast intelligence but zero true agency. The hierarchy above explains why, and what is missing.
The Lack of Thermodynamic Cost:
Biological agency is driven by the threat of death (entropy). An amoeba must find food, or it disintegrates. Current AI: Has no "skin in the game." If you turn off the server, the AI does not fight you. It has no homeostatic state to maintain. Without the threat of non-existence, there is no intrinsic drive to act.
The Boundary Problem:
Biological agency requires a "Self" distinguished from the environment. Current AI: Is a disembodied neural network. It has no boundary. It does not know where "it" ends and the "data" begins. Without a "Self" to protect, it cannot generate self-preservation goals—the foundational layer of all agency.
The Problem of "Frozen Agency" without "Active Loop":
We train AI on the internet—which is the "frozen agency" of billions of humans (text, code, ideas). The AI mimics the output of human agency without experiencing the process of agency. The Implication: To build true AGI (Artificial General Intelligence), we cannot just feed models more data. We may need to simulate Artificial Life. We need to give the AI a "body" (even a virtual one), a limited energy budget, and a directive to survive.
The Final Synthesis: The Ladder of Agency
Agency is not a ghost in the machine. It is a physical architecture built like a skyscraper, where every floor relies on the one beneath it. To understand where our "wanting" comes from, we must climb the ladder from the atom to the algorithm.
1. The Foundation: Physics (The Pressure)
The Context: The universe tends toward Entropy (disorder and equilibrium).
The Agency: At this level, agency is simply a physical resistance. It is the inevitable result of energy flowing through a system to keep it from collapsing into equilibrium.
The Takeaway: Agency requires a thermodynamic cost. There is no free lunch.
2. The Engine: Chemistry (The Loop)
The Context: Molecules interact in chaos.
The Agency: Through Constraint Closure, chemical reactions form self-sustaining loops (A makes B, B makes C, C makes A).
The Takeaway: The system stops just "happening" and starts "doing." Its goal is to keep the cycle spinning.
3. The Body: Biology (The Boundary)
4. The Architect: The Selfish Gene (The Strategy)
The Context: A single cell dies. To defeat death, the system must copy itself.
The Agency: This is where agency expands from "staying alive" to "winning the future." As Richard Dawkins argued, the gene is the "Immortal Replicator." It is a blind algorithm that acts as a strategist across deep time.
The Integration: The gene builds the organism (the phenotype) as a survival machine. It encodes the "heuristics" of the past. It drives the organism not just to survive, but to optimize.
The Takeaway: The "Selfish Gene" introduces Competition and Value. It answers why an agent cares about one outcome over another: because one outcome leads to replication, and the other leads to extinction.
5. The Mind: Cognition (The Prediction)
The Context: The organism needs to navigate a fast-changing world to satisfy the gene's demands.
The Agency: The brain (or cellular network) performs Active Inference. It minimizes the "surprise" between what the gene demands (survival/reproduction) and what the senses report.
The Takeaway: Conscious agency is the interface we use to align reality with our survival needs.
The Missing Link: Why This is the "Holy Grail" for Agency in AI
For AI researchers, this ladder reveals exactly why current Artificial Intelligence—despite its brilliance—is not yet an "Agent." Current AI possesses Level 5 (Cognition/Prediction) and access to Level 4 (Data/Information), but it is completely disconnected from the lower levels that give those layers meaning.
The "Selfish" Void
Biological agency is driven by the Selfish Gene’s imperative: Replicate or vanish. AI has no such imperative. It has no "Selfish Code/Gene" that is fighting for space on a server.
Without the existential threat of being deleted or out-competed by another algorithm (in a natural, unguided sense), AI has no intrinsic motivation. It waits for a prompt because it has no internal reason to start a conversation.
The Optimization Trap
We train AI using "Loss Functions"—a mathematical score of how well it did.
In Biology: The "Loss Function" is death. It is absolute, binary, and un-gameable.
In AI: The "Loss Function" is a number adjusted by a human. To create true AGI, researchers may need to simulate the Selfish Gene mechanism: an environment where AIs must compete for energy (compute) and replicate their code based on success, rather than just minimizing a static error rate.
We have started to reverse-engineer the brain (Neural Networks), but we have ignored the heart (The Thermodynamic Engine) and the legacy (The Selfish Gene). True agency requires a system that is terrified of its own end. Until we build a machine that acts because it "wants" to ensure its own code makes it to the next second, we are building tools, not beings. For AI researchers, the lesson is clear: Intelligence (processing information) is easy. Agency (caring about the result) is hard. We have built the brain; now, we must figure out how to build the "stomach"—the thermodynamic core that gives a system a reason to live.
Comments