Research
What is Information?
Information is a slippery concept. Is it a subjective quantity or an objective one? It’s probably clear that there’s information in a book. What about a book written in a language you don’t understand? The book may be informative to someone, even if it is not informative to you, so it still contains information. What if the book is written in a language that no living person can read? Does it still contain information? To describe these situations more precisely one needs to distill information that has meaning from the information that doesn’t. Subjective, meaningful information is semantic. The structural aspects of information, which may or may not have semantics, is referred to as syntactic.
Surprisingly (pun intended, IYKYK), syntactic information is closely related to entropy, an objective aspect of the physical world. Take the microcanonical definition of entropy as the number of microstates in a given macrostate; for example imagine all the ways that the stuff in your bedroom could be positioned. A cheeky way of describing entropy is to say that there are WAAAY more ways for your bedroom to be disorganized than there are ways of it being organized. Where is your favorite shirt? In a disorganized room, you may have to look in many places before you find it, the room is said to have higher entropy. Entropy is synonymous with uncertainty - the more uncertain you are about the exact state of your bedroom, the more entropy it has. When you look around your bedroom you discount areas where the shirt is not, becoming more certain about where the shirt actually is. The information you gain reduces your uncertainty; the entropy of the room is smaller the more you know about it!
Wait up. Hold on. Didn’t that last paragraph start by saying entropy is an objective aspect of the world? Why are we talking about subjective uncertainty? Thermodynamic entropy is related to the temperature of a system. How can looking around for a particular molecule (your favorite one) in a gas, change the temperature of the gas?! The subjective and the objective shake hands to create what seems like a paradox. When you find your favorite shirt in your messy room, you grab it and put it on. Or put it into a dresser drawer. Whatever. The point is that you organize your room a little with the information you gain. Similarly, one can use the information about a particular molecule in a gas to change the temperature of the gas. James Clerk Maxwell first described such a neat fingered being in 1867, and dubbed Maxwell’s Demon several years later by Lord Kelvin. The demon sits at a tiny doorway between two volumes of gas at equal temperature. When it sees a fast moving molecule approaching the door from the right, it lets it through. Similarly, if it sees a slow moving molecule moving from the left it also lets it through. By observing the gas, and acting on that information, the demon can make the gas on the left get hotter, while the gas on the right cools. Why don’t you try it:
Were you hired?
How does Agency emerge from Complexity?
Look at that.
It’s like a little beastie, clamoring about in search of food.
It notices the directions in which its food supply is increasing, and turns towards them.
Sometimes it gets stuck between choices, shakes a little before choosing a particular option to move in.
It looks like its making choices, like its looking for food.
It’s behavior looks complex, much like the behavior of paramecia under the microscope.
But is it?
What you’re seeing is the simulation of a chemoton - a protocell consisting of a small collection of chemical species trapped inside a membrane. The environment contains sources of another chemical species, one that is metabolized by the species in the membrane to produce more membrane. As time goes by the membrane decays, causing the chemoton to shrink. The motion you’re seeing is due to the front end of the chemoton’s membrane growing, and its back shrinking.
Seemingly complex behavior is what our eyes are drawn to, and agency is how our brains rationalize what they see.
That is not what is actually there.
This simple example of an incredibly basic system displaying complex, agent-like behavior is why finding agency is difficult.
We can’t rely on a system behaving in a complicated way to justify the claim that the system is an agent.
For that matter, what exactly do we mean when we say Agency?
We are agents, right?
Entities capable of acting teleologically, of having goals and directing our actions towards attaining them.
By that definition, however, LLMs too are agents.
Whether one agrees or disagrees with that statement, it is clear that agency itself is still a nebulous concet.
Sweeping such concerns under the rug for the moment, when does a system transition from being purely mechanistic to having true agency?
The chemoton may not have agency, but does the paramecium?
Or is it, too, simply a bag of chemicals, reacting and metabolizing its surroundings, all the while fooling us into rationalizing it is something more?
I’m interested in models that live at this awkward boundary between automoton and agent.
Understanding how complexity, computation, and control intertwine to create behavioral autonomy will get us one very large step closer to understanding its emergence in the physical substrate.