rationalskepticism.org

GrahamH wrote:

John Platko wrote:
2) Agential states are knowledge

3) Therefore agential states must be able to survive noise, to some extent, and therefore be error correctable.

I'm not disputing that mental states (agential states) are multiply realisable) but I think your logic here is a bit of.

Although we might accept that knowledge tends to have these characteristics, I don't think they are necessary. I can know something for a moment then forget it (fail to retain it for any of various reasons).

I don't see why we must discount it as knowledge merely because it can be forgotten.

3.1. An accurate constructor must contain a replicator

A task T being possible means that for any given accuracy (short of perfection) the laws of physics permit an approximate constructor capable of performing the task to that accuracy.

Consider a possible, non-elementary task T and an object F that can perform T to a high accuracy ε.9 For instance, T could be the task of constructing a car from generic substrates and F a generalized car factory, including all the processes converting raw materials, such as iron, etc., into a car.

The approximate constructor F executes a procedure—a recipe—to perform the task T to accuracy ε. I will show that F must include a replicator and a programmable constructor; and that the recipe must have a hierarchical structure and be instantiated in the replicator.

No-design laws contain no good constructor for T, such as F—neither in the elementary interactions, nor in the generic resources. Hence, the recipe used by F to perform T must be decomposable into steps (not necessarily sequential) that are allowed by no-design laws, i.e. sub-recipes—procedures to perform sub-tasks that are executed by sub-constructors contained in F. To avoid infinite regress, two conditions must be fulfilled.

One is that the sub-tasks be non-specific to T. For instance, when T is the task of constructing a car, the sub-tasks are those of constructing sub-parts of the car—e.g. door handles, windows, etc. Hence, the constructor F must include two parts: one—called V—performs T blindly, i.e. subtask by subtask, and it is non-specific to T, because so are the sub-tasks. The rest of F—called P—is specific to T and instantiates the recipe for T—the sequence of sub-tasks controlling V. Hence, F is a programmable constructor, V, programmed with a program P having the same logic as the recipe: it has a modular structure P = (p1, p2, … , pN) where each instruction pi takes values in an information variable and tells V which sub-task to perform, when, on the substrates10. V is non-specific to T because it must also be capable of executing other programs—different combinations of the elementary units pi. For example, a car factory contains robots executing sub-recipes to construct the car's doors. These robots contain sub-robots to construct handles, windows, etc., usable to construct other objects than cars.

The other condition is obtained by applying the same reasoning recursively to the sub-tasks. If they, too, are non-elementary, they require a recipe that is decomposable into non-specific sub-recipes. The base for the recursion—for T to be performable to that particular accuracy—is provided by the elementary sub-recipes of the recipe for T being elementary tasks—which can be performed by (approximations to) constructors that are available in nature, as generic resources.

These elementary sub-tasks need not be specified in the recipe: they are implicit in the laws of physics. For instance, the elementary steps in the car recipe are tasks like, say, ‘oxidize the aluminium coating’, occurring simply by leaving the substrate exposed to air.

Under no-design laws, any (approximation to a) constructor wears out after a finite time. Therefore F, to perform the task T to the accuracy ε, must undergo a process of maintenance, defined as one whereby a new instance of F—i.e. of P and V—is constructed, from generic materials, before the former one stops working. In the car factory, this is achieved by replacing old sub-parts of the robots, assembly lines, etc., and by preserving the programs they run.

To avoid an infinite regress, the maintenance must not in turn require the recipe P for T. Also, the recipe's design cannot be in the laws of physics. Thus, the only other possibility is that the new instance of the recipe is brought about by blind replication of the recipe contained in the former instance—i.e. by replicating its subunits pi (non-specific to T). We conclude that, under no-design laws, the recipe must necessarily be instantiated in a modular replicator: a physical object that can be copied blindly, an elementary subunit at a time. By contrast, V—the non-specific component of F—is constructed anew from generic resources.

Moreover, under no-design laws errors can occur: thus, to achieve high and improvable accuracy, the recipe must include error-correction. In the car factory, this includes, say, controlling the functionalities of the subcomponents (e.g. fine checks on the position of doors, wheels, etc.). Hence, the recipe P must contain information about the task T, informing the criterion for error detection and correction.

The information in the recipe is an abstract constructor that I shall call knowledge (without a knowing subject [29]). Knowledge has an exact characterization in constructor theory: it is information that can act as a constructor and cause itself to remain instantiated in physical substrates. Crucially, error-correcting the replication is necessary. Hence, the subunits pi must assume values in a discrete (digital) information variable: one whose attributes are separated by non-allowed attributes. For, if all values in a continuum were allowed, error-correction would be logically impossible.

We can conceieve of many cases of 'knowledge' that can be destroyed by noise, that lacks error correction and so on.

Do 4 does not follow.

John Platko wrote:4) For 3 to be possible there must be more than one physical set of states that map to a given agential state so that if an error is introduced by noise in the set of physical states that define an agential state, the state can still be recovered. There are many ways agential states can be encoded to provide for error correction but suffice it to say that if there is only one physical state mapping to an agential state, then if that physical state mapping changes in any way, then that agential state is lost. Such an information mechanism does not meet the definition of Knowledge because it can't meet the survivability criteria.

If you are merely claiming that mental states are not precise arrangements of quantum particles such that displacement of a single quark alters the state substantially then of course that is trivially true but I don't accept your deduction from CT.