The universe of sets, existential quantification in set theory Yesterday, I posted a question that was received in a different way than I intended it. I would like to ask it again by adding some context. 
In ZF one can prove $\not\exists x (\forall y (y\in x)).$ This statement can be read in many ways, such as (1) "there is no set of all sets" (2) "the class of all sets is proper (i.e. is not a set)" etc. and I believe that there is a substantial philosophical difference between (1) and (2). The former suggests that the existential quantifier refers to the actual existence of something intended in a platonic way, while the latter interprets $\exists$ as meaning "it is a set". So, in the second case, I would say that the existential quantifier is a way of singling out things that are sets from things that are not sets, rather than a way to claim actual existence of something. 
I am a set theorist and I always intended the statement above as (2) because I don't think existential quantification in set theory refers to actual existence. I suspect that also Zermelo intended existential quantifications as a way of singling out sets from things that are not sets, because in its original formulation he introduced "urelements" i.e. objects that are not sets but could be elements of a set. But I am interested in what is the most common interpretation among contemporary set theorists and I have the impression that my colleagues in set theory use (1) more often. 
So my question is: from the point of view of someone who believes that existential quantifiers in set theory refer to actual existence, does the statement above mean "the class of all sets does not exist"? Does this interpretation appear anywhere in the literature?   
Thank you in advance. 
 A: I regard ZF (or better ZFC) as a (partial) description of the behavior of actual sets.  The theorem you quoted says, in that context, that there is no set containing everything.  In the same context, I might sometimes talk about classes, but I would regard such talk as an abbreviation for statements that are only about sets, as explained, for example, in Jensen's book "Modelle der Mengenlehre."  In other words, I don't think of classes as actual entities.
Concerning urelements, I would use a slightly modified version of ZF to describe a world of sets and urelements; see for example the theory ZFA in Jech's "Axiom of Choice" book.  The theorem you quoted still holds in the presence of urelements, and it still has the interpretation that there is no set containing everything.  
If some people (not me) wanted to work with sets and proper classes as genuinely existing entities, they would probably use a theory like Morse-Kelley to formalize their ideas.  The theorem you quoted is still available; now it says that there is no class containing everything (including all classes).  
There are, of course, set theories in which the theorem you quoted is not true; Quine's New Foundations and its variants are the most prominent of these. Here there is a set of everything.  Unfortunately, I have no idea what sort of entities NF is "intended" to describe; perhaps Holmes's consistency proof will eventually lead me to such an idea.
A: Even so a model of ZFC might be given together with a collection of classes, only the sets should count as real. So if one would modify the collection of classes without modifying the sets, it would still be the same model. The formula $\forall y (y\in X)$ defines a class $X$ which cannot be modified at will, but a formula using impredicative quantification over classes will not define such a fixed class.
I feel this is similar to how a manifold in differential geometry can be embedded into a simpler but larger space. The simplest way to describe a manifold like the $n$-sphere might be its obvious embedding into $\mathbb R^{n+1}$, but the embedding is still arbitrary and unimportant. The simplest way to describe a manifold (like a projective space) might also be as equivalence classes of a simpler but larger space. This sort of external description is just as arbitrary and unimportant as the external description provided by an embedding. The models of ZFC provided by the model existence theorem are given as equivalence classes of simple syntactic terms, so this analogy still works.
A: Considering what you wrote in your slide presentation "On the definitional character of axioms.", you might be interested in the following preprint by John L. Bell (found on his Homepage) titled "SETS AND CLASSES AS MANY".  In it, he initially gives the following naive definition of set (following Cantor in his book, Contributions to the Founding of the Theory of Transfinite Numbers):

Set theory is sometimes formulated by starting with two sorts of entities called individuals and classes, and then defining a set  to be a class as one, that is, a class which is at the same time an individual...If on the other hand we insist--as we shall here--that classes are to be taken in the sense of multitudes, pluralities, of classes as many, then no class can be an individual and so, in particular, the concept of set will need to be redefined.

What does "class which is at the same time an individual" mean?  Well, following Cantor,  since a set ("aggregate") is

...any collection into a whole $M$ of definite and separate objects m  of our intuition and our thought.  These objects are called the "elements" of $M$...In signs we express this thus:  $M$={m}.

it seems reasonable to infer that a set is then a class which is a object, i.e. that which can be an element of another set (or even, possibly, of itself).  This, however, produces an interesting variant of the Russell paradox, defined entirely in terms of class as object:
Is {x| x$\notin$x} both a class and object, that is, can {x| x$\notin$x} be an element of another set (including itself)?  If {x| $\notin$x} is both a class and an object, a contradiction follows because then {x| x$\notin$x} can be an element of itself.  On the other hand, since {x| x$\notin$x} is, in some sense a 'name' (label) of its elements, this 'name' can be an 'element' of some other collection, it must be deemed an 'object' and therefore can again be deemed an element of itself, and the contradiction again follows (perhaps the paradox ensues through a confusion between 'label' and that which is labelled, but then, can a class that is not itself an 'object' be labelled?).
Bell solves this problem in the following manner (though he does not explicitly mention the above version of the paradox):

Now while we shall require a set to be a class of  some kind, construing the class concept as "class as many" entails that sets can no longer literally be taken as individuals.  So instead we shall take sets to be classes that are are represented, or labelled, by individuals in an appropriate way.  For simplicity we shall suppose that labels are attached, not just to sets, but to all classes:  thus each class $X$ will be assigned an individual $\lambda$$X$ called its label.  Now in view of Cantor's theorem that the number of classes of individuals exceeds the number of individuals, it is not possible for different classes always to be assigned distinct labels [consider now the argument put forth in Stanford Encyclopedia of Philosophy's  (SEP's) entry  "Frege's Theorem and Foundations for Arithmetic" (Edward Zalta's entry) that Russell's paradox is engendered because Second-order Logic+ Basic Law V requires the impossible situation in which the domain of concepts (labels) has to be strictly larger than  the domain of extensions (classes) while at the same time the domain of extensions has to be as large as the domain of concepts--my comment].  This being the case, we single out a subdomain $S$ of the domain of classes on which the labelling map $\lambda$ is one-to-one.  The classes falling under $S$ will be identified as sets; and an individual which is the label of a set will be called an identifier.

Bell now defines the dual notion of colabelling:

For reasons of symmetry, it will be convenient (although not strictly necessary) to assume that, in addition to the operation of labelling each class by an individual, there is a reverse process--colabelling--which assigns a class [note here that classes then, of necessity, must exist as extensions--my comment] to each individual.  Thus we shall suppose that to each individual x there corresponds a unique class x* called its colabel.  Again, because of Cantor's theorem, not every class can be the colabel of an individual (although every individual can be the label of a class).  However, it seems natural enough to stipulate that each set be the colabel of some individual, and indeed that this individual may be taken to be the label of the set in question.  Thus we shall require that $X$=$\lambda$$($$X$$)^{*}$ for every set $X$.  In that event, for any identifier x in the above sense, we shall have x = $\lambda$(x*); that is, the colabel of an identifier is the set of which it is the label, or the set labelled by the identifier.  Another way of putting this is to say that the restriction of the colabelling map to identifiers acts as an inverse to the restriction of he labelling map to sets.

After dealing with singletons and the empty set in the following fashion

Singletons and the empty class--"multitudes" with just one, or no members respectively--are here regarded, like the "numbers" $\mathbf 1$ and $\mathbf 0$, as "ideal" entities introduced to enable the theory to be developed smoothly.

he considers the problem of adequately defining the $\in$ relation:

The membership relation $\in$ between individuals and classes is a primitive of our system.  It will be taken as an objective relation in the sense suggested, for example, by the assertion that Lazare Carnot was a member of the Committee of Public Safety, or Polaris is a member of the constellation Ursa Minor.  The fact that $\in$ is not iterable--there are no "$\in$-chains"--means that it can have very few intrinsic properties.  This is to be contrasted with the relation $\epsilon$ of "membership" between individuals, defined by x $\epsilon$ y $\leftrightarrow$ x $\in$ y*:  x is a member of the class labelled by y.  This relation links the entities of the same sort and is, accordingly, iterable.  It should be noted, however, that the presence of the colabelling map * in the definition of $\epsilon$ gives the latter a purely formal, arbitrary character

Next, Bell uses the $\epsilon$-relation to present the notion of nonwellfounded set.  However,

In the usual set theories it is difficult to grasp the nature of a set which is, for example, identical with its own singleton since a set cannot be "formed" by assembling individuals.  In the present scheme, on the other hand, the assertion $\alpha$={$\alpha$}--which is, as remarked above, not well-formed--is replaced by the assertion $\forall$$x$($x$$\epsilon$$\alpha$ $\leftrightarrow$ $x$=$\alpha$), that is, $\alpha^*$={$\alpha$}, which asserts that {$\alpha$} is identical, not with $\alpha$ itself, but rather with its colabel.  Similarly, the self-membership assertion $\alpha$$\in$$\alpha$ is transformed into the statement $\alpha$ $\epsilon$ $\alpha$, that is, $\alpha$$\in$$\alpha^*$, which asserts that $\alpha$ belongs, not to itself, but merely to its colabel.  And an assertion of cyclic membership $\alpha$$\in$$\mathop b$$\in$$\alpha$ is transformed into the assertion $\alpha$ $\epsilon$ $\mathop b$ $\in$ $\alpha$, or $\alpha$$\in$$\mathop b^*$& $\mathop b$$\in$$\alpha^*$, that is, "$\alpha$ (respectively $\mathop b$) is a member of the colabel of $\mathop b$(respectively $\alpha$)."  These rephrasings appear much more natural in that they only impute the possession of curious properties to the colabelling map, rather than to the objective membership relation $\in$ itself.

Considering the "naturalness" of the rephrasings, and the version of the Russell paradox stated above, one might reasonably infer that the confusion between labels and colabels is at the heart of the derivation of Russell's paradox, and by making the distinction between the two, Bell has found the 'best' way of ridding systems of set theory from it.
I say this because of what Prof. Bell says in the concluding paragraph of the "Introduction" to his paper:

We shall also see that, in addition to nonwellfounded set theories, a number of other theories familiar from the literature can be provided with natural formulations within the system to be presented here [the theory $\mathbf M$ of multitudes or classes as many--my comment from the begining sentence of Bell's Section 1].  These include second-order arithmetic, the set theories of Zermelo-Frankel, Morse-Kelly, and Ackermann, as well as a system in which Frege's construction of the natural numbers can be carried out.  Each of these theories can therefore be seen as the result of imposing a particular condition on a common apparatus of labelling classes by individuals [these conditions therefore define what sets 'exist' within each individual theory, a result similar to yours --my comment].

