> in the world of semantic networks and frames, what are some distinctions
> that people originally failed to make clear, causing a bit of confusion?

most of the problems derived from the lack of a clear and concrete
semantics for semantic nets, leaving people to guess the semantics from
the programs behavior. 

one distinction not made in early systems was that existing between
"subset of" and "member of", using the isa relation to represent both.
another was the distinction between properties of a set as a whole and
properties of the members of the set. also, a distinction mentioned in
rich&knight is that of a link defining a new entity and one relating
existing entities.

as mentioned in the krypton paper, another ambiguity due to the lack of
semantics was whether semantic nets represent linguistic information which
only describes relationships between objects, or factual information that
makes general, exhaustive statements about the world.  furthermore,
because semantic networks allowed manipulation of the data structures
forming them, it was easy for users to confuse assertional information
with structural information, thus leading to unwarrented conclusions based
on node- or link-counting (see the path-length algorithm for inheritance
in rich&knight, for example), or on the presence or absence of these
structural elements. 

these were the ideas i originally had in mind. eyal's answer, it seems,
took a slightly different approach, so i include it here.

eyal
----
People neglected to mention what semantics was actually presented by the
frames/sn. They appealed to a general notion of inheritance, but did not
specify exactly what does each notion means. This in turn, contributed to
the two major ontologies of frames. The first (Hayes-like) said that a frame
is actually a formula that is universally quantified, where the instances
are existential evidences. The second (KL-ONE-like) said that the structure
has nothing to do with the semantics, besides what can be "compiled" into
a frame (i.e., the set of values/restrictions/defaults that it has). There
is no reasoning possible over the structure of the net itself (and thus, we
can not non-monotonically infer that an entity is comprised of its
subentities).


references:

russell&norvig, p.317
rich&knight, p.254,255,271-273
krypton paper, p.3,4

others: mcdermott's 1976 article "artificial intelligence meets natural
stupidity" on the isa ambiguity, and woods' 1975 "what's in a link?" on
the assertional/structural distinction