Information Technology Reference
In-Depth Information
Given that these aspects of speech regularly affect the import of utterances,
a written sentence must be considered a poor representation of a spoken ut-
terance. But most philosophical and linguistic analyses have proceeded on the
basis of this idealized representation, a tradition that AI has carried on. The
point is not that these fields talk about writing; only that they concentrate on
the aspects of representation that writing normally captures. As a result, the-
ories will naturally tend to lean on distinctions that writing captures, and not
on the many distinctions it doesn't.
Among the many routes by which the writing metaphor entered AI prac-
tice, one moment stands out: Newell and Simon's (1963, 1972) invention
of symbolic programming. Most of Newell and Simon's domains, especially
in their earlier work, have been domains like cryptarithmetic in which the
“world” consists of a sheet of scratch paper. Newell's production system models
do not contain separate mechanisms for the scratch paper and for the agent's
“short-term memory”. Newell and Simon invented symbolic programming in
order to implement the sorts of structures and operations that their models
specified. List structures, like scratch paper, and like the symbolic structures
of all subsequent AI programming languages, have many properties of writing
and few properties of speech. People invented writing because there's nothing
in their heads that's anything like paper.
Writing in the head
AI research is often caught in a pattern whereby mechanisms that seem ex-
tremely “expressive”, “powerful”, and “general” refuse to scale up. Let's return
to the properties that AI has ascribed to symbolic representations - object-like,
passive, static, structured, visible, and portable - and consider how they lead to
difficulties of scaling and implementation.
Symbolic programming languages endow their datastructures with all six
properties of writing. They implement these properties using pointers that
metaphorically make objects visible to processes. Pointers connect the com-
ponents of structures. One transports a structure by “passing” a pointer to
it. Structures only change when processes change them. Pointers do not obey
any locality beyond that of their own connectivity. Thus they are eminently
reconfigurable.
Pointers cause two sorts of difficulties. First, they require their implemen-
tation medium to be infinitely reconfigurable (Chapman 1991: 35-41). They
fight both against the locality of physical space and against the inertia of phys-
ical machinery and its interconnections. On serial machines we observe this
