The wonder and promise of general-purpose computers

As I argue in the intro to this book, the advent of general-purpose computers has altered the limitations of the human condition in ways both unforeseen and still poorly understood.

In the middle of the 20th century, something exceptional in the history of mankind happened: a universal tool was invented: the general-purpose computer. For the first time, any human could potentially become an inventor and solve arbitrarily complex problems using pictures of their thought processes projected as information patterns, ie. programs, without the cost of manufacturing new physical artifacts.

Although Charles Babbage can be acknowledged as the forefather of this artifact, the full scope of its generality was only understood by Turing at the start of the 20th century. Indeed, our current general-purpose computers approximating Turing’s abstract machine are one of only two ways that we currently know to make a computer that can compute most of what a human may want to compute; the other way being queue machines, invented later. Furthermore, when the computer is connected to input/output interfaces, it becomes interactive and can convert to and from phenomena of the physical world. Conversely, there is no known other way to assemble artifacts, other than those that can be described by interactive Turing and queue machines, which can compute most of what humans can think of in a way that can influence, or can be influenced by, physical phenomena.

Obviously one does not need to invoke Alan Turing nor his thoughts when building a device that accomplishes a specific task. Census count machines reading punched cards were built and used successfully long before Turing was born. The reason why Turing’s contribution was remarkable is that it created theoretical confidence that a general-purpose hardware platform could be successfully reused, after it is fabricated, to solve problems not defined yet, and thus guaranteeing perpetual employment for software creators.

This confidence stems from the formal thesis of Alonzo Church and Alan Turing on computability–although it really comes from collective work with Post, Kleene and Gödel, cf. `The undecidable <http://isbndb.com/d/book/the_undecidable.html>`__ by Martin Davis for the full story–which establishes that all expressible formulas of arithmetic, which by definition are all possible computations that humans can phrase intelligibly ever (NB: computations (computable functions) are a subset of all functions that can be phrased by humans. In particular there exist intelligibly phraseable non-computable functions, such as the `busy beaver function <http://www.alcatel-lucent.com/bstj/vol41-1962/articles/bstj41-3-877.pdf>`__), can be computed by either an abstract Turing machine, Church’s λ-calculus or partial recursive mathematical functions. (It further establishes that neither of these formalisms can compute anything more, that is, everything we can compute using either can be described by an intelligible, valid formula of arithmetic; but this point is irrelevant here. See G.E.B. by D. Hofstadter for an accessible review of the theoretical and practical consequences.) Moreover, when the machine is universal, the function it computes can become a run-time input, ie. software, while preserving the full generality of the model. Because of this, a hardware platform that resembles a universal Turing machine gives us confidence that it can be reused in the future by new software to solve problems that have not been formulated yet. Since the only conceptual difference between a universal Turing machine and a concrete implementation is the finiteness of storage capacity (vs. the infinite length of Turing’s tape), it is possible to approximate the abstract machine increasingly accurately by simply adding more addresses to the storage, which seems to be technically tractable for the foreseeable future.

This is the crux of what general-purpose computing is about: design and build a hardware platform now, with reasonable confidence founded in logic that they can be used to solve future problems in software.

Since then, other formalisms distinct from Turing machines and λ-calculus have been developed and subsequently proven to be Turing complete, that is, at least as powerful as Turing machines. The models that have received most attention are queue machines (mentioned above), graph reduction machines able to reduce terms of λ-calculus, register-based variations on the Turing machine, the π-calculus, and specific initial states of cellular automata. Any of these models can be made universal, ie. programmable via software, by modeling a single-function machine whose behavior is to read a program from an input device and then evaluate it. They can furthermore be made interactive, ie. connected to the physical world, by introducing basic actions in their evaluation rules that effect external interactions. However, the only computing artifacts built so far have been either Turing-like (register machines) or queue-like (stack machines). All implementations of other formally as powerful models have only existed as simulations of their semantics in programs running on hardware devices that can be primarily described as register or stack machines. It is the Turing-completeness of these specific two models that guarantees the future utility of the constructed artifacts.