[At some point I noticed that many of my essays contained digressions about the net's effect on our minds and lives. These digressions were faults as digressions, but the topic was interesting, so I began collecting these scraps with the plan of welding them into a coherent essay. The result was on a larger scale than I had expected: around 10 000 words in 13 parts.
This poses a problem of formatting. In the past I have run essays in several parts over as many days. But such a bombardment would be ridiculous. Instead I have gathered these 13 essays into 4 parts, to be run at the usual near-weekly intervals. In this way readers may plausibly have time to digest and follow the argument.]
Is intelligence obsolete? I mean: do the digital technologies of intellectual augmentation make exceptional intelligence obsolete, in the same way that the mechanical technologies of physical augmentation made exceptional strength obsolete? Not, "is the net is making us stupid?" but "does the net make it as impossible to be stupid, as the grid makes it impossible to be powerless?" I want to be free to develop the argument without suspense, so I will conclude now: yes—with reservations about the concept of obsolescence.
I say intellectual augmentation to reference Douglas Engelbart's 1962 Augmenting Human Intellect. I will use this book as the scaffold for the first part of my argument. Anyone who has investigated the origins of the net will know Vannevar Bush's 1945 As We May Think, a prophesy of the Internet in light-table and microfilm. Augmenting Intellect is explicitly an attempt to show how Bush's vision could be made workable in electronic form. It is not a marginal document; six years after it was published the author, head of the Augmentation Research Center at Stanford, gave what is now known as the "Mother of All Demos", where he débuted, among other things, the mouse, email, and hypertext.
Some of the possibilities that Augmenting Human Intellect proposes have been fulfilled; some have been overtaken; some have failed; and some remain untried. The interesting ones are the untried.
The relevant part of Augmenting Human Intellect begins with Engelbart's description of the system he used to write the book--edge-notched cards, coded with the book or person from whom the content was derived. I say "content" because, as anyone who has attempted to maintain a system of notes allowing small, disparate pieces of information to be conserved usefully will realize, it is impossible to strictly distinguish thoughts and facts—the very act of selecting a fact for inclusion implies a thought about it. Engelbart calls these thought-facts kernels. He would arrange these cards into single-subject stacks, or notedecks. In the book he summarizes the frustrations of creating a memo using these cards—the lack of a mechanism for making associations (links, that is, but in both directions), the tedium of copying the links out, the confusion of keeping track of what linked to what. He considers some mechanical system for leaving trails between cards and for copying them, but objects:
It is plain that even if the equipment (artifacts) appeared on the market tomorrow, a good deal of empirical research would be needed to develop a methodology that would capitalize upon the artifact process capabilities. New concepts need to be conceived and tested relative to the way the "thought kernels" could be knitted together into working structures, and relative to the conceptual presentations which become available and the symbol-manipulation processes which provide these presentations.
He proceeds to further object that by the time some such mechanical system could be perfected, electronics would be better suited to the job. And we are off.
But let us pause first to and consider the concept of the kernel. Engelbart is explicit that the kernel itself represents a "structure of symbols" subject to mapping. Yet for purposes of inclusion in a larger symbolic structure the kernel must be treated as smooth and integral. Every symbolic structure is composed of smooth kernels, yet all kernels are spiky. This tension can be dealt with in more than one way.
Imagine a series of levels in the computer's awareness of a kernel's internal structure. (These levels are my own coinage for this essay; they probably correspond to a known mathematical structure, but it seemed easier to reinvent than research.)
Level zero is the simplest possible organization of nuggets, an anonymous jumble where the only thing the system knows about the content of a kernel is that it exists. An unsorted inbox or a directory of temp files are level 0 structures.
Level 1 is a simple filing system: the system knows exactly one thing about the kernel: which folder it belongs to.
A level 2 structure approaches the limits of what is possible with paper: the system knows that a single kernel can be in several places at once. A physical file system where documents are both uniquely identified in some master reference, and where copies of these documents are present in multiple folders, is a level 2 system; so is double-entry bookkeeping, where the kernels have no internal structure at all. Tagging is the computerized equivalent; in library science, faceted classification.
A level 3 structure is possible with paper—using edge-notched cards and pin-sort operations—but in practice it requires a computer. In a level 3 structure the system can retrieve a kernel conditionally, based on which folders it is in—Edward Teach was a pirate and a bearded man, so he is found in the pseudo-folder "Pirates and bearded men". Basically a level 3 structure knows enough about its kernels to do basic set operations—the Venn diagram level.
(Full-text search is a level 3 structure, where each kernel is indexed by every word that it contains.)
In a level 4 structure folders are themselves included in folders. This sounds trivial—manila folders inside hanging folders, or sub-directories in directories—but that would be a level 1 system. In a level 2 structure, each folder is included in multiple folders, including itself. This is impractical on paper, and just practical using symlinks—the POSIX file system is arguably a level 4 structure. But the only nontrivial level 4 structure in operation is Google's search engine, which is smart enough that it could retrieve Edward Teach for "Pirate who did not shave"—it is capable of including the folder "bearded" in the folder "not shaving." (It doesn't, alas).
Understand the distinction between levels 3 and 4: in level 4 the maximum number of searches that can retrieve a kernel is a function of the number of folders and the length of the search query. At level 4 because folders can be retrieved by other folders—not just pseudo-folders—the maximum number of searches assumes at every folder includes every pseudofolder.
But why stop with level 4? Computers could do better. Level 0 jumbles kernels; level 1 puts them in folders; level 2 puts them in multiple folders, level 3 puts folders in folders, level 4 puts folders in multiple folders. The simplest case of a level 5 structure would be the search: "searches that return Edward Teach." This sounds useless until you consider the benefit of gradually narrowing a search by doing one search after another, each on the results of the last. When you think about it this seems very hierarchical—like folders inside folders. With a level 5 structure you could retrieve multiple searches the way you search multiple tags. For example, suppose you want to know how the Queen Anne's Revenge was rigged; and suppose there is a website about this. Now, of course, you could search "Queen Anne's revenge rigging", or search "Queen Anne's Revenge" and then rigging, or "sailing ship" and then "Queen Anne's Revenge"—but no luck. But suppose you could search "searches that return Queen Anne's Revenge + searches that return sailing ship + rigging". Now if it happened that the Queen Anne's Revenge was a French frigate, and French frigates of the early 1700s were rigged in a characteristic way, this search could shake that connection out.
(Note that a strictu sensu blog is actually a sort of level 5 search—a blog that collates links on a certain set of topics is a handmade equivalent to a search on searches that return those topics.)
But enough speculation. The point of proposing these levels is to show that ascending levels in the sophistication of search are independent of the internal structure of a kernel. Even the most sophisticated searches possible, and those not yet possible, are still a matter of folders and contents. And putting a kernel into one or many folders is not the same as parsing it.
Indeed parsing is an impediment to search, not an aid. Certainly it is good when we search on Edward Teach and are directed to a "Blackbeard" chapter in a book about pirates. For our purposes the book as a whole is a kernel; perhaps the chapter is too—we may print it out, or find the book and photocopy it, or collect it screenshot by screenshot. But how far can we break it down? It may be true that half of the chapter is not about Blackbeard at all—this paragraph tells about the town where he was born, this paragraph tells about his ship, this paragraph tells about his competitors—and it may be true that of the paragraphs about him half the sentences are not about him—here is a thought on the nature of brutality, here is a thought about why bearded men are threatening. Yet if you isolate only the sentences that are about Blackbeard specifically, the result is gibberish. You wanted something about Blackbeard? Well, this chapter as a whole is about Blackbeard—but no part of it is actually about him.
This is why PIM is hard: there need not exist any necessary connection between a kernel's internal structure and the folders where it is classified. The relationship is unpredictable. This unpredictability makes PIM hard—/hard/ not as in difficult, but hard as in insoluble, in a way that is revealing of some human limitation. Classification is irreducibly contingent.
Accordingly PIM is always tendentious, always fallible, and not always comprehensible outside of a specific context, or to any other but a specific person. And the most useful abstract classifications are not the best, but the most conventional—like the Dewey Decimal system, whose only advantage was that of existing.
Now I return to Engelbart and his "quick summary of relevant computer technology." It would be tempting to pass over this section of Augmenting Human Intellect as pointless. We know computers; we know what they can do. The introductions necessary in 1962 are needless for us. And true, some of it is funny.
For presenting computer-stored information to the human, techniques have been developed by which a cathode-ray-tube (of which the television picture tube is a familiar example) can be made to present symbols on their screens of quite good brightness, clarity, and with considerable freedom as to the form of the symbol. Under computer control an arbitrary collection of symbols may be arranged on the screen, with considerable freedom as to relative location, size, and brightness.
But we should look, because Augmenting Human Intellect predates a great schism in the design and use of computers. Two sects emerged from that schism. The technologies that Engelbart thought would make augmentation practical largely ended up in the possession of one side of this schism—the losing side.
Engelbart thinks of computers as symbol-manipulating engines. This strikes one in the face when he talks about simulation:
[T]hey discovered that the symbol structures and the process structures required for such simulation became exceedingly complex, and the burden of organizing these was a terrific impediment to their simulation research. They devised a structuring technique for their symbols that is basically simple but from which stem results that are very elegant. Their basic symbol structure is what they call a 'list," a string of substructures that are linked serially in exactly the manner proposed by Bush for the associative trails in his Memex—i.e., each substructure contains the necessary information for locating the next substructure on the list. Here, though, each substructure could also be a list of substructures, and each of these could also, etc. Their standard manner for organizing the data which the computer was to operate upon is thus what they term "list structuring."
This is in reference to IPL-V. A few paragraphs later he writes, with frustrating understatement, "Other languages and techniques for the manipulation of list structures have been described by McCarthy"—followed by eight other names. But McCarthy's is the name to notice; and his language, LISP (LIst Processing) would become the standard tool for this kind of work.
There is a famous essay about the schism, source of the maxim "Worse is Better." It contrasts two styles of programming: the "MIT style"—the style of the MIT AI Lab, with the "New Jersey style"—the Bell Labs style. Software as we know it—based around the C programming language and the Unix family of operating systems, derives from the New Jersey style. Gabriel's essay actually characterizes the New Jersey style as a virus.
But how does this difference in style relate to the concept of "symbolic structures"? Lisp is focused on the manipulation of symbolic structures; and Lisp is the language best suited for this because Lisp code is in fact itself a symbolic structure—every Lisp program is an exhaustive description of itself. C-like languages are a set of instructions to a compiler or interpreter. The instructions are discrete and serial. The resultant symbolic structure exists only when the program is run. (Note that the difference is one of tendency, not of capacity. It is an axiom that any program can be written in any programming language that has the property of being Turing-complete —as all these languages are.)
Why C-like languages won may be summarized by a point of jargon. In Lisp-like languages anything besides manipulating symbolic structures—say, writing a file to disk or rendering it to the screen—is called a side effect. What are side effects to Lisp programmers are the business of C programmers. So instead of symbols and trails we deal with files and windows and websites, and have to hold the structures they are supposed to fit into in our own heads.
Coincidentally in housebuilding the quick and dirty style is called "New Jersey framing." The standard way is to frame a wall as a unit—a grid of studs nailed through their ends at right angles—then stand it up and nail it into place. Jersey framing instead maneuvers each stud into its final position before toenailing it in place—that is, hammering nails in at an angle. The standard style is more secure, but involves delay and forethought; New Jersey framing is less secure, but makes constant progress. New Jersey programming has essentially the same advantages and drawbacks.