forwarded message from drmacro@vnet.IBM.COM (Marc Andreessen)
Date: Wed, 5 May 93 15:05:02 -0500
From: (Marc Andreessen)
Message-id: <>
Subject: forwarded message from drmacro@vnet.IBM.COM
X-Md4-Signature: 752065daba17e0609d9c89c0a57805d5
Forwarded from comp.text.sgml... 

------- Start of forwarded message -------
Newsgroups: comp.text.sgml
Disclaimer: This posting represents the poster's views, not those of IBM
From: drmacro@vnet.IBM.COM
Subject: Explicit Linking is Impossible
Date: Tue, 4 May 93 11:10:35 EDT

At relatively small scales, it is possible to create hypertext
where authors hand craft the links between various places, say
via typical IDREF mechanisms.  Given a "standard" document of
100 pages, one author can reasonably manage to create say 1000
links within that 100 pages.  We do it as a matter of course
within IBM.  Across a library of 10 such books, we can manage
crafting 100 links from each book to the other books, but just
barely.  The cost is the cost not of initially creating the
links, but of maintaining them as the documents change during
development.  This can just be managed within one library where
all the books are finished and delivered at about the same time.

The next order of magnitude up would be linking across libraries.
This is impossible once the number of cross-library links goes
above about 50 because of maintenance problems--there are simply
not enough hours in the day or people on the job to create, maintain,
and test these links.  The next order of magnitude up (cross-enterprise
linking) is so clearly impossible as to be not worth considering.
               ^^^^^^^^^^^^^^^^^^ [emphasis added -marc]

Thus my contention that explicit linking is impossible at large
scales.  Even if you apply tools at one order of magnitude to
make the impossible possible (say a database system to track
links and manage changes), the next order of magnitude increase
will be impossible.  With change management and tracking, the
human element cannot be eliminated and you quickly reach the
practical limit at which humans can manage change.

Note that within IBM, and specifically within the Networking
Systems line of business, we are now at the cross-library linking
stage.  Using the BookManager product, we are creating, by hand,
explicit hyperlinks between the NetView, VTAM, NCP, and related
product libraries.  We think we can manage a relatively small
number of links at this level, but there's no way we could do
it at any larger scale.  We need to move up to the next order
of magnitude (cross-enterprise linking), but we can't with the
explicit-link-based tools we have today.

Note that when I say explicit or hand-crafted links, I mean
only from the author's perspective.  Exoterica's experience
with the Cinemania project is instructive here.  If I understand
it correctly, they used declarative markup in the SGML source
to enable creation of all links associatively, and then bound
those links explicitly in the derived form that was actually
delivered.  From the author's viewpoint, all the links were
implicit, but from the final application's viewpoint, they
were all explicit.

The cost of hypertext is the cost of creating and maintaining
the links.  I don't have any numbers, but it's pretty clear
>from my experience with the extensive hypertext we've created
in Networking Systems, that the cost of maintenance is not
a linear function of the number of links, but increases dramatically
(geometrically?) with the number of links, compounded by the
degree to which the linked objects' schedules are not synchronized
and the life span of the linked objects.

Eliot Kimber                      Internet:
Dept E14/B500                     IBMMAIL:   USIB2DK9@IBMMAIL
Network Programs Information Development     Phone: 1-919-254-5160
IBM Corporation
Research Triangle Park, NC 27709
------- End of forwarded message -------

You know, if I didn't know any better, I'd say he probably works for
IBM.  (Well?  Didn't *you* lose $5 billion last year?)


Marc Andreessen
Software Development Group
National Center for Supercomputing Applications