VSN-C Considered Harmful

Joel Gustafson / Thoughts

Winston’s Contributions

Professor Winston’s AI seminar serves two parallel purposes: a survey of seminal papers in artificial intelligence, and a vehicle for what Winston calls the “VSN-C mantra”. This mantra, Vision-Steps-News and Contributions, is at once a way of writing, a way of reading, and a way of life. Winston argues that good writing, particularly in a research context: 1) first lays out the author's Vision, then 2) lists the Steps required to realize that vision, then 3) updates the reader with some News of original content, and 4) concludes by summarizing the author's Contributions to the field. This is marketed as a quick-and-dirty form for all of a student's communication needs, and is emphasized so as to border on dogma, with weekly exercises in writing various pieces in VSN-C and extracting VSN-C from insufficiently compliant works.

I propose that this is harmful.

Goodhart’s Law

The SAT was temporarily a good test. Its purpose was to evaluate “college readiness”, which it did well: even though the content of the test (e.g. obscure vocabulary knowledge) didn’t directly represent college readiness, the two were so closely correlated that they became interchangeable for evaluation. College readiness was testable by proxy via trivia, because trivia wasn’t known except by those ready for college.

But soon, high school students bought practice books and memorized trivia directly, without the trouble of readying for college beforehand. Those convenient proxies became gameable, so the thing that the SAT really tested stopped being “college readiness” and became “proficiency at taking SAT tests”. This phenomenon is so common (and so commonly discussed) that it has least three names (Goodhart’s Law, Campbell’s Law, and the Cobra Effect). If names give you power over things, we certainly should have tamed this one by now! Yet it lives on, in high school exams, teacher evaluations, or any metric susceptible to feedback loops.

Academic publication is another such metric. Publication, after all, should not be the goal of any researcher. Publication is the evaluation of research done well. Research is done for advancement, for discovery, for exploration, for science! But if research is ever done to be published, then publications become journals not of the best research, but only of the most publishable. “When a measure becomes a target,” Charles Goodhart wrote, “it ceases to be a good measure.”

This is one insidious danger of VSN-C. Just as the SAT discreetly substituted college readiness for vocabulary memorization, emphasizing frameworks like VSN-C substitute an idea for its presentation. These may be correlated, but are not at all the same, thereby rendering writers vulnerable to feedback loops in pursuit of VSN-C conformance. Many good papers may happen to follow VSN-C, and VSN-C may make some writing more compelling or memorable or publishable, but identifying those as primary goals to write for is dangerous.

Another side effect of the SAT’s vocabulary obsession is that high schoolers feel pressured to use large words in essays that they might not feel comfortable with. Similarly, an obsession with VSN-C pressures young academics to hallucinate visions and pull contributions from out of the void to justify what might already have been a perfectly self-contained idea. Applying VSN-C too liberally inflates papers and distracts from whatever core they originally wanted to communicate.

Intentional or not, VSN-C implies priorities that value being read and easily summarized over the authentic expression of an original idea. Even worse, these priorities propagate backwards to prime students to seek ideas that can easily fit the VSN-C form, severely restricting the concept-space they might otherwise explore.

McLuhan’s Message

In Computer Science, code is strictly separated from the data that passes through it. Structure is distinct from content. Information can be cast from type to type, transmuted from form to form, serialized and deserialized - always perfectly recoverable. Its representation seems independent of its essence.

The real world is not like this. The Mona Lisa cannot be reconstructed from a poem, no matter how vividly written. No infographic can recreate Ozymandias, and no single picture can capture the kinesthetics of a ballet. Any message is inextricably intertwined with its medium.

But it is tempting for computer scientists to deny this. It is particularly tempting with pieces not normally considered artistic, like research papers or grant applications or PhD theses. "Have an idea?", Winston asks, "Just serialize it into VSN-C! Then anyone can read it!". But that's not true! Like a haiku retelling of Swan Lake, it may be good, but it’s no longer the same.

Translating a message across mediums preserves some of its aspects, but irreversibly mutates others. Worse, the new medium projects its own connotations onto the message. By writing “in” VSN-C, your ideas get overshadowed by the narrative that VSN-C imposes. This may be desirable for some, but putting the medium before the message doesn’t give your ideas a chance: they’re immediately branded with VSN-C, associated with all of the other VSN-C papers the audience has read, and subconsciously aligned with previous work. For incremental contributions, this could be an advantage, as it lets readers quickly parse the paper and place it in existing context, but for truly novel works, this positioning detracts from its apparent significance. VSN-C as a medium reeks of bite-sized second-semester MEng theses written just before deadlines.

Instead, effectively communicating an idea means guarding every shred of originality and designing the method of communication around the message itself. Human perception is wonderfully sensitive, but with it comes great responsibility to share our ideas carefully, with clarity and authenticity.

Minsky’s Medium

Marvin Minsky’s perspective on the architecture of the mind is not easily stated. It is abstract and difficult to reify. It is not one single idea, but many pieces that fit together to form a cohesive view that only makes sense holistically. It is not supported by experimental evidence, and requires the supposition of completely new objects like agents and k-lines.

Minsky could have written Society of Mind serially, like any other book. But Minsky knew that he was not trying to write about a serial idea - instead, he was writing about a collection of little ideas that create a whole. His message was fundamentally one of synthesis, so rather than twist it into a wall of text, he wrote it as he imagined it: a scattering of bite-sized atoms, with flexible dependencies instead of rigid order, and explicit relations between each other. Minsky created a new, novel medium around his new, novel message. A VSN-C treatment of Society may have checked all of Winston’s technical boxes, but would have nowhere near its intrigue or allure.

Gustafson’s Objection

Professor Winston has been teaching at MIT for longer than I have been alive, and reading and writing papers for much longer. Certainly he, more than anyone else, can be expected to give safe, reliable advice on how to produce a reasonable paper that will pass thesis defence or get accepted to a journal. VSN-C is an effective formula. This I do not doubt.

But I object to the premise of VSN-C. I object to the idea of writing with a purpose to be read or remembered or published. Writing, and particularly academic writing, is not persuasion: it is communication, and communicating well means framing your writing around your ideas, wherever that may lead.

Even more serious than my ethical objection is the harm VSN-C poses to students, whose thoughts are more creative and minds more expressive than they ever will be again. VSN-C has a chilling effect on this creativity, through academic feedback loops and backpropagation of VSN-C’s formulaic style. While frameworks can certainly be useful, great care must be taken to limit their influences and qualify their application.

It strikes me as the strangest of ironies that a professor pursuing human story-understanding would propose a writing style that seems designed for a machine to parse. We want computers to become more human - not vice versa! If creativity is our greatest power, then it is the wild and chaotic and expressive that we must value, not the mechanical and robotic.