I’ve been aware of the Dvorak Simplified Keyboard for a long time, but only in the past few days have I decided to try the layout for myself. Like any cognitive realignment pushing against the momentum of a lifelong habit, the initial adjustment process has been slow and occasionally punishing. When you are acccustomed to the fluidity of the keyboard as an invisible extension of the mind, it’s terrifying to find it amputated and clumsily reattached. I expect this overwhelming self-consciousness to be the norm someday when future generations willingly trade in their limbs for more dynamic cyborg substitutes.

Up to now, the closest I’ve come to this awkward stumbling was when I attempted to train my left-hand dexterity on Charlie Parker melodies I would normally play with my right. A kind of impotence, really: I was willing myself to do things that I was used to executing at dizzying velocities with ease, but my body just wouldn’t respond. The trick, I discovered, is to force yourself to slow down, clean up the suddenly naked particulars, and not rely so much on your established ‘chunks’ of muscle memory. My left hand is still a shambles, mind you, but as the lesser automaton it invents the more colourful passages.

That’s why I’m still plugging away in Dvorak. It may be slow-going at first—this post you are reading now is taking an eternity to punch in—but within minutes of playing with it, you begin to perceive all sorts of qualitative pleasures that simply don’t exist in QWERTY-land. It’s like switching to an Apple Macintosh, complete with the moment of epiphany where the cultishness of the already indoctrinated looks reasonable all of a sudden. (Or so I’ve heard. Having been a Mac user on and off since the age of five, I can’t really say.)

Patented in 1936, the Dvorak keyboard was designed around a handful of basic principles. High-frequency characters reside on the home row (middle row) to minimize squashing and stretching. Vowels and common punctuation marks sit together on the left, encouraging the alternation of hands from one character to the next; one hand presses keys while the other hand repositions. Finally, synergistic pairs like the digraphs ch and th are packed in close proximity. (In the original design, the arrangement of the number row fell on the axis of an outward spiral, reading 7531902468 from left to right. Even Dvorak’s adherents conceded that was silly, and it has largely been dropped.)

Dvorak users will often tell you two things. The first is that the layout’s ergonomics are a vast improvement on QWERTY, allowing you to push your typing to record speeds without incurring nearly the same risk of repetitive strain injuries. I can’t verify this myself; as someone who pulls a respectable 120 wpm in QWERTY, it’s unlikely that I’ll see efficiency gains in my typing habits anytime soon, and RSI has never been a problem for me thanks to my exclusive preference for lightweight, shallow keyboards.

Intuitively, the claims about Dvorak’s top-speed advantage sound plausible. Although the credibility of the original pro-Dvorak study has been questioned, notably in the 1990 paper “The Fable of the Keys” by SJ Liebowitz and Stephen E Margolis, the fact remains that the QWERTY layout was specifically “anti-engineered” by its inventor, Christopher Sholes, to split digraphs and spread common letters apart and thus avert key-jamming. In other words, it was designed to slow you down.

The second thing you’ll hear is that the Dvorak keyboard has nothing to do with the most notable figure to bear that name, the great romantic composer Antonín Dvořák. The keyboard’s designer, the Seattle-based educational psychologist August Dvorak, was a distant cousin at most—and that, we’re told, is all there is to the story.

This is where I disagree.

Key Largo

Most of the conversation you will find about the Dvorak layout portrays it as a case study in economics. If mass commercial standardization precludes the adoption of a considerable improvement in design, the argument goes, do free markets really foster innovation? Jared Diamond’s 1997 essay in Discover, “The Curse of QWERTY”, is a classic of the genre. Liebowitz and Margolis, in contrast, stay on the tack that QWERTY has remained triumphant simply because the alternatives aren’t discernibly better.

In either case, the way people tend to talk about Dvorak is invariably utilitarian, balancing the costs and benefits of adoption in the quantifiable parlance of character frequencies, finger mileages, relative activity by row, and above all, words per minute. Rarely will you hear specifics about the intangibles of the overall Dvorak experience, even among the few who swear by it. My impression is that many who praise Dvorak on principle don’t actively use it themselves—”I wish I knew how to qwert you!” rings the cri de cœur on Backspace Mountain—and testimonials among those who do typically say a few words about speed and comfort and leave it at that.

The closest thing I’ve seen to a lucid experiential observation is this article by Nicholas Thompson, in which he says:

Using a Dvorak after a lifetime of banging on a Qwerty is like removing a tiny pebble from your shoe. Writing a word such as “the” gives me a buzz as I roll my fingers to the left in a fluid, natural motion. The the the the.

The the the the. Thompson couldn’t have picked a better example; ‘the’ is the word that sold me on Dvorak. It rolls off your fingers like the spoken word rolls off your tongue as you flick it against the back of your teeth. Teeth teeth teeth teeth. But then he blunders:

For musicians, think about trying to play “Blowing in the Wind” starting with a B-flat ninth. That’s a Qwerty board. Now think about starting on a G chord. That’s a Dvorak board.

This makes no sense to me whatsoever. Were you to play “Blowing in the Wind” with a ninth on the initial tonic, the chord would reduce to a F6 or Dm7 over a B-flat root. Easy, comfortable, and far better than G on many instruments. I suspect Mr Thompson was a guitarist.

However poorly he may have worded it, Thompson had the right idea. Dvorak’s layout is more than a mere ergonomic reconfiguration. It proposes an entirely different way of thinking about typing. It makes the activity of typing musical. Dvorak, in a word, is like Dvořák.

Major Major Major Major

Experienced pianists have a way of detecting whether a composer is catering to their needs. In this respect, Frédéric Chopin comes up most often as the model composer for the instrument. Playing Chopin is like revenge: it isn’t easy by any means, but everybody covets the satisfaction of pulling it off. It’s easy to see why once you practice his works—his chords and patterns have an uncanny knack for fitting in the curve of your hand like a volley of fastballs perfectly aimed at the palm of a catcher’s mitt. The fingerings by and large suggest themselves. In the jazz world, Duke Ellington is the same way: construct the chords under a tune like “Mood Indigo” or “Prelude to a Kiss” and you find yourself pulled towards brilliant, open structures with voice-leading that magically clicks.

One neat little morsel of trivia about Chopin is that he liked to start new piano students on B major, which has the most idiosyncratic fingerings of any major scale if you learn the instrument according to the common pattern of starting with C major (no sharps or flats, and therefore no black keys) and adding accidentals as you get better, spreading out along the circle of fifths. In the usual progression, B major with its five sharps is introduced relatively late, and thus it has developed a reputation for being difficult.

Chopin, who frankly knew better than everyone, taught C major last. It’s the easiest key to read, he reasoned, but the hardest key to play. C-oriented thinking creates obstacles in the long run in real-world performance scenarios; better instead to begin with B, which develops the proper contour of the hand. This won’t seem like a big deal if you are anything like Eva van Crommelynck from David Mitchell’s novels and “couldn’t tell C major from a sergeant major,” but believe me, it is.

For Chopin, training for the eventual practicalities of expressing real ideas took priority over taking advantage of conventions that happened to be convenient now. You can probably see where I’m going with this. The Dvorak keyboard, you may notice, was conceived along similar pedagogic lines. It is a system where to work on fundamentals is to prepare yourself to tackle practical scenarios efficiently. Learn a few neighbouring characters at a time, starting with your hands in the rest position, and within minutes you already have the building blocks of words and phrases.

But it’s the clever arrangement of digraphs where Dvorak truly shines. This is something you pick up right away: drum the right hand on its natural resting place and you instantly glimpse the potential of legato typing. The t in nth is a passing tone; the s in sh, a colourful appoggiatura.

Even as a Dvorak novice, you don’t hunt and peck a character at a time. Instead, as you practice the layout you rapidly come to visualize phonemes and syllables, hammering them out in clumps. Strings like ‘Schubert’ feel like five keystrokes, not eight, but acronyms like QWERTY remain a nasty pickle. Pronounceables skip like stones on a pond; abbreviations are minefields daring you to tiptoe across. In essence, the rhythm of Dvorak imitates the rhythm of speech. Rhythm rhythm rhythm rhythm.

QWERTY has a rhythm of its own once you’re fluent, but as you accelerate you converge on the uniform staccato of a Gilbert and Sullivan patter song. There isn’t a way around this, either, as your pace is bounded by your fingers’ travel time. Typing in QWERTY is atomic at heart, decomposing into a succession of meaningless independent characters—quite unlike Dvorak, where vowels and consonants are demarcated by their very placement, and the phoneme reigns supreme.

We could almost think of the QWERTY map’s decentredness as twelve-tone serialism over a wider alphabet of possible notes, none of them privileged, no combination outwardly consonant or dissonant. By this analogy, switching to Dvorak is akin to witnessing music history play out in reverse, returning to a classical pianistic scheme of vowels in the left hand harmonizing a punctuated melody of consonants in the right.

And in tactile terms, that’s really how typing in Dvorak feels, only all the letter-chords are broken down sequentially. There are many obscure alternatives to QWERTY in the keyboard ecosystem, but perhaps why Dvorak has endured as the representative champion is its essential musicality. It’s the romantic keyboard, reminding us that beneath every typewritten palimpsest sleeps a sound.

It reveals an odd kind of poetry, too, when you first practice it in fridge-magnetic increments. “The idea that nineteen studious Dadaists assisted Einstein is asinine,” reads the cryptic aphorism of a home-row exercise. “This session is tedious on the tendons,” reads another.

A few weeks ago, Colby Cosh—a friend of a friend of sorts who ordinarily writes reasonable things for a chap who still thinks the Edmonton Oilers are a real sports team—penned an article in his Maclean’s blog about Watson, IBM’s Jeopardy!-playing machine (“I’ll take ‘Cheap Publicity Stunts’ for $1000, Alex”, 16 January 2011), that I found to be dreadfully uninformed. The thrust of his argument is that Watson is a corporate “gimmick”—a fancy plea for media coverage by the faceless villains at IBM, with nothing of scientific interest going on underneath. Keep in mind that by the standards of this article, nothing in the “perpetually disappointing history of AI” will ever be interesting until we’ve graduated from tightly delimited objectives to Big Problems like the Turing Test:

Every article about Watson, IBM’s Jeopardy!-playing device, should really lead off with the sentence “It’s the year 2011, for God’s sake.” In the wondrous science-fiction future we occupy, even human brains have instant broadband access to a staggeringly comprehensive library of general knowledge. But the horrible natural-language skills of a computer, even one with an essentially unlimited store of facts, still compromise its function to the point of near-parity in a trivia competition against unassisted humans.

This isn’t far off from saying that particle physics will be perpetually disappointing until we’ve observed the Higgs boson, or that manned spaceflight is merely an expensive publicity stunt that will never be scientifically interesting until we’ve colonized the Moon: it leans heavily on popular culture as the ultimate barometer of scientific achievement, and it requires both ignorance of methodology and apathy towards specifics.

Colby and I had a five-minute skirmish about the article on Twitter, which as a format for debate is unwieldy as piss. I promised a proper response as soon as I cleared some other priorities off my plate. Those other priorities are still, to my annoyance, on my plate; but having finally paid good money to register my copy of MarsEdit, I’m thirsting for a scrap.

This topic will do as well as any. Reluctant as I am to swing the pretentious hammer of “I know what I’m talking about,” this really is (as the idiom goes) a chance for Faramir, Captain of Gondor, to show his quality. Computational linguistics happens to be my onetime research area, popular misunderstanding of science happens to be one of my favourite bugbears, and Kasparov’s anticomputer strategies against Deep Blue happened to make a cameo appearance in the meandering slop of my master’s dissertation. None of this matters a great deal, mind you. One should never be dismissive of journalists from a position of relative expertise; they’re the ones people actually read, and it’s vital to engage with what they say.

(It is a little game we play: they put it on the bill, I tear up the bill.)

When simplifications attack

What concerns me is not so much Colby’s perspective as a non-expert (invaluable), his resort to the familiar hand-waving sophistries of Dreyfus and Searle (expected), or even whether I should call him Colby when I don’t really know the fellow and haven’t gotten around to amending my unwritten style guide to arbitrate matters of semi-personal address (pedantic). The bigger problem, one that is endemic in journalism about science, is his exclusive reliance on popular simplifications by corporate PR, other journalists, and cherry-picked philosophers for pictures of what AI research is all about.

Surely it wouldn’t have hurt to consult a real computing scientist; there are plenty of those to choose from the public sector with no vested interest in the fortunes of IBM. The only thing this would have jeopardized is a premeditated thesis founded on dismissive assertions about the entire field of research. Why talk to someone credible when they’re unlikely to agree with you?

Here, there are several bad assertions in play—all of which are traceable to the selective consultation of sources.

Let’s consider this one paragraph alone—the crux of Colby’s entire argument that nothing terribly fascinating is going on inside the box:

Jeopardy!, after all, doesn’t demand that much in the way of language interpretation. Watson has to, at most, interpret text questions of no more than 25 or 30 words—questions which, by design, have only a single answer. It handles puns and figures of speech impressively, for a computer. But it doesn’t do so in anything like the way humans do. IBM’s ads would have you believe the opposite, but it bears emphasizing that Watson is not “getting” the jokes and wordplay of the Jeopardy! writers. It’s using Bayesian math on the fly to pick out key nouns and phrases and pass them to a lookup table. If it sees “1564″ and “Pisa”, it’s going to say “Galileo”.

Now let’s put some numbers beside the assertions:

1. Jeopardy! is a trivia game, and all there is to trivia is looking up keywords. We know computers can do that.
2. When Watson handles wordplay, it doesn’t do it like humans do. It isn’t really thinking; it doesn’t really understand the puns. Furthermore, this somehow matters.
3. IBM would like us to believe that Watson really gets the jokes. If Watson doesn’t really get the jokes, the project is a hollow exercise in corporate self-promotion.

The first assertion vastly understates the complexity of what Jeopardy! demands. The nature of the game—a time-constrained, multi-agent affair—radically alters the straightforward problem of answering a question (or in this case, questioning an answer). Even simple pattern-matching is far from trivial when every millisecond counts.

Let’s run with Colby’s caricature for a moment. With a database of facts as gargantuan as the one Watson requires, looking up “1564″ in conjunction with “Pisa” is a surprisingly time-consuming task, never mind the inference to Galileo’s date of birth. This isn’t something tractable via faster processors or larger memory banks: there are theoretical lower bounds on the efficiency of searching and sorting algorithms in proportion to the dataset’s size. Exhaustive traversals that perform perfectly on small scales are out of the question here. The algorithms have to take shortcuts and make approximate guesses. Semantic associations must be efficiently structured in the software’s abstract maps as well as the physical database in order to best distribute searches in parallel. When you consider these factors, drawing semantic inferences from the natural-language clues becomes a heuristic necessity if the approximate search queries are to be any good.

Crucially, the time constraint on a response is not a static value, but a dynamic one that depends on the performance of the other competitors. This is why a match against the most successful Jeopardy! players in history is an essential proof of concept. Every contestant who appears on the television show has to pass a solo audition first, and any of them could tell you—particularly if they meet with little success—that in a competitive setting, the game becomes a different kettle of fish.

This is to say nothing of the other decisions Watson has to make in order to be competitive in a live test. It has to assess the risk of answering a question, considering not only its confidence in its own correctness but the standing scores of both itself and the other players. It has to set wagers for Double and Final Jeopardy, which requires an assessment of confidence based on the category title alone; in the case of Double Jeopardy, this will also have to consider the money still up for grabs on the board. One of the reasons Ken Jennings had such an astonishing run on the show was that he was able to make excellent strategic wagers on the fly.

Contrary to what Colby suggests, if the structured decomposition of the process of taking a Jeopardy! clue all the way from answer to question is able to match and surpass the blazing speed of human intuition at its best, that would be a tremendous accomplishment indeed. Without the capacity to parse natural language in terms of meaningful semantic chunks—a task well beyond mere symbol manipulation—Watson wouldn’t have a prayer of displaying a fraction of the competence that it has already shown.

Trapped in the Chinese Room

The second assertion is a real howler, and one that has become downright boring to swat aside over the course of the past thirty years. That’s right, folks: say hello to John Searle and the Chinese Room. The Chinese Room objection to AI is this: a computer translating between English and Chinese is like an English speaker who knows no Chinese, but who sits in a room looking up symbol tables and matching the syntactic elements correctly. Even if the translation looks perfect to the outsider, argued Searle, you couldn’t say that the symbol-manipulating translator (i.e. the computer) understands Chinese.

In a general sense, the Chinese Room stands for a whole class of arguments that boil down to saying, it doesn’t matter how well the computer performs—it’s not really thinking because on the inside, it’s not processing information in the same way humans do. Colby makes an argument about Watson identical to the Chinese Room when he says that the system doesn’t “get” the jokes and puns in Jeopardy!‘s more puzzling clues. Apparently, it doesn’t matter if Watson solves the clues correctly: it still isn’t behaving like a human inside the box, so the whole shebang is all just smoke and mirrors.

The logic of the Chinese Room is spurious in many respects, and I won’t go through all of the embedded fallacies here. For those of you new to the debate, here are two of the more serious ones. The first is that the analogy is false. The appeal of the argument comes from how it personifies a particular component of the system to highlight its dissimilarity to real human understanding. This fallacy endures unchecked because its proponents are free to move the goalposts however they like: no matter how robust the system is, the critics can isolate a piece of the syntactic machinery, put a human face on it, and complain about the absence of high-level, humanlike semantics. The second fallacy lies in the deceptive assertion that the syntactic internals of a computer are completely unlike the internals of the human mind. In truth, we still know next to nothing about how the latter works. Our understanding of how we get from the low-level operations of neuroscience to the high-level processes of cognitive psychology is at least as discontinuous as our best notions of how semantic structures might emerge from the symbolic structures of computer systems.

I alluded to this in my initial salvo on Twitter:

Shockingly poor article by @colbycosh on Watson, IBM’s #Jeopardy AI. Apparently, Chinese Room fallacies never get old. http://t.co/VHzLzTX

To which Colby offered this astonishing reply:

@Nicholas_Tam It’s got nada to do with the Chinese room. The Turing test is the one most everyone agrees on & there’s NO progress toward it.

Completely apart from the fact that one of Colby’s objections was precisely the Chinese Room, there’s a logical contradiction here along with a factual error. (Not bad, all in all, for 140 characters or less.) The contradiction arises from the failure to distinguish between external behaviours and internal thought processes. Let’s suppose, for a moment, that the goal for whichever AI system we’re talking about is to pass the Turing Test—that is, to be misidentified as the human being in a double-blind question-and-answer test where the questioner knows that one respondent is human and the other is a machine. If you read the original paper in Mind where Alan Turing introduced his “imitation game”, Turing’s whole point was to black-box the internals and take them out of the picture. The premise of the Turing Test is that if you can’t tell the difference between man and machine in terms of external behaviour, then functionally there may as well be no difference at all; this suffices as intelligence.

The Chinese Room argument, on the other hand, is a direct attack on the validity of the Turing Test. It seeks to establish that thoughts don’t supervene on actions: that is to say, identical external behaviours do not imply identical internal machinations.

Turing’s and Searle’s positions are more or less incommensurable. You can’t have it both ways. You can’t hold up the Turing Test (which is entirely about exterior performance) as the standard of achievement while griping, as Searle does, that even in a successful performance that passes for humanlike, symbol manipulation doesn’t really count. Contrariwise, Turing ventured that if a machine’s behaviour is indistinguishable from a human’s, it’s pointless to squabble over whether it qualifies as intelligent; from the available evidence, we might as well treat it as such.

If you accept the Chinese Room argument—and you really shouldn’t—the only function of bringing up the Turing Test at all is to set up a straw man. It has not escaped me that this may have been the intent.

Acting inside the box

Unfortunately for this transparent rhetorical tactic, the Turing Test is not the accepted benchmark for artificial intelligence research, nor is it even a commonly desired objective. AI is not one monolithic project that either has or hasn’t been achieved.

The goals of AI research have historically diversified along two separate axes (a schema for thinking about AI that most students of intelligent systems pick up from Russell and Norvig). The first key distinction is between acting (what a system does on the outside) and thinking (how a system gets there on the inside). The second distinction is between performing like humans and performing rationally or optimally (which may be entirely unlike humans, but may provide solutions to well-defined problems that outstrip the capacities of human agents).

This yields four quadrants that loosely circumscribe your garden-variety intelligent agents: systems that aim to think like humans, act like humans, think rationally, or act rationally. (Think of these categories more as design goals than as discrete kinds of agents, which in practice lie all over the map.) The first quadrant, systems that think like humans, is the area of interest for much of cognitive science. This is the type of system that the Chinese Room argument contends will in principle never succeed; Hubert Dreyfus’s objection, the thesis that human thought is fundamentally unformalizable, applies specifically to this category as well. The second quadrant, systems that act like humans, is the one where the Turing Test applies.

It must be said that the Turing Test is relevant here with specific reference to the indistinguishability of external behaviours—not to the requirement of aptitude in natural languages, as Colby seems to believe. Turing’s original imitation game was framed purely in terms of language, which remains an overwhelming challenge to this day, but it has since been expanded to other problem domains. (Jeopardy! is one of them.) To pluck out one example, natural language is hardly suitable as a test for computer vision, the branch of AI concerning how computers can perceive objects in photographs or positions in 3D space from the raw data of images. It would be preposterous to say that a robust system in computer vision fails as AI or marginalize its significance as a scientific accomplishment simply because it can’t pass for a human on the telephone.

Natural language is a particular problem domain—indeed, an umbrella category for all sorts of subproblems that are fascinating in their own right. It is not the essence of the Turing Test, nor is there any consensus that linguistic aptitude is the essence of intelligence.

It’s convenient for our discussion, however, that Jeopardy! involves natural language to the extent that it does. It should attract comparisons to Turing’s imitation game, and it has. Yet it bears mentioning that whether a system is really thinking is a completely incidental consideration for the vast majority of practical work in AI, just as it was for Turing. Nobody says, “Let’s build a system that possesses general intelligence.” What they actually say is this: “Let’s identify a chunky, intuitive problem that demands high-level thought and see if we can’t build a system to break it down and tackle it.”

Watson’s aim is clear: perform well enough in Jeopardy! to defeat the best human players. Any consequences for our beliefs about the nature of human intelligence is a byproduct and not the ultimate goal. That said, it is perfectly valid to speak of a Jeopardy! Turing Test. Watson would clearly fail the test not if it fell short of champion-level play, but if it ventured solutions to clues that don’t even make sense as guesses. (Consider the early test at about 1:50 into this video. The clue, from the category on I Love Lucy: “It was Ricky’s signature tune and later the name of his club.” Watson: “What is song?”)

But if indistinguishability from human-level performance is what we are looking for, Watson is already doing fairly well. There is a very important difference between defeating humans in Jeopardy! and passing for a human player, although the goals are intertwined. There is an even wider gulf between passing for a human Jeopardy! player and passing for a human being in toto. Everybody knows the latter goal is as far off as colonizing Mars, and nowhere in the promotional materials does IBM suggest otherwise.

Colby has a problem with this:

So why, one might ask, are we still throwing computer power at such tightly delimited tasks, ones that lie many layers of complexity below what a human accomplishes in having a simple phone conversation?

And one might also ask, why study nuclear physics when we seem to be no closer to harnessing fusion power than we were fifty years ago? First of all, in both cases, we are substantially closer in terms of how we understand the problem, even if our estimates for when the endpoint will show up on the horizon haven’t necessarily shortened. The achievements that scientists think of as the most significant may not be fixtures in popular culture, but that doesn’t mean they were pointless. Far more importantly: computing science, like nuclear physics, is inherently interesting. Designing AI systems for delimited problem spaces is an activity that leads us to all sorts of discoveries about the nature and structure of those problems, and of the minutiae of problem-solving processes in general. We learn all sorts of things about comparative strategies for structuring, representing, and manipulating information—and how they measure up to the relatively black-boxed processes of human minds.

So to answer Colby’s question:

@Nicholas_Tam So we can’t test AI by scrutiny of interior process OR the curtained-black-box Turing test? What’s left, religious revelation?

We “test” AI in the context of its performance with respect to well-defined goals. Those goals could certainly involve a Turing Test, be it for answering natural-language questions or some other specified task. Whether an artificial system has a human-like mind of its own, along everything that implies—consciousness, self-awareness, semantic understanding—is a problem we leave to the philosophers; and no, it’s not empirically testable. But neither is the problem of whether other humans have minds.

The inverted pyramid scheme

Now let us turn to the third assertion: that IBM is making outlandish promotional claims that oversell Watson in the name of fuelling a publicity blitz.

What does it mean to say that something is a “gimmick”? We mean to accuse it of being all dressing and no salad. We mean to expose its failure to accomplish what we are told it does on the surface. We mean to insist that we will not be duped into believing that something humdrum is, in truth, extraordinary.

The trouble for Colby’s argument is that Watson is extraordinary—just not in the way that he thinks IBM has misled him to expect. “AI researchers have arguably the highest conceivable standards to meet when it comes to thinking about thinking,” remarked one commenter at Maclean’s, “and it’s hard to fault them for failing to live up to the naive expectations of science fiction.” Colby replied: “By ‘the naive expectations of science fiction’ I presume you mean ‘the naive expectations deliberately created by IBM promotional materials and employees’.

I received a similar response:

@Nicholas_Tam Maybe you should look at the IBM ads. Your claims for Watson are a LOT more modest than theirs.

At the time of our repartee, I was admittedly only familiar with IBM’s own materials in passing; most of what I knew about Watson was from sources that discussed it in greater detail. I found it odd that Colby’s point of engagement was exclusively with the advertising and not the technology itself, but this was understandable: he was making a statement about hype, after all, and it’s very common nowadays that the implications of scientific accomplishments are exaggerated in the public sphere. (Refer to Jorge Cham’s excellent illustration of the science news cycle, which concerns university research but applies equally as well to corporate and governmental laboratories.)

By and large, this is a product of two sets of behaviour—one on the part of journalistic reporting, the other on the part of the research organizations. Let’s begin with the journalists.

The dominant template for journalistic narrative is the inverted pyramid: begin with the most important information, and continue to points that are less and less essential on the assumption that the reader could stop at any time. (Before the age of desktop publishing, this also made it easy for newspaper editors to literally snip away the last paragraph or two when assembling the columns on the page.) The trouble is the gulf between what journalists deem most relevant to non-expert readers and what scientists consider to be important contributions to their field.

The end result is sensationalism—and too many articles about science wind up looking like Martin Robbins’ parody. They begin with far-reaching implications that may or may not be related to the research at hand, and work their way down to the specifics that matter most. This is a narrative framework that is seriously divorced from the reality of research, which operates on the level of local challenges and goals. (This post by Greg Lusk on the inverted pyramid and the conflicting priorities of journalists and scientists is highly relevant here.)

Because long-term, big-picture implications like the performance gap between artificial and human intelligence (in Watson’s case) become the centrepiece of the story, they become the focus of media attention and debate, often with no consideration of the specifics of what has been accomplished. And this is why we see casual expressions of dissent like Colby Cosh’s criticisms of Watson: wildly off the mark, selectively researched from Wikipedia with an a priori verdict already in mind, and laced with a sprinkle of pseudo-expertly mumbling about Bayesian combinatorics that are far more involved than the author makes them out to be. Criticisms like these respond to the news stories, not to the science.

Of greed and gimmickry

Colby is convinced, however, that his projected misunderstandings of what Watson claims to achieve are fundamentally IBM’s fault. And it’s no use pretending that IBM isn’t a self-interested organization: like NASA in their recent fiasco over arsenic-based lifeforms (a discredited paper, but one that was widely misreported when people still thought it looked shipshape), if people take their promotional materials and statements to the press the wrong way, they have no incentive to correct anyone so long as their project is still in a positive light. Watson is a proof of concept for IBM’s enterprise hardware and the DeepQA question-answering system, both of which the company intends to license and sell.

Not all of the problems with science journalism is the fault of journalists: research laboratories, public as well as private, are often complacent about inaccuracies in secondary reporting because the attention (and the concomitant prospects for funding) are too attractive to throw away.

Let’s be very clear about one thing, however: IBM’s profit motive as an organization does not negate the intellectual interests of its researchers. As fashionable as it is these days to appeal to the trope of corporations that are only responsible to their shareholders and therefore can’t be interested in anything but the bottom line, the truth is that corporate laboratories in private industry are invaluable centres of research. Projects like Watson attract contributions from university scientists not because they all want to see IBM succeed, and not even necessarily because the pay is so much better (though it is), but because they provide access to hardware that enables large-scale work. Computing scientists in industry are taken every bit as seriously as their compatriots in the university world, and the two regularly cooperate on grand initiatives.

But what does that say about the marketing? Complacency aside, is IBM actively making Watson sound like a much bigger deal than it is?

I have now combed through IBM’s promotional videos, articles, and FAQs, and I would like to retract my earlier concession that their claims may have gone too far. IBM’s statements about Watson are fair reflections of what AI can realistically achieve and what a successful performance by Watson will demonstrate. About the most outlandish thing they say—the one that treads the furthest into the minefield of the philosophy of AI—is that Watson performs well in Jeopardy because it understands natural language. And strictly speaking, it does. The clues in Jeopardy! are undeniably in natural language, and differ from formal or heavily restricted sentences by a significant degree of complexity. About the only restriction on the clues is length. Discard the puns and puzzles and you still have challenging problems like binding indefinite pronouns to objects (or classes of objects) that fit.

Whether Watson’s “understanding” of natural language is analogous to that of humans doesn’t figure into the discussion here. Nobody is saying that Watson actually has a conscious mind; AI researchers don’t think on those airy-fairy ontological terms when they are designing systems for specific tasks. They participate in the debates over the philosophy of artificial minds, yes, and they’re usually on the optimistic side, but everyone is aware of the separation between that conversation and the immediate challenge of defeating humans on a robust, open-domain answer-questioning game show.

We are not even remotely in Dreyfus territory. Still, I can understand why layperson readers might think we are when they read the story in The Globe and Mail and come across a juicy quotation like this:

“We can use computers to find documents with keywords, but the computers don’t know what those documents say,” Dr. Ferrucci says. “What if they did?”

People whose notions about AI come entirely from Battlestar Galactica could easily misread Ferrucci’s statement as referring to sentience or consciousness. But anybody who knows a thing or two about AI can read this and correctly interpret it to refer to semantic-level knowledge representation—concepts on a larger scale than string matching or keyword search. It’s entirely agnostic on the problem of whether artificial minds can exist. I’m not deliberately reading this as a modest apologist: this is actually what Ferrucci is obviously saying.

If you get all your science from Hollywood and you think cloning has to do with developed bodies and selves rather than the raw data in your genes, it’s not the responsibility of geneticists to clarify their work for you every time they speak. Similarly, you can’t expect scientists and engineers in AI to explicitly backpedal from the philosophical question of conscious machines every time they talk about their work.

Or can you? What we desperately need is a greater public understanding of what scientists do, and what they mean when they use everyday words to talk about their fields. Readers dive into news stories about science with popular preconceptions that are often wrong, but nobody takes up the responsibility of correcting them until the discourse goes seriously awry. We’ve seen this before with how the hysteria over genetically modified foods or embryonic stem cell research obfuscated the real issues deserving of policy attention. There are even some dark corners of the world where creationists are wreaking havoc on schools because they still think evolution by natural selection is some kind of affront to their god.

Sooner or later, this will happen with AI: we’ll explore the possibility of delegating something big and very public to an autonomous system, and legitimate policy concerns will drown in a sea of hysteria about machines taking over the world. If scientifically knowledgable people do not shoulder the burden of sober clarification, that role will become occupied by contrarian journalists who don’t really know what they’re talking about, but still take pleasure in posturing as the voice of reason in the room.

If you are going to take the position of someone who sees through the publicity and understands the underlying science, you have to understand the underlying science. No matter how bombastic IBM’s promotional claims are, or how submissively the media repeats the press releases with a dash of unchecked sensationalism on top, Watson is more than a “gimmick” if it’s computationally interesting—and by any informed and reasonable standard, it is. Watson is a nontrivial system, and Jeopardy! is a nontrivial pursuit.

Yesterday, Apple launched the Mac App Store, the latest interface refinement capitalizing on the observation that increasingly, Mac OS X is not likely to be a new user’s first Apple product. Just as we saw iTunes navigation features such as Cover Flow migrate over to the OS X Finder, now we are seeing OS X take after the iPhone/iPad user experience by delivering software via one-click installations. Click on the button to purchase or download an app, and the App Store dumps it in your Applications folder and on the Dock.

OS X veterans will note that this is, in theory, a 200% improvement on what was formerly a three-click installation: download the DMG, drag the app into the Applications folder, and optionally dump it on your Dock as well if you are still oblivious to the eldritch wonders of Quicksilver.

And in theory, one would think this is one of the best features yet to arrive on the Mac for users and developers alike. The end-user software culture for Mac users has always been very distinctive: unlike the unfortunate bifurcation in the savage lands of Windows, where software is often either a) homegrown and free or b) professional and exorbitantly priced with corporate site licenses in mind (and therefore often pirated), Mac software for the individual consumer is pretty much where it was in the early 1990s: practically anything that Apple didn’t hand you with the system comes from independent development houses, usually in the form of try-now, buy-later shareware, their products reasonably priced. Compared to other platforms, good free software is much harder to find.

(Incidentally, the Mac App Store has already delivered a 1990s time-capsule feeling of its own, raising the hungering corpses of products long believed dead. I mean, StuffIt Expander? Kid Pix?!)

For this model, digital distribution was a dream come true from its inception, and it would make sense to believe that a centralized distribution channel for downloads and updates only improves on it. In practice, however, there is no advantage to using the Mac App Store for anything that is already available directly from the developers.

Apple’s approval process effectively ensures that software updates through the App Store will lag behind the automatic updaters that already exist. Buying directly from the developer relieves them of Apple’s 30% cut for products sold via the App Store. There is also no support for time-constrained shareware trials, which are far and away the best way for developers to demonstrate why their software is worth paying for.

Product licenses, thank goodness, are bound to your Apple account instead of your machine; if something disastrous happens to your computer, you can always download your purchases again later, and there is no limit to the number of machines you can install them on. This is still more annoying than the DRM-free status quo of “punch in your product code and we’ll trust you the rest of the way,” but at least with a centralized ID you don’t need to worry about losing your product code.

In any case, informed users accustomed to hunting for quality Mac software that didn’t come pre-installed have no incentive to use the App Store at all except for software that is otherwise unavailable. The App Store’s function is to inform everyone else that third-party software even exists. Developers are effectively compelled to push their products onto the App Store in order to remain exposed and competitive, but if they make their products available directly, it’s hard to think of a reason why one wouldn’t obtain them that way instead.

Postscriptum. Speaking of third-party Mac software, I am presently composing this post in MarsEdit and finding it wonderful. I may end up blogging more frequently again purely for the pleasure of using it. Daniel Jalkut, the man behind MarsEdit, wrote an informative FAQ about the Mac App Store and what it means for his product. It confirms most of my sentiments above.

Somewhere along the way, I penned a dissertation of sorts, but let’s not talk about that. Here is the crust of readings that has built up in the meantime. There are more, but the list below was becoming rather overgrown and at some point I had to stop.

• Two of the great figures in things I care about passed away in May, both of them at ripe old ages after leading fulfilling lives: jazz pianist Hank Jones at 91; mathematical popularizer and Lewis Carroll expert Martin Gardner at 95. I came to both Jones’ and Gardner’s works late in life but quickly—very quickly—came to understand their immeasurable impacts on music and mathematics, respectively, which I had previously felt secondhand without being aware of it. More on Jones here and here; more on Gardner here and here.

• It speaks volumes for how long I’ve been away from saturating this page with hyperlinks that sitting atop the pile in my draft box is an ominous article by Dominic Lawson on David Cameron and Nick Clegg’s public-school upbringings at Eton and Westminster, written the week of the first televised debate.

• IBM has developed a Jeopardy!-playing computer. Observe the promotional video. From an AI perspective, this is orders of magnitude more exciting than Deep Blue, and takes us deep into Turing Test territory. I hope to say more about this should I find the time.

• One of the disadvantages of being in the United Kingdom—indeed, the most serious one I have yet encountered apart from the absence of fine, extravagant steaks—is that for the first time since 1998, I was unable to see a new Pixar film on or before the date of its release. Two Pixar films of note, in fact: Toy Story 3 and the accompanying Teddy Newton short Day and Night. That hasn’t stopped me from following the resurgence of coverage of Pixar’s process of perfection in this Wired piece and this interview with Lee Unkrich.

• Typesetting matters, folks. Just ask the consummate professionals behind these two book-size online resources: Typography for Lawyers, and LaTeX for Logicians.

• Everyone with an interest in the romance of modern international affairs has read it already, but Raffi Khatchadourian’s profile of WikiLeaks founder Julian Assange is an outstanding piece of storytelling, if also one that tends towards the making of myth.

• And while on the subject of journalism and international intrigue, here is the Rolling Stone feature on Stanley McChrystal that led him to be sacked from command in Afghanistan.

• Civilization V is on its way, but there’s still plenty to say about Civilization IV. Troy Goodfellow shares a letter from Christopher Tin about composing music for the game. Kotaku asks lead designer Soren Johnson about the mathematization of religion.

• Jeremy Parish reflects on this year’s Electronic Entertainment Expo (E3) and calls out much of the game industry for the creative bankruptcy of video game violence.

• Neil Swidey of The Boston Globe courageously explores the mind of the anonymous comment-box troll.

• As this year’s graduate session at Singularity University gets underway, The New York Times talks to Ray Kurzweil and gang about the posthuman lifestyle.

• John Naughton writes in The Guardian about what the Internet has really changed.

• England has been swept up in the pathos and misery of football fever, as usual, and one may as well get some World Cup readings out of the way before the Three Lions have truly met with yet another ignominious doom. (Or, preferably, they could win.) Tim de Lisle enquires into the origins of spectator sport’s global draw. And then there’s this article on the North Korean national team, published in timely fashion just before Portugal blanked them 7-0.

• Finally, the only thing that can stop this asteroid is your liberal arts degree.

And now for something completely different:

• Start with Michael Weingrad’s piece in The Jewish Review of Books about why there is no Jewish Narnia. Then proceed to Israeli sci-fi reviewer Abigail Nussbaum’s response and her survey of the conversation.

• My friend Stephen McCarthy, who is coaching Korean schoolchildren in the art of debate, writes about his cultural collision with corporal punishment.

• Anthony Gottlieb digests a survey of what philosophers believe. The data set covers English-speaking academia and skews heavily analytic, but I’m not one to complain.

• Not exactly “reading” per se, but it’s election time, and I can’t stop playing with The Guardian‘s lovely polling widget.

• Wikileaks is in the news again after releasing footage of American troops firing upon a Reuters photographer in Iraq. The BBC profiles who they are.

• John McWhorter parses Sarah Palin. Typically the way the print media scrubs audio quotations into coherent, well-formed sentences (or doesn’t) is a good indicator of media bias, but the thing about Palin is that it can’t be done.

• Julie Just asks where the parents have gone in fiction for young adults.

• What are marching bands playing these days? Shostakovich, that’s what.

• Dale Dougherty writes about the iPad and misses HyperCard. He’s not the only one.

• Cartoonist James Sturm leaves the Internet. I should do that too.

• Turn up your speakers and read Jan Swafford’s article in Slate about performing classical piano repertoire on classical pianos, which is full of audio comparisons that will make you wonder if the homogenized ideal of the modern Steinway grand is really a good thing.

• The Guardian asks a wide selection of novelists for their writing tips, which have a way of telling us more about the authors than about writing. Some of my favourites: Geoff Dyer (“Don’t be one of those writers who sentence themselves to a lifetime of sucking up to Nabokov”), Anne Enright (“The first 12 years are the worst”), Philip Pullman (“My main rule is to say no to things like this, which tempt me away from my proper work”).

• Ben Goldacre shows us how regulating alternative folk medicine through requiring certification is no use at all when we don’t know what’s being certified.

• From The New York Times: Canadians shoot left, Americans shoot right. The article is about hockey players but I think there’s something bigger in this.

• Teresa Nielsen Hayden remarks on the imaginative poverty of failed authors who think suing J.K. Rowling for plagiarism is a good idea.

• Jonah Lehrer wonders if the direction of funding towards older scientists hinders us from tapping into the creativity of youth. Also read the followup on his Frontal Cortex blog.

• Civilization IV lead designer Soren Johnson talks about designing strategy games around our intuitions about probability (or lack thereof).

• Mark Chu-Carroll explains why computer simulations of biological phenomena will never replace animal testing.

• Joel Stickley’s explorations of bad writing by example finally catch on to my fatal flaw.

This week’s bag of links:

• In a rare sighting of the man behind Calvin and Hobbes, Cleveland newspaper The Plain Dealer interviews Bill Watterson fifteen years after the legendary comic strip ended its run.

• Peter Hum ruminates on the “ugly beauty” of avant-garde jazz.

• The big news coming out of Barack Obama’s 2011 budget was the abandonment of NASA’s plan for the resumption of manned spaceflight to the moon. SPACE.com has the analysis.

• Jonathan McCalmont, caught between the debate over high/low culture and his vehement dislike of the popular video game Bayonetta (“a game so dumb that it makes a weekend spent masturbating and sniffing glue seem like an animated discussion of Wittgenstein’s Tractatus Logico-Philosophicus (1921)”), spun it all into a compelling essay on hypnotism and lowbrow art.

• This Charles Petersen piece in The New York Review of Books is one of the better histories you will find of where Facebook came from and how it has transformed, and offers a thorough look at the content-pushing pressures facing the social-network model of a nominally private Internet.

• Mark Sarvas identifies some common problems of debut novels from the perspective of a prize-committee veteran.

• In The Guardian, Darrel Ince implores scientists who rely on internally developed software to publish their source code.

Many are rightly wondering if Apple’s iPad really does fill a niche that isn’t already better served by a laptop and a phone (specifically, Apple laptops and Apple phones). I can’t speak for anyone else, but I’ve long desired a device that allows me to do two things:

• Read full-size PDFs and websites in bed;
• Prop up music books on the piano in digital form, complete with gestural page-turns and sheets that don’t blow this way and that.

The iPhone and iPod Touch can’t do this because the screen is too small.

Laptops can’t do this because they go practically anywhere but on your lap, you can’t set them on the bed because they’ll set your house on fire, and the keyboard juts out and gets in your way. The screen orientation is also unsuitable for most PDFs. If you use a laptop, you are practically tethered to a desk. (I have, incidentally, seen a few musicians who put their laptops on the piano as a substitute for lugging a bagful of Real Books around. The form factor leaves much to be desired.)

What about e-readers? I’m astonished at how poorly existing e-readers have handled PDF support. The Kindle, last I heard, allows you to convert PDFs into its proprietary format so you can interact with the text the way you do with any of the books available for the device, but this completely fails to handle the kind of documents I tend to read as PDFs in the first place: music, articles in academic journals (often with diagrams, footnotes, and figures all over the place), and other scans that are sensitive to their original layouts. While the iPad can’t hope to match the battery life and screen texture of dedicated e-book readers for, well, reading books, a bright full-colour screen is exactly what I need for the kind of documents that wind up on my drive as PDFs.

The iPad is perfect for both of these tasks. By the looks of it, I can hold it in any orientation in the laziest of postures without strain, and it will sit nicely on any music stand. It’s an absolute dream for musicians, and the ideal device for someone who needs to pack a lot of stray documents on the go. Who knows—it may even save The New York Times, and I wouldn’t be in the least surprised if this was a major reason why the Times had the gumption to announce it would move its online edition to a paid subscription model next year.

What’s odd, then, is that Apple is falling short of its usual marketing savvy in promoting the features of the iPad as if it were merely an iPod Touch with a bigger screen. The company is clearly expecting the revolution to come from third-party application developers, as was the case with the iPhone, and banking to a lesser extent on the massive content push of its iBooks store, but this seriously undersells the potential of the device.

Combined with a keyboard dock, the iPad is potentially a complete computer replacement for everything I do except a few heavy design/development applications, World of Warcraft, and Civilization—essentially, every reason I have a MacBook Pro instead of the lightweight standard line. And as comfortable as I have become with using LaTeX for all of my document preparation, I am even willing to go back to a word processor like Pages if someone develops a good implementation of speech-to-text, so I can try the Richard Powers method of prose composition. Most people don’t use their computers for any of these tasks, and so long as there is an adequate file management system—something we have yet to see—the iPad could be viable as a standalone device. Keyboards are around to stay, but it’s only a matter of time before the mouse paradigm is dead.

• I was already aware of Ren and Stimpy creator John Kricfalusi as a superb animation educator via the meticulous frame-by-frame studies at his blog, but Letters of Note has a real treat: a letter from Kricfalusi to a 14-year-old aspiring cartoonist.

• Rohan Maitzen makes a passionate argument that the value of a literary education is in the study of literature, not just the ancillary job skills that English departments cite to defend their own worth. (Continued here and here.)

• Jeff Foust surveys the debate over the scientific value of human spaceflight and what it means for NASA policymaking now.

• Sarah Eve Kelly, whose Anne Boleyn novel got picked up by an agent and is currently being shopped around, tells writers inundated with industry advice to shove it aside and get cracking on a draft.

• In the Columbia Journalism Review, Miles Corwin gives us a look at the young Gabriel García Márquez as journalist.

• The Economist‘s Democracy in America blog muses on the conservatism of the Muppets.

• Finally, in anticipation of whatever Apple is announcing this week, Beat-era poet Gary Snyder shares a poem about his Mac.

This just in: I can’t believe I crapped out a thousand words on a niche issue that might have mattered to half a dozen of my acquaintances three or four years ago, when I have academic work to do and a deadline coming up fast. It bothers me. It is costing me sleep.

But I know why I did it. It was Georgia.

You won’t have noticed it if you only read my site on Facebook or your RSS reader, but this week I made some minor tweaks to my layout, and it got me thinking about web design again for the first time in many moons. It began innocently enough. I’d fallen in love with Georgia again, thanks to the online edition of the The New York Times. This is what happens when you read the Times as much as I do: the typesetting becomes inseparable from the text, the text indivisible from the Web; and so daintily, transitively, your memories of other faces slip away like dingbats in the cold, long night.

I’m usually cautious around these upstart fonts for the screen—Georgia is practically an infant, designed in 1993—and ever since I brought my site into its present incarnation I’d stuck with old, reliable Garamond the whole way through. Garamond the Wise, Garamond of Many Colours! How soon had I forgotten that in my former locale, Georgia was once my face of choice. Maybe this is why typographic fashion has borrowed the language of haute couture: one look at the Times and you tell yourself, I want to look like that. Those curves, those stately majuscules.

So I opened up my stylesheet and changed the type. Before I knew it I was fiddling with a margin here, a colour there—minor cosmetic obsessions, nothing big. Then the title image; then a plank for recent comments along the starboard side. It’s not perfect, but it’s tidier now and I felt an overpowering urge to write some copy just to give myself an excuse to look at it.

Don’t worry, though—I’ve not cast away my classical tastes. Jenson remains the champion of the page.