Monday, May 22, 2017

Deep Donald

(New Google Translate poems at bottom)

Doughnuts by vegan chef Sam Melbourne, via Lost at E Minor.
In contemplating the science-fiction Singularity, we always picture the machines learning to think, and becoming more and more like people; not the possibility that the alienated humans will become more mechanical, but a cool idea from the neuroscientist Robert A. Burton in the Times's philosophy department suggests thinking this way about our Emperor Trump could account not only for his deep strangeness but also for his incomprehensible success:
If conventional psychology isn’t up to the task, perhaps we should step back and consider a tantalizing sci-fi alternative — that Trump doesn’t operate within conventional human cognitive constraints, but rather is a new life form, a rudimentary artificial intelligence-based learning machine. When we strip away all moral, ethical and ideological considerations from his decisions and see them strictly in the light of machine learning, his behavior makes perfect sense.
A "deep learning" program like Google's Deep Mind or IBM's Deep Blue, programmed to accomplish a specific task (like winning a chess game, or an election) by mapping the data of previous efforts onto the background of the current situation, the same kind of heuristic that is used by the Google Translate algorithm we've been having fun with.

Burton can suggest a startlingly persuasive account of how a Deep Donald could have won the election, by having the single objective of winning and undisturbed by any other motivations or calculations of consequence after the election:
imagine that the Trump A.I. machine determines that the opinions most likely to get him elected have slim chances of being carried out once he is president. To the extent that traditional politicians are embarrassed by flip-flops or bound by ethical and moral values or both, they are likely to display a degree of restraint and hedging of controversial positions. Winning at any cost is at least partially balanced by underlying principles. Not so for a neural network. There is no cringe factor, no sense of anticipated embarrassment or humiliation with seemingly random changes of mind. There is no concern with subsequent disclosures of misrepresentations, falsifications and outright lies. As the U.C.L.A. Bruins football coach Henry Sanders, known as Red, once said, “Winning isn’t everything; it’s the only thing.”
But for what happens to Deep Donald after the election, there isn't such a clear model, presumably because he hasn't been programmed with an adequately narrowed goal; a president doesn't just "win", though you'd think Trump might be focused in that way on poll numbers, say. Of course poll numbers are much more difficult than elections, as a game-theoretical exercise, and the technique isn't very well suited to moving them. In the election you have two fundamentally simple tasks, to get enough votes (which is only about 26% of the available electorate) and to see that your opponent doesn't; in becoming popular you haven't got the advantage of competing against anybody but yourself, and thousands of untethered variables to tack into.

Like Deep Blue against Kasparov, Trump was superbly designed to win the election: Kasparov's particular shock in the last tournament (May 1997: just 20 years ago!) sounds like something Hillary Clinton must have felt, a little too late, as the returns started coming in:
Kasparov began showing the effects of facing an opponent whose limitations were no longer certain. When the machine would make an unexpected, possibly foolish move, he'd look stunned. You could almost see his inner dialogue: "Is this idiocy? Or something so brilliant that it only looks stupid?''
But now that Trump's task is the supremely ill-defined job of being president, it's not surprising that his behavior should be seeming more random and erratic after the inauguration, less frightening and less effective. And that all the idiots—Mark Halperins and Mike Allens and Chris Cillizzas—who decided Trump was a genius during the campaign because of his apparent skill at the only thing they value, winning, are baffled now by his floundering inability to embark on anything or defend himself from attack or even conceal his criminality, as apparently helpless as a porpoise dumped on a miniature golf course.

Back to the Google Translate analogy, I'd love to see the Senate Judiciary Committee trying out the technique of prompting him with a single repeated word to see if they could break him down in the same way as we break down the translation algorithm:
FRANKEN: Doughnut, Mr. President.
TRUMP: Excuse me?
FRANKEN: Doughnut. Doughnut.
TRUMP: I would rather not see.
FRANKEN: Doughnut. Doughnut. Doughnut.
TRUMP: I would rather not see the frangipani.
And so on. I'm not going to offer the Ionesco play this procedure could produce, but a couple of new Google Translate poems are appended below. I can't put my finger on the post-Imagist style the Japanese one reminds me of, but the Korean is pure Sam Beckett.

A lukewarm
A rainy season
A warmth
A blovering thing
A dudley
A brief bloom
Ineem Mu
Muddy asleep
(Hiragana ぬ/nu, algorithm invited to translate from Japanese)

Speak to Not to Be
ぬ ぬ ぬ ぬ
I do not want to go
I do not want to go
To be able to
Speak to not to be
Speak to not to be
To be able to see
Speak to not to be
To be able to run in a short time
To be able to be in a hurry
(same Japanese character, algorithm asked to translate from Korean)

No comments:

Post a Comment