macksting

Layman’s Musings on AI

I’d like to preface this by saying I don’t actually care much about the difference between strong AI and weak AI. The nature of consciousness is already up for debate before we start saying a machine can or cannot achieve it. We wouldn’t even judge this by the same measures we judge our own. (Setting aside your feelings on Descartes, how long do you think it would take a computer to say “Cogito, ergo sum?”)
In the novel Do Androids Dream Of Electric Sheep?, Philip K. Dick, who of course was not a roboticist, wrote Deckard’s observations on the new Nexus-6 type as follows:
‘The Nexus-6 androids types, Rick reflected, surpassed several classes of human specials in terms of intelligence. In other words, androids equipped with the new Nexus-6 brain unit had from a sort of rough, pragmatic, no-nonsense standpoint evolved beyond a major – but inferior – segment of mankind.’
If you notice some failures of empathy and ill-treatment of the neurodivergent, congratulations, you’ve cottoned to one of the themes of the novel, and some of Deckard’s many flaws. I won’t revisit the question of Deckard’s assessment of people as inferior; Deckard is Deckard, and kind of an ass.
This brings me to Mario.
I was watching a relatively simple AI learn to play Super Mario Bros in a YouTube video. It took a while to learn how to go to the right, and then to jump. It was trained initially on the player’s own gameplay, which was objectively inferior to my own. He fully admitted he was rather bad at the game. Before long the AI was actually quite decent at the game by a few measures, and wouldn’t take long to outstrip the teacher.
In the 1980s, a “self-driving car” was a huge device, and not especially skillful. They could achieve speeds of nearly 40 mph on a road with no other traffic. Not useless, but not amazing, either. Now, of course, we have self-driving cars which are primarily hampered not by the existence of traffic, but by human error in the traffic around them. (In short, the cars apparently work very well together, and if we got the humans out from behind the wheel they’d be even better.)
Of course, at present an artificial intelligence tends to give itself away in conversation. Its subject matter is often repetitive, not showing any real understanding of the words it’s speaking. Tay the twitter bot is a fine example, able to parrot hate speech with variation and a high degree of success, but only because it was taught how to do so by a rather warped segment of the population.
At MIT, however, there’s an algorithm which has studied hours upon hours of situation comedies. Given still shots, it can predict the next five seconds of human behavior 11 percent of the time. Humans in this test performed better, predicting what would happen (for example a handshake) about 71 percent of the time.
I have Asperger’s Syndrome. I do have some gifts as a result of my mind’s odd manner of thinking, but it’s not exactly a fair or even trade-off, and certainly in this regard I’m what Deckard would call a “special.” I might even judgmentally be called a chickenhead by some folks in the setting.
We don’t have any Nexus-6 types yet, but I can’t help but think it won’t be terribly long before we do. All else being equal, we will eventually have artificial intelligences trained on police dashcams, war games, game theory scenarios, and just plain old conversation which will outstrip its teachers in many regards.
They might end up being pretty darn good conversationalists. They’ll certainly be very good at Galaga.