The Turing Game started as a thought experiment called the Imitation Game.
Alan Turing proposed that machines could be intelligent in his 1950 paper, Computing Machinery and Intelligence [22 pages]. It's one of the most idiosyncratic papers in the history of science and philosophy. It's a fun read.
Turing basically said: Look, we can’t even define what "intelligence" is. But if a machine can convincingly write like a human under unscripted circumstances, isn’t that good enough to call it "intelligent"?
This came to be known as the Turing Test and it's proven useful. The big example is CAPTCHA, which stands for "Completely Automated Public Turing test to tell Computers and Humans Apart."
But for the purposes of machine learning, the Turing Test has serious practical disadvantages.
The Turing Test is adversarial.
It literally trains the machine to view humans as opponents, and vice-versa. I don’t know about you, but I like to avoid having a combative relationship with my computer.
The game pits machine vs. human in a battle of apparent human-ness. No wonder human players would sometimes show up for a fight, as though defending their title.
The Turing Test is zero-sum, winner-take-all.
This severely limits the scope of play. When winning is the only thing that matters, alignment becomes irrelevant. Not much learning happens either.
Cooperative (or variable-sum) gameplay increases alignment, training the machine to work with people for greater mutual benefit. In real life, most decisions are about trade-offs of varying degrees.
The Turing Test rewards the machine for deception.
If you're A/B testing, I'd rather an intelligent machine be honest. That's better alignment.
The Turing Test isn't much of a test.
The first machine to successfully pass itself off as human was ELIZA in the early 60s. It talked in the language of Rogerian psychology. This was an easy Q-and-A format to emulate because it would mirror humans' statements back at them in the form of questions. This allowed for fluid conversation with a machine – fluid enough so that it felt like a person. But, like a magic trick, once you're shown how easy it actually was, the magic disappeared.
It's a predictable result of what Dan Dennett called the "intentional stance." It's our hyper-sensitive human-agency-detection system. (Dogs have it too. That's why they bark and growl at unfamiliar noises.) And it certainly also plays into our ability (eagerness?) to project ourselves into our environment.
"As human as..."
Instead, Five Pines asks people – artists, scientists, or people who just have a project they want to do – "try working with this system and see if you can find some use in it." And FAC usually does help, even if it's only as a neutral sounding board. Each Turing Game is a exercise in creative problem solving.
We try to keep the scope of play as wide as possible. Just as a Universal Turing Machine is a machine that can be countless other machines, the Turing Game can be made of any number of games. Part of my job running Games is to get the human and FAC to play constructively with one another.
This is where priciples of stage improv become useful. It's not widely appreciated but what we call "improv" started as educational and childhood development techniques in the early 1900s. It uses gameplay for creative problem solving and it's something enjoyable to do. I highly recommend the work of Neva Boyd at Chicago's Hull house and her books on gameplay with kids.
[BTW, I suspect Turing read Boyd's work. His intelligen machine idea got him preoccupied with childhood mental development and games. So he would have been familiar with her work. Although I've found no direct reference to Boyd from him, Turing's description of the Imitation Game is very similar to entries in Boyd's book on children's games. Forgive me, but this touches on two documentation nightmares from my thesis research days. A Turing-Boyd connection was never found. Also, I never found a real-world example of an imitation game, which could have been an actual game, in Victorian culture. But I found no evidence. It seems Turing made it up himself.]
I'll close with a quote from game designer Brenda Laurel, from The long-term impact of IT culture [36 mins]
"The questions and the answers are human. And the computer needs to be human, too. As human as language. As human as a film. As a talisman. As a fairy tale. As a song." — Brenda Laurel
- - -{@
No comments:
Post a Comment