Meet on the ledge
|
|
|
AI creeps nearer via bee algorithms and crowdsourcing
JEOPARDY CHAMPION Ken Jennings quipped, "I for one welcome our computer overlords," when IBM's Watson computer defeated him in February 2011.
"It's hard to make a joke on Jeopardy," he said Sunday, "because there isn't much time." He was speaking at the Singularity Summit, the sixth annual gathering to check on progress to a fistful of possible futures that might be science fiction or might be…real?
The physicist and science fiction author Vernor Vinge posited in 1993 that, "In 30 years we will have the means to create superhuman intelligence. Shortly after, the human era will be ended." Vinge called the transition the Singularity after the physics phenomenon. The concept was taken up by the engineer and inventor Ray Kurzweil, who perennially predicts a completion date of 2045.
Is Watson a seminal moment? Many contemporaneous news reports were simultaneously vague and hyperbolic. Explained up close by IBM's David Ferrucci, Watson is more impressive: it's not a database trying to anticipate all possible questions but a harnessed team of hundreds of different algorithms whose results must be weighted and assessed with a degree of confidence.
But consider the imbalance. Watson: 2,280 processor cores, 15TB of RAM, 20 tons of air conditioning, requiring 80KW of electricity to power it all.
"Compare that to a brain that weighs two pounds, fits in a shoebox, and that you can power with a tuna fish sandwich," said Ferrucci. Plus, as he didn't say, much cheaper reproduction. Watson cost upwards of $3 million.
What makes Watson difficult to appreciate is that the action is all internal. Rice University professor and former Irobot engineer James McLurkin, by contrast, believes that if everyone loves a robot, then everyone will love swarms of robots so much more. While his team of 20 or so wheeled square robots sorted themselves, lights flashing, into a line by ID number and set out on a treasure hunt, McLurkin said his inspirations are swarms of bees and internet routing.
Bee algorithms are good, he explained, "because they've been debugged over 160 million years." Twenty robots can look for hot spots in forest fires; 200 can search earthquake-stricken areas for survivors; 2,000 could explore Mars; 10 million could perhaps reclaim a landfill. All jobs humans can't do.
Tags: Software
Share this:
Share
We are getting closer to the processing power required for laregscale real AI tasks, but assembling the 'logic as data' than makes up the complexity of solving broad problems is the real limitation.
The decision logic has to be formed and corrected by hand and any generalized AI has billions of interrelations that need to be built up and that for partial problem domains.
Watson only identified concepts by secondary descriptions and that is only a small part of actually solving problems from broad categories.
Solutions come from human thought and extracting those billions of possible solutions and getting them into algorithmic and data form is the real task.
The whole world is already run by AI systems, it's just not self-aware AI.
And to actually get a self-aware AI similar to humans you'd would need to bake in drives (no, not the storage kind) and wants and mental limitations and even fears/worries I think, since that's what it is all about for lifeform.
And strangely I didn't hear that mentioned much in the things I hear about AI research, but maybe that's the fault of journalists rather than that it's really not researched.