ARTIFICIAL INTELLIGENCE
Once again, all the action behind this year's Loebner Prize - the annual search for a computer smart enough to win the Turing test - centered around second place. The contest for the $100,000 top award, which is given to the machine that meets Turing's criteria of duping event judges with both text and audio, was too hard. That left the $2,000 consolation prize, awarded to the "most human" computer, no matter how unconvincing it was.
After 11 years without a top winner, it's no wonder some observers are getting annoyed at the whole exercise. "The Turing test is a red herring," says AI pioneer Douglas Lenat, founder and president of Cycorp, who gave the keynote address at last year's awards. "Anthropomorphizing a computer program isn't a useful goal." The former Stanford prof is one of a handful of contentious brainiacs vying to create a program with a corpus of common sense.
Right now, all efforts point to the Web, where programmers hope to unleash bots that can troll for information, learning the ticks of language and meaning along the way. In the end, it probably won't matter if the voice sounds human - so long as the intelligent, natural language-speaking automatons manning shopping sites, help desks, call centers, and emergency hot lines tell us what we need to know. Here are some of the top AI projects in progress.
Project: CYC (1984 to present)
Creator: Cycorp (Austin, Texas; www.cyc.com)
Profile: Spock without the attitude; an inference engine plus database that contains 1.5 million "true" common sense statements (e.g., "Water is wet")
Philosophy: Intelligence/language as a construct
Method: Logical: Build a brain one connection at a time
Human Brainpower: 70 employees
Bankroll: $50 million, primarily from the military
Project: Generic Artificial Consciousness, or GAC (1994 to present)
Creator: MindPixel Modeling Project (Atacama Desert, Chile; www.mindpixel.com)
Profile: A scan of Netheads' shared reality; database with more than 415,000 statements tested on 20 people selected at random; each statement has a probability of being either true or false (e.g., "Swimming with ski pants is difficult" is 79 percent true, 21 percent false)
Philosophy: Intelligence/language as a statistical process
Method: Computational: Train a neural net to mimic a human using the GAC (pronounced "jack") database as a Dawkinsian fitness test
Human Brainpower: 1 project leader and about 43,000 online participants
Bankroll: More than $10,000 out of pocket
Project: Common Sense (1999 to present)
Creator: Open Mind Common Sense (Cambridge, Massachusetts; www.openmind.org)
Profile: Database of one-liners; open source database containing 362,000-plus unverified statements (e.g., "The highest point in China is Mount Everest
at 8,848 meters") Philosophy: Intelligence/language as data
Method: Cultural: Compile an open source database of common sense statements for use by AI researchers/companies
Human Brainpower: Several researchers and more than 7,300 online participants
Bankroll : More than $10,000 out of pocket
Project: HAL (2000 to present)
Creator: Artificial Intelligence NV (Tel Aviv; www.a-i.com)
Profile: B. F. Skinner's toddler in a box; stochastic-learning algorithm: HAL predicts which output sentence is most likely to be rewarded based on its past experience - it currently communicates in two- to four-word sentences similar to those used by an 18-month-old (e.g., "Where is daddy?")
Philosophy: Language as learned behavior (like tango dancing)
Method: Top-down: Train a learning program to respond to textual stimuli by rewarding "good" utterances
Human Brainpower: 40 employees
Bankroll : A few mil in angel money
MUST READ
Electronic Paper Pushers
Dial M for Mobile MPEG
Don't Throw Away That Disk!
Olympics Rapid Response System
Sticky Business
People
Jargon Watch
America's New-Venture Capitals
Lucky Star Accommodations
Battle of the Brains
Build Your Own Franken-PDA
Raw Data
