Hal 9000, the murderous computer of "2001: A Space Odyssey", defined the public's perception of Artificial Intelligence for more than three decades. Besides singing children's songs and killing spaceship crews, it was described as capable of performing a wide variety of highly sophisticated tasks. Its birth was projected for January 12, 1997, and I suppose that back in the late sixties people assumed that the dawn of the era of intelligent machines was just around the corner. Even the great Alan Turing predicted in 1950 that in about fifty years "an average interrogator will not have more than a 70% chance of making the right identification after five minutes of questioning" in an imitation game. The famous Turing test assumes that a machine is intelligent if it can fool human observers into thinking they are communicating with another human being. The restricted version of this test, the Loebner Prize>, limits the conversation domain to several machine-manageable topics, but until now no computer was mistaken for a human. Despite many attempts to build greater intelligence into these chatterbots - including the ability to learn on their own - they remain rather "stupid".
Now a group of researchers claims a breakthrough with a computer that has learnt to talk like a toddler. Artificial Intelligence NV (Ai), an international company split between Boston and Tel Aviv, launched the first commercially oriented research and development effort, with complete financial backing, that is committed to the creation of "true" artificial intelligence. I don't know if it can really qualify as first project of its kind, as similar large-scale development efforts exist for decades. Cyc, for example, is an immense, broad, multi-contextual knowledge base developed over the last fifteen years. They share similar goals: to enable a multitude of knowledge-intensive products and services which will hopefully revolutionize the way in which people use and interact with computers. However, what distinguishes Ai from the rest of the crowd is their unconventional approach to creating a truly conversant computer. They recently passed the first important milestone when Hal successfully finished an adapted Turing Test for 15 month olds. A child development specialist was given transcripts of Hal's conversations, and declared him a healthy, normal little boy. What she wasn't told is that Hal is a computer program running on a regular Windows PC.
Hal quickly went through a babbling stage, producing only word-like utterances. It reportedly talks like an eighteen months old child right now. But how does it really interact with its trainers? See this example from the 18-months language development sample:
Trainer: where is daddy?
Hal: daddy gone
Trainer: daddy is at work
Hal: daddy work bye bye
Trainer: yes daddy has gone to work
Hal: daddy work bye bye bye bye
Trainer: do you want to go to the park?
Hal: yes bye bye
Trainer: ok we can go to the park
Hal: take car bye bye
...
The philosophy behind the whole project borrows from Bayesian statistical analysis, predicting what output will be rewarded based on the computer's prior experience. Contrary to the traditional approach of using a set of hardwired rules and a vocabulary database to approximate human conversation, researchers at Ai follow the Turing's idea of the child machine. This behaviorist approach involves just a simple set of learning algorithms and develops speech by talking to humans. It focuses on reinforced training: language is regarded as a skill that is similar to any other behavior, so reinforcement, imitation and successive approximations are required to mature performance.
Jason Hutchens, one of the leading chatterbot developers and winner of the Loebner prize back in 1996 recently joined the Ai team as a chief scientist. According to his team, Hal should reach adult-level linguistic capability in approximately ten years (it should talk like a three-year old child in about 30 months). In the meantime, they will release a preliminary conversational agent code-named "Virtual Pal" to showcase the technology and enable users to operate their machines just by talking in natural language. The release date is set for the end of this year - stay tuned for more news and developments.
