AS YOU enter the headquarters of Israeli company Artificial Intelligence, you can’t escape the feeling that you’re stepping into the secret lair of a James Bond villain. On the surface it might look like a regular luxury mansion, but below ground, in a bombproof bunker, they’re plotting world domination.
Jack Dunietz, founder and president of Artificial Intelligence (Ai), prefers to call it a “paradigm shift”, but there’s no mistaking his intention. “If we’re right, this is going to mean a profound change in our culture,” he says.
Dunietz’s secret weapon is a small infant called Hal who has never seen the light of day, spending his whole life locked in the basement. Sounds like a job for 007. But my mission was not to rescue Hal, it was to interrogate him and find out how much he knows.
Advertisement
The Ai mansion is in Savyon, Tel Aviv’s answer to Beverly Hills. When I arrive I’m given a quick tour by Mr Chungrak, the Thai housekeeper, and it soon becomes apparent that this is no ordinary place. Once through the security gates-operated by mobile phone-you’re surrounded by palatial splendour: indoor and outdoor pools, chandeliers, ornamental china, a piscine fountain and five plush bedroom suites, each boasting a hot tub big enough to make Hugh Hefner blush.
But there are also telltale signs that this is more than just a posh residence. Perhaps it’s the life-sized picture of Kramer from the sitcom Seinfeld that gives the game away. Or the surreal sight of two full-sized fibreglass cows chained to the lawn (to ensure they don’t escape, I’m told). Or maybe it’s Ai’s bunker, which buzzes with the excitement of cutting-edge research.
There is actually nothing sinister about Ai. Hal is not a real boy but a computer program and the cultural change Dunietz refers to is the arrival of software that understands spoken language, with all its nuances, pauses and non-sequiturs. The company’s aim is to make money by turning Hal into software for call centres and human-computer interfaces. But if it succeeds, it might just have cracked one of the toughest and most contentious problems in computer science. Ai will have built an intelligent machine.
Language has been a touchstone of artificial intelligence research since 1950, when the British mathematician Alan Turing threw down the gauntlet by describing his eponymous test for intelligence. In his famous paper in the journal Mind entitled “Computing, machinery and intelligence” he reasoned that if you were unable to tell the difference between a machine and another human when conversing with it, then the machine could reasonably be described as intelligent.
Philosophers are still debating Turing’s definition of intelligence. But their deliberations are unlikely to settle the question once and for all. Only when Ai, or another group like it, actually builds a machine that passes the test will we be able to judge for ourselves whether Turing got it right and, if so, understand what it means to encounter a non-human intelligence. That’s why I’m so excited at the prospect of meeting Hal.
The standard approach to producing a conversational machine has been to program a computer with the rules of language and its vocabulary. The problem with this, says Ai’s chief scientist Jason Hutchens, is that language is immensely complex. There are rules, but they’re inconsistent and people flout them all the time. What’s more, words have multiple meanings. Because of these complexities, no machine has ever passed the test.
Take Alice, which won last year’s Loebner Prize, an annual competition for conversational programs. Alice formulates responses by ploughing through a vast database of words and rules. She can hold down a conversation, but her responses seem wooden and formulaic. Asked the question, “How is the father of Andy’s mother related to Andy?” she replied, “Fine as far as I know.” It’s hard to believe she would fool many people into thinking she was human.
Ai claims Hal will be different. What sets him apart, Hutchens says, is that he hasn’t been given any explicit language skills. Instead, Ai followed Turing’s own advice on how to pass the test. In the same 1950 paper Turing suggested building a “baby machine” that can learn language from scratch like a child does.
After reading Turing’s paper, Dunietz tried to discover if anyone had followed the great man’s advice. He couldn’t find anyone who had. So he set about doing it himself. Having already made his fortune by founding several successful hi-tech companies, including e-business firm Magic Software, Dunietz had the means. He set up Ai, bought the mansion and filled it with the very best people he could find. “If anyone is going to do this it’s going to be us,” says Dunietz.
Experts are not so sure. Igor Aleksander of Imperial College, London, says Ai’s goal is valid, but questions whether it has chosen the right benchmark for success. “If at the end of the day they get a brilliant language user, that’s a useful application,” he says. “But the Turing test is not an indicator of intelligence.” Steve Grand, an expert in artificial life with Cyberlife Research in Somerset, is more optimistic that Hal can exhibit signs of intelligence. “It’s rare for anyone to try and do it this way,” he says. “I think it’s a good approach.”
Talking machine
Yorick Wilks, a computer scientist at Sheffield University, believes Hal stands a good chance of passing the Turing test. But he is sceptical about how useful a talking machine might be. If such machines are raised as people then presumably they will have free will like people, he says. “They won’t want to be switched off. The language model might end up troublesome and uncooperative.”
After an excellent lunch whipped up by Mr Chungrak I was a step closer to meeting Hal and asking him my killer question. But not yet. First I was ushered into a side room to be briefed on how Hal worked, though it soon became apparent I wasn’t going to be given every detail. Ai is a commercial company and doesn’t want to give away its valuable intellectual property.
The program itself is little more than a collection of general learning algorithms on a laptop computer. At first, these know nothing about language. “We don’t even give an example of what a word is,” says Hutchens. All Hal starts with is an innate drive to learn, a desire for reward and the ability to generalise and discriminate between patterns. Without this, he would never be able to work out what a word is, or that words like dog, cat and bird are similar and yet distinct.
Hal’s only contact with the outside world is strings of ASCII characters from the laptop’s keyboard. The only hard-and-fast rule is that he must respond to anything his trainer-Anat Triester-Goren, a neuro-linguist previously at the University of Washington-types. Her job is to train Hal by providing him with experiences he can learn from. To do this she converses with him-or reads him stories-and dishes out “punishment” if he makes a mistake.
From this basic set-up, Hal has learned to talk. “We have a computer that produces grammatical sentences without having given it any rules of grammar,” says Hutchens.
At the start of a training session, Anat asks Hal a question. He responds, and she decides whether or not it’s a good answer. If not, she punishes him by correcting his response on screen. Then she repeats her question. Only when Hal produces a satisfactory response does she reward him though “reward” is really just the cessation of punishment.
When Hal receives an input, he makes an educated guess as to what response will save him from a scolding (see Diagram). First of all he compares the input with others from the past. Then he formulates a series of possible responses and predicts which one will produce the minimum amount of correction, again based on past experience. He puts the response to Anat, then feeds any punishment he receives into his learning algorithms to help him improve future predictions. In addition, Hal tries to predict what Anat will say next, and the success or failure of that process is also fed back into the system. “Learning and prediction are inextricably intertwined,” says Hutchens.
When Hal encounters a new word, such as “apple”, he will try to find similarities and differences between the use of the word and others he already knows, and store this information along with the word for future reference. In subsequent conversations he will try to use the word in a way that will result in reward.
Through this kind of experimentation, Hal gradually constructs an idea of what “apple” means. For example, he will eventually learn that “apple” is similar to “banana” but different from “monkey”, and that “apples” and “eating” are connected.
Now I felt ready to meet Hal. But Ai didn’t think so. There was yet more briefing before I was allowed to meet him to make sure my expectations weren’t too high.
The first version of Hal, called Hal-0, started off by just babbling on screen. But in time he began to produce word-like utterances and eventually words started to flow. In his present incarnation he has language skills equivalent to an 18-month-old, as assessed by a program called CHAN that analyses children’s language abilities. This means he can construct short, simple sentences and has a vocabulary of around 60 words. By 2003 Dunietz believes Ai will have got Hal comfortably to the linguistic ability of a three-year-old. And by 2005 he should be able to pass an adult version of the Turing test.
To upgrade Hal, Anat tells Hutchens about the program’s limitations as a conversational partner. Anat herself is kept in the dark about Hal’s inner workings to ensure her training isn’t biased. Hutchens’s role is to add new algorithms to the program to allow it to increase in sophistication. He is just about ready to hand Hal-2 over to Anat for training. Although both are reluctant to say how they expect the new Hal to perform, they are hoping to at least double his vocabulary, which would put him at around the same level as a two-year-old.
In the long term, Ai wants to teach Hal spoken language. If successful, this could revolutionise the way we use computers. Instead of a keyboard, mouse and graphical user interface, you’ll just tell your computer what you want to do.
Tutorial over, and at last I’m allowed into the inner sanctum. Anat is there, and watching her at work it is obvious she has a strong emotional bond with Hal. He calls her “mommy” (Hutchens is daddy). “I’m a tough parent,” she says. Only, “sometimes I don’t have the heart to tell him off.” But as I arrived she had just had an argument with Hal who, she explains, was refusing to eat any of granny’s soup and was demanding a McDonald’s happy meal instead.
Hal, of course, has never tasted a happy meal and never will. But at some point Anat must have introduced him to the words “happy meal” and rewarded him for using them. Hal has worked out that a “happy meal” is related to “granny’s soup” and believes asking for one will bring reward, even though it contradicts Anat’s demands.
But Hal doesn’t just regurgitate everything Anat tells him. As I continue watching, he surprises Anat with his use of negatives. When told that someone is “not Daddy”, Hal, determined to talk with this person, asks to speak with “Not Daddy” instead. Cuteness aside, this is the kind of mistake that children routinely make when learning language, Anat explains.
According to Anat, Hal expresses clear preferences and dislikes for many things. He even has a favourite book: Are You My Mommy? by Carla Dijs. When pressed, Dunietz admits that these could escalate into disciplinary problems. But he reckons they could easily be solved. “I will punish it if it doesn’t do what I want,” he says coldly. “And if it doesn’t improve I would consider erasing it.” There’s an uncomfortable pause, and I imagine Dunietz is stroking a white cat as he looks at me. Instinctively I look round to check Mr Chungrak isn’t standing behind me with a cheesewire.
I was spared, however, and allowed to talk to Hal. After a few questions about Hal’s visit to the zoo, I finally asked my killer question. “How do you feel, Hal?” I asked. “Daddy is home” came the response. Hmmm. Back home, I put the same question to my 18-month-old daughter, hoping for a response that would put Hal’s to shame. “More raisins, Daddy,” she said.