Is AI not really a thing ?

I agree with both of the experts quoted. AI is advanced pattern recognition, not yet intelligence or understanding. It's genuinely very useful for searching large data sets that would otherwise be difficult - make no mistake about that - but no threat at all in terms of killer robot driving instructors and the like.

In radio astronomy we have search algorithms that are useful additions to manual techniques. But they are not really automatic as they are often erroneously described - they are at best semi-automatic, because a person has to check their results. Their catalogues can be either complete (containing most of the real sources) or reliable (most of their catalogues consist of real detections and not wrong identifications of noise). You can't really have both, and neither are perfect. But this is still extremely useful because it gives you another, faster look at the data in a way which has different biases to that of people. Oh, it's objective and repeatable, certainly, but it's still biased.

Pattern recognition is great for some things, but it isn't the whole story. Some signals can look exactly like galaxies and it's impossible to distinguish them from real galaxies without other observations. Algorithms that claim to spot personality and other social traits (e.g. sexuality, criminal intent) actually just spot expected observable societal norms for certain traits, not the traits themselves. Hence they're easy to fool. They may be considerably more useful where features are more objective and less variable, but they will still be more like assistants (of varying levels of training) than expert replacements.

Mr Chishti is not some contrary pundit, but a computer scientist who has built two multi-billion dollar businesses, the latest devoted to using AI. He accepts that machine learning, where computers are given vast amounts of data and learn to spot patterns in it, has advanced as processing power has leaped forward. But he seems dubious that these techniques have much to do with the kind of AI we have been warned about by the likes of Ray Kurzweil and Elon Musk, where machines surpass humans and make them redundant.

"Really there's nothing here that's tremendously groundbreaking - the same principles that were at work even 40 years ago are still at work today," he says.

Artificial General Intelligence, he says, would need computers to acquire human characteristics such as understanding meaning and context, or even becoming conscious. "Nobody serious thinks we've made any progress on those fronts," he says. "There's not one person in the field that has any view that we're any closer to generalised intelligence than we were two decades ago."

Kriti Sharma, who oversees artificial intelligence development for the business software firm Sage, tells us we should not get too hung up on terminology when we talk about AI: "Machine learning, predictive analytics, AI - in reality it's about the problems that AI can solve for humanity."

She says even the smallest businesses are now finding that AI techniques can save them money and make them more productive. As a counter to Mr Chishti's cynical take on DeepMind's Go victory as a marketing exercise, she points to the same company's project to automate the diagnosis of eye scans, potentially bringing more rapid treatment to millions of patients.


She thinks we need to focus on short-term gains rather than dreams of sentient machines: "We're too fixated on creating AI that is as good as humans - in reality we need something that can support and make human lives better," she says.


http://www.bbc.co.uk/news/technology-45219902