Creating systems that can be used for a variety of problems could take decades
Artificial intelligence is having a moment.
Startups that claim to be using AI are attracting record levels of investment. Big tech companies are going all-in, draining universities of entire departments. Nearly 140 AI companies have been acquired since 2011, including 40 this year alone.
AI is showing up in our everyday lives, as voice-recognition technology in our devices and image recognition in our Facebookand Google accounts.
Now, Google parent Alphabet Inc., Amazon.com Inc. and MicrosoftCorp. are making some of their smarts available to other businesses, on a for-hire basis. Want to make your app or gadget respond to voice commands, and answer in its own “voice?” These services can do that. Need to transcribe those conversations so they can be analyzed? This new breed of services can do this and many other things, from face recognition to identifying objectionable content in images.
But wringing measurable utility from these new AI toys can be hard. “Everyone wants to think the AI spring is going to blossom into the AI summer, but I think it’s 10 years away,” says Angela Bassa, head of the data-science team at energy-intelligence-software companyEnerNOC Inc.
Before switching to her new role, Ms. Bassa led a team at EnerNOC that used AI techniques such as machine learning and deep learning, which feed massive amounts of data into computer programs to “train” them. But the company found that customers were more interested in analytics than in the incremental value that sophisticated AI-powered algorithms could provide.
AI, says Ms. Bassa, requires three things that most companies don’t have in sufficient quantities. The first is enough data. Companies like Facebook, Amazon, Alphabet, General Electric Co. and others are harvesting enormous amounts of data, but they are exceptions.
The second is problems where making a small difference can justify the expense of creating an AI system. If AI can improve the fraud-detection system at a credit-card company by 1%, that could be worth tens of millions of dollars. For a midsize manufacturer that makes many different products, however, a 1% improvement in productivity of a particular line might not justify the cost of hiring a half-dozen highly paid engineers.
That leads to the third scarcity: People to build systems. The war for AI talent is driving up the cost. “There are maybe 5,000 people in the world who can put together one of these machine-learning systems in a way that saves money, even if only incrementally,” says Ms. Bassa.
This doesn’t mean that AI is useless to businesses. But it does suggest that AI is being oversold. Creating systems that can be used for a variety of problems, and not just the narrow applications to which AI has been put so far, could take decades. Systems have to be built and trained. Like educating a child, this takes time.
Most of what’s available now are “pre-trained” systems, built by companies like Microsoft, Amazon and Google, and reflecting the data those companies have. Those companies have billions of images, so they offer commercial image-recognition services for others. Amazon, having compiled a vast trove of spoken language from its Alexa personal assistant, offers services to process spoken language—and generate replies—for others.
Some startups are beginning to offer broader AI systems that require neither a machine-learning expert nor a pre-trained system constructed by the likes of Google. Israel’s n-Join sells manufacturers a small box that collects data from machines on an assembly line, and then uses machine learning to spot aberrations that could presage a breakdown.
The key to n-Join’s utility, says Guy Tsur, a senior technologist at Strauss Group Ltd., one of Israel’s largest manufacturers of dairy products and an early n-Join customer, is that it doesn’t have to know the type of assembly line it’s attached to, or what the sensors feeding it data are measuring. It’s simply looking for correlations that indicate the manufacturing process is operating differently than usual. It then alerts its human supervisors, who can use their own experience and judgment to diagnose a specific problem.
One thing an astute reader has noted by now is that none of these triumphs and shortcomings of AI resemble the sci-fi visions of machines taking over the world. Reflecting on my own brief experience as an invertebrate neuroscientist, I’d say that today’s AI is at the jellyfish stage in the evolution of biological intelligence. Real brains—and genuine intelligence—are so far in the future as to be beyond any reasonable horizon of prediction.
Or, as chief scientist and AI guru Andrew Ng of Chinese search giantBaidu Inc. once put it, worrying about takeover by some kind of intelligent, autonomous, evil AI is about as rational as worrying about overpopulation on Mars.