Defining artificial intelligence isn’t just difficult; it’s impossible, not the least because we don’t really understand human intelligence. Paradoxically, advances in AI will help more to define what human intelligence isn’t than what artificial intelligence is.
But whatever AI is, we’ve clearly made a lot of progress in the past few years, in areas ranging from computer vision to game playing. AI is making the transition from a research topic to the early stages of enterprise adoption. Companies such as Google and Facebook have placed huge bets on AI and are already using it in their products. But Google and Facebook are only the beginning: over the next decade, we’ll see AI steadily creep into one product after another. We’ll be communicating with bots, rather than scripted robo-dialers, and not realizing that they aren’t human. We’ll be relying on cars to plan routes and respond to road hazards. It’s a good bet that in the next decades, some features of AI will be incorporated into every application that we touch and that we won’t be able to do anything without touching an application.
Given that our future will inevitably be tied up with AI, it’s imperative that we ask: Where are we now? What is the state of AI? And where are we heading?
Descriptions of AI span several axes: strength (how intelligent is it?), breadth (does it solve a narrowly defined problem, or is it general?), training (how does it learn?), capabilities (what kinds of problems are we asking it to solve?), and autonomy (are AIs assistive technologies, or do they act on their own?). Each of these axes is a spectrum, and each point in this many-dimensional space represents a different way of understanding the goals and capabilities of an AI system. On the strength axis, it’s very easy to look at the results of the last 20 years and realize that we’ve made some extremely powerful programs. Deep Blue beat Garry Kasparov in chess; Watson beat the best Jeopardy champions of all time; AlphaGo beat Lee Sedol, arguably the world’s best Go player. But all of these successes are limited. Deep Blue, Watson, and AlphaGo were all highly specialized, single-purpose machines that did one thing extremely well. Deep Blue and Watson can’t play Go, and AlphaGo can’t play chess or Jeopardy, even on a basic level. Their intelligence is very narrow, and can’t be generalized. A lot of work has gone into using Watson for applications such as medical diagnosis, but it’s still fundamentally a question-and-answer machine that must be tuned for a specific domain. Deep Blue has a lot of specialized knowledge about chess strategy and an encyclopedic knowledge of openings. AlphaGo was built with a more general architecture, but a lot of hand-crafted knowledge still made its way into the code. I don’t mean to trivialize or undervalue their accomplishments, but it’s important to realize what they haven’t done. We haven’t yet created an artificial general intelligence that can solve a multiplicity of different kinds of problems. We still don’t have a machine that can listen to recordings of humans for a year or two, and start speaking. While AlphaGo “learned” to play Go by analyzing thousands of games, and then playing thousands more against itself, the same software couldn’t be used to master chess. The same general approach? Probably. But our best current efforts are far from a general intelligence that is flexible enough to learn without supervision, or flexible enough to choose what it wants to learn, whether that’s playing board games or designing PC boards.
How do we get from narrow, domain-specific intelligence to more general intelligence? By “general intelligence,” we don’t necessarily mean human intelligence; but we do want machines that can solve different kinds of problems without being programmed with domain-specific knowledge. We want machines that can make human judgments and decisions.
