Crazy singularities, robot rebellions, falling in love with computers … artificial intelligence conjures up a multitude of wild what-ifs. But in the real world, AI involves machine learning, deep learning, and many other programmable capabilities that we’re just beginning to explore. Let’s put the fantasy stuff on hold — at least for now — and talk about this real-world AI. Here’s how it works, and where it’s going.
What is artificial intelligence?
Today’s AI systems seek to process or respond to data in human-like ways. It’s a broad definition, but it needs to be as broad as possible, because there are a lot of different AI projects currently in existence. If you want a little more classification, there are two types of AI to consider.
- General: The purpose of general AI is to mimic human behavior as much as possible. Developers actually care about the Turing Test, and the goal is to make a system as life-like as possible. That also makes it less useful, however, and often unable to specialize. General AI systems are good for showy demonstrations and sales — Siri and Cortana are prime examples — but they are inherently limited because of their interaction requirements.
- Narrow: Narrow AI is focused on a specific problem or situation, and designed to analyze data and form conclusions far more efficiently than humans can. An automatic translator that converts Spanish to English can be considered a type of narrow AI, or software that analyzes stock options and recommends investment ideas. These aren’t very flashy and are usually confined to simple interfaces, but they are far more useful in a practical sense.
AI can also be classified by how it operates, which is particularly important when considering how complex an AI system is and the ultimate costs of that software. If a company is creating an AI solution, the first question must be, “Will it learn through training or inference?”
- Training: These AIs are designed to learn and improve over time, and will adjust their data sets and certain parts of their processes to become more efficient. This takes a lot of processing power, so most training features in commercial AI are very simple.
- Inference: These AIs are designed to look at data and draw conclusions in careful steps. For a casual example, an AI might infer, “To answer this question, data for yesterday’s game scores must be found; searching list of reliable sports data sets; comparing to favorite teams listed in settings; reporting scores in audio.” But they have no or little ability to change by themselves over time. This takes far less processing power (and lower costs).
There have been books and books written about what specific features AI must include to be truly AI, and unsurprisingly, no one really agrees on what these features are; every description of AI is a little different. But there are several examples of successful AIs in our current landscape worth looking at.