Part 1 of "How We Got Here," Chapter 1 of the The Machine Intelligence Primer
The first challenge to understanding Machine Intelligence (MI) is defining it. A simple web search for a definition will return enough conflicting explanations and bad illustrations of robots to make you regret trying in the first place. In this section, we’ll help you gain a realistic understanding of what MI is—and is not.
The best place to start is by understanding what academics, consultants, and internet savants can agree on: the “why” and the “how” of MI. First, the why. The purpose of MI is to produce machines able to perform tasks that would normally require human intelligence. Put simply, MI researchers want to build machines that can perform tasks we do—ideally better than we do them. Next, the how. MI works by enabling machines to perceive, learn from, abstract, and act on data (machine learning). MI and machine learning are often conflated, but machine learning is a technique for achieving MI—not the entire field. Deep learning, for its part, is a machine learning technique. (Fig. 1.1) These techniques are described in greater depth in the next section, A Brief History of MI.
MI and artificial intelligence (AI) are interchangeable terms. MI has grown popular because it lacks negative connotations sometimes associated with AI. Companies don’t want you thinking about The Terminator when you’re deciding whether to buy a “virtual” home assistant. More importantly, the “artificial” in AI suggests machines possess a fake form of human intelligence, the way a good reproduction of a painting could be confused with the original. In fact, intelligent machines’ smarts are much different from human smarts. While our know-how comes from evolution, experiences, culture, and more, machines’ intelligence comes mostly from math. Making this distinction in terminology helps remind us how machines differ from us.
Today’s MI is not anything like the killer robots we see in movies or on TV. In fact, the cardinal limitation of MI is that—at least for now—it can only perform narrowly defined tasks well. That’s why today’s MI is referred to as Machine Narrow Intelligence. Machine General Intelligence (MGI), in which machines would match humans’ capacity to perform many tasks (think, C-3P0 from Star Wars) does not yet exist. Despite recent acceleration in MI development, most researchers agree there is no clear path to creating MGI today. Therefore, claims that MI will soon “replace” us should be viewed with healthy skepticism. We may find our jobs boring at times, but few of us perform just a single task (and thank goodness for that).
Another frequent point of confusion is the relationship between MI and data science. Both fields are rooted in computer science and mathematics, and both are concerned with finding patterns in data to produce action- able insights. The critical difference between the two is that MI is concerned with enabling machines to autonomously perceive, reason, and act, while data science is the human tradecraft of producing insights from data that enable us to act. Online store managers might use data science to review click rates and shipping data, identifying opportunities to create efficiencies. Or they might use MI to parse other vendors’ pricing data and automatically adjust their own store’s pricing in real time. The magic of MI is in ceding some work to machines to avoid tedious tasks and allow us to focus on higher-order work.
Introduction
How We Got Here, Part 2:
Understand the History of MI
The Machine Intelligence Primer provides a foundational understanding of where MI came from, how it got to where it is today, and where it’s likely going. It explains and dispels common myths that surround MI. It is meant to help executives, practitioners, and curious skeptics alike consider what MI will mean for them and their teams, and create a world that is more efficient, equitable, meaningful, and verdant.
