WHAT ARTIFICIAL INTELLIGENCE (AI) IS AND IS NOT

EXAMPLES Example 1: The thermostat versus the chatbot. Imagine a thermostat that follows the rule "if the temperature is below 68 degrees, turn on the heat." A programmer wrote that rule, and the thermostat does exactly that and nothing more. It never improves, and it never meets a situation it was not told about. This is a rule-based system, and it is not AI in the modern sense. Now imagine a chatbot that learned the patterns of language by reading billions of documents. Nobody wrote its answers. It builds each reply from patterns it absorbed. That is narrow AI, specifically a language model. Walking through the difference, the thermostat is predictable but limited, while the chatbot is flexible but sometimes wrong. Knowing which one a task needs is the first judgment a harness engineer makes. Example 2: Two generations of chess programs. Chess programs from the 1990s mostly relied on hand-coded rules and a method of searching through many possible moves. Programmers wrote how to value a pawn, a knight, or a safe king, and the computer searched deeply using those hand-written valuations. This is narrow, symbolic AI. Later chess programs began to use deep learning and self-play. Self-play means the program plays enormous numbers of games against itself and learns from the results. These modern engines learn which positions are good without being told, and they sometimes find strategies that human players had never considered. The shift illustrates the move from hand-written rules to learned strategy. Both kinds of program are narrow, since neither can do anything beyond games. Example 3: A coding agent in 2026. Picture a coding agent that receives a bug report saying "the checkout button crashes when the cart is empty." The agent reads the report, searches the codebase for the checkout code, writes a fix, runs the project's automated tests, notices that one test fails, adjusts its fix, runs the tests again, and finally opens a pull request. A pull request is a formal proposal, within a code-sharing system, that asks human teammates to review and merge a change. This is a good example of narrow AI operating as an agent. It is not general intelligence, because it only works on software tasks and can be fooled in ways a human colleague would not be. Yet it takes many steps by itself, uses tools, and responds to results, which is what makes it an agent. A harness engineer would be responsible for the loop, the tool permissions, the test runner connection, and the limit on how many steps the agent may take. Example 4: Self-driving perception. A self-driving car uses camera and sensor data to recognize stop signs, pedestrians, and lane markings. That recognition is done by trained neural networks, and it is a very good example of narrow AI, because the networks are expert at perceiving road scenes but cannot write a poem or plan a vacation. As of this book's preparation, no deployed system anywhere is general AI. Keeping that fact in mind protects you from both exaggerated fears and exaggerated promises.