Early AI mastered logic puzzles but stumbled over everyday common sense
The first researchers in artificial intelligence taught machines to reason step by step, the way people work through a puzzle. It turned out humans rarely think that way. Most of our decisions come from quick intuition and a vast store of unspoken everyday knowledge, and capturing that proved far harder than solving logic problems.
Artificial intelligence became an academic field in 1956. Its aim is to build systems that learn, reason, plan, understand language and perceive the world well enough to pursue goals. The history has swung between bursts of optimism and stretches of disappointment and vanishing funding, periods nicknamed AI winters.
Step-by-step reasoning programs hit a wall called combinatorial explosion: as a problem grows, the possibilities multiply so quickly that the program slows exponentially or never finishes. Rules of thumb known as heuristics help by steering the search toward promising paths. By the late 1980s and 1990s, researchers were borrowing ideas from probability and economics to cope with uncertain or incomplete information. An agent might score every situation by how much it prefers it, then pick the action with the best expected payoff.
Knowledge caused its own trouble. Symbol-based systems stored facts explicitly, but the set of things an ordinary person knows is enormous, and much of it is never put into words at all. Early language programs built on Noam Chomsky's grammar could handle word meanings only inside tiny artificial micro-worlds. The linguist Margaret Masterman argued that meaning, not grammar, was the real key, and that thesauri should underpin computer language.
A different route took over after 2012, when graphics processors began speeding up neural networks and deep learning beat older methods; the transformer design from 2017 pushed things further. Large language models absorb knowledge from millions of books and billions of web pages rather than hand-written rules, and AlphaZero learned game strategy by playing itself. That approach sidesteps the common sense bottleneck, yet it still struggles with accurate recall and sound reasoning, and it can produce confident errors known as hallucinations.
Source: Artificial intelligence
