How the field developed.
THE BEGINNING
The field was named in the 1950s, with early optimism that human-level intelligence was decades away.
That optimism proved substantially premature.
EARLY APPROACHES
Symbolic AI: encoding knowledge as rules and logic.
Achieved narrow successes and failed to scale to messy real-world problems.
THE AI WINTERS
Periods when funding and interest collapsed after expectations were not met.
Twice, notably in the 1970s and late 1980s.
Worth remembering when assessing current claims.
EXPERT SYSTEMS
In the 1980s, rule-based systems encoding specialist knowledge.
Commercially useful in narrow domains, and brittle and expensive to maintain.
THE STATISTICAL TURN
From the 1990s, learning from data rather than encoding rules.
Steadily more successful as data and computation grew.
THE DEEP LEARNING ERA
From around 2012, neural networks with many layers transformed image and speech recognition.
THE LANGUAGE MODEL ERA
From the late 2010s, transformer models produced the language capabilities now widely used.
WHAT THE HISTORY SUGGESTS
Progress is real, uneven, and consistently slower than predicted.
