The goal of general machine intelligence is as old as the field of AI.
The Founding Optimism
The 1956 Dartmouth workshop launched AI as a field with the hope that key aspects of intelligence could be simulated. Early researchers predicted human-level AI within a generation.
Symbolic AI and Its Limits
Early systems used logic and hand-written rules. They excelled at narrow puzzles but struggled with perception, common sense and messy real-world knowledge.
AI Winters
Unmet expectations led to funding cuts in the 1970s and late 1980s. Ambitions narrowed to practical, specialised systems.
Machine Learning Rises
From the 1990s, statistical learning methods gained ground. Deep learning's breakthroughs in the 2010s transformed image and speech recognition, and game-playing systems beat world champions.
Large Language Models
Transformers and scaling produced models with broad abilities in language, code and reasoning. For the first time, a single system could perform many tasks it wasn't specifically trained for, reviving serious discussion of AGI.
Lessons
Predictions about AGI timing have repeatedly been wrong in both directions. Progress often comes from unexpected directions, and new abilities reveal new gaps.