The Transformative Breakthroughs of Early AI

Key Insight

The realization of Artificial General Intelligence (AGI) began in 2012 with DeepMind's Deep Q-Network (DQN) algorithm, which aimed to create general learning agents surpassing human cognitive performance. This self-learning system was trained to play classic Atari games, such as Breakout, by receiving only raw pixels and the score, then learning paddle control through trial and error. Initially, DQN performed poorly, but over time, it developed basic game-playing skills.

A pivotal moment occurred when DQN autonomously discovered a sophisticated 'tunneling' strategy in Breakout, which involved targeting a single column of bricks to create a path for the ball to bounce off the back wall, systematically clearing the screen for maximum score with minimum effort. This tactic, known to serious gamers but not immediately obvious, demonstrated the AI's ability to learn valuable, novel knowledge without explicit programming. This breakthrough showed that an AI agent could discover new, seemingly superhuman insights.

Within a few months of further development, DQN reached superhuman performance levels, validating the core mission. This success was followed by AlphaGo in 2016, an AI designed for the ancient game of Go, which is exponentially more complex than chess with 10^170 possible board configurations. AlphaGo initially learned from 150000 human expert games, then vastly improved by playing against itself millions of times. It famously defeated world champion Lee Sedol 4–1, with its 'move 37' in the second game being a seemingly irrational play that ultimately proved decisive, rewriting millennia-old Go strategy and demonstrating AI's capacity to uncover unprecedented ideas. Later, AlphaZero learned from scratch without human input, surpassing the original AlphaGo's performance within a day.