What surprised PG about AI
AI developed opposite to expectations: starting human-like and improving toward precision, not vice versa
“instead of starting with the ideal and then moving towards the person, you start with the person and then move towards the ideal”
Paul Graham reflects that AI development confounded expert expectations: researchers assumed AI would scale from simple to complex intelligence, but instead got human-breadth capability riddled with hallucination from the start. He introduces the 'ragged frontier' concept to describe AI's uneven capability profile—solving hard math problems while failing basic factual lookups. The observation is historically interesting but offers no new technical or strategic signal beyond well-known LLM limitations.