Overview
- Kaplan predicts a turning point between 2027 and 2030 when AI begins training and improving its own successors in a self-improvement loop.
- He says Anthropic’s newest models can build software agents, sustain hours-long programming tasks, and produce extended reasoning without human input.
- He identifies loss of control over highly capable systems as a core danger, raising doubts about future alignment with human agency.
- He warns of misuse by bad actors, including efforts to co-opt powerful systems to enforce personal will.
- He argues AI will outpace students in academic tasks, noting even young learners will not outperform systems at essays or maths exams.