Overview
- AI as computing paradigm: The strongest analogy for AI is a new computing paradigm, not electricity or industrial revolution
- Historical parallel: In the 1980s, specifiability predicted which jobs computing would automate (typing, bookkeeping, calculators)
- Software 1.0 vs 2.0: Hand-written programs automate what you can specify; neural networks automate what you can verify
- Verifiability requirements: Tasks must be resettable, efficient (many attempts), and rewardable via automated process
- Jagged frontier: Verifiable tasks (math, code, puzzles) progress rapidly—even beyond experts—while creative and strategic tasks lag
Takeaways
Andrej Karpathy wrote this essay on AI automation. The key insight is that verifiability—whether AI can "practice" a task with automated feedback—predicts which jobs will be automated next.
Software 1.0 easily automates what you can specify. Software 2.0 easily automates what you can verify.