Verifiability: The Key Factor in AI Automation

Artificial intelligence operates as a new computing paradigm fundamentally centered on automating digital information processing. Unlike traditional manually-written programs—termed "Software 1.0"—modern AI follows a distinct model ("Software 2.0") where objectives are specified and neural networks are trained via gradient descent to achieve those objectives. Verifiability emerges as the strongest predictor of whether a task can be successfully automated by AI systems. When tasks are verifiable, they become directly optimizable through standard approaches or reinforcement learning, enabling neural networks to achieve high performance through repeated practice. This explains the uneven progress of large language models: verifiable tasks such as mathematics and code generation advance rapidly, potentially surpassing expert-level capabilities, while creative and context-dependent tasks requiring real-world knowledge advance more slowly.