Verifiability (karpathy.bearblog.dev)

This piece reframes AI as a new programming paradigm—“Software 2.0”—and argues that the key predictor of which tasks will be automated is verifiability rather than specifiability. In the 1980s, whether a job could be automated hinged on whether its steps could be hand-coded (Software 1.0). Today, we specify objectives (e.g., accuracy or reward functions) and search program space via gradient descent to produce neural networks. Tasks that are verifiable—those where performance can be automatically measured and improved—are much easier to optimize and thus progress rapidly, sometimes surpassing human experts. Verifiability is defined concretely: the environment must be resettable (you can run many attempts), efficient (you can make many trials), and rewardable (you can automatically score attempts). This explains the “jagged” frontier in LLM and AI capabilities: domains with clear, repeatable rewards (math problems, coding, puzzle-like tasks, or any task with objective feedback) advance quickly via supervised learning or reinforcement learning, while creative, strategic, or context-heavy tasks lag and rely on weaker mechanisms like imitation or brittle generalization. The practical implication for the AI/ML community is to prioritize creating verifiable environments, reward signals, and benchmarks to accelerate robust automation—and to recognize the limits where verification is inherently hard.