Verification becomes the scarce resource in the AI era
The author argues that as generation gets cheap, verification is the bottleneck, and notes that Jev can score candidate actions in sufficiently parseable environments, though it is not yet a mainstream agent paradigm.
The Scarcity of Verification As AI solutions become abundant, the scarcity of verification becomes increasingly apparent. The challenge is not only evaluating individual solutions, but also selecting the best among a massive number of plausible alternatives. For LLM development, benchmarks and environments with reliable verification are the first step toward understanding model capabilities. For software engineering, beyond requirements analysis, testing becomes increasingly important as code generation gets cheaper. For scientific research, human reviewers may no longer be able to keep up with the growing volume of AI-assisted research. And the reviewers themselves may increasingly rely on AI. For CUAs, when the environment is sufficiently parseable, Jev offers an approach to scoring candidate actions, although this has yet to become a mainstream agent paradigm. For mathematics, the pace of human peer review is struggling to keep up with AI-generated proofs. Interestingly, this makes P vs. NP feel even more fascinating: it essentially asks whether efficiently verifiable solutions can also be efficiently found. Verification itself introduces another challenge: agents are remarkably good at passing verification, even when the human-defined verification criteria are flawed. I guess verification needs to scale with generation!