Elvis Saravia urges building evals for open-ended AI improvement

Elvis Saravia, an AI researcher posting on X as @omarsar0, says recursive self-improvement depends on being able to measure what is being improved, whether that is a model, a codebase or a business function.

He argues AI performs best on verifiable problems and that open-ended work needs different methods and rigorous evaluations, not short-term benchmark chasing. He says DAIR.AI is researching and building evals in this area and invites collaboration.