Open-Ended AI Needs New Vocabulary and Verifier Capabilities

This research delves into the fundamental limitations of contemporary AI systems, particularly their struggle with open-ended innovation. The authors contend that while AI excels at tasks like reasoning and coding within predefined frameworks, its capacity for genuine novelty is constrained by fixed conceptual vocabularies and evaluation criteria. They introduce two critical concepts: the "vocabulary gap," which describes the difficulty AI has in creating and stabilizing new representational primitives, and the "verifier gap," which highlights the challenge of assessing the long-term value of these new primitives. The paper proposes a unified framework viewing intelligence as cognitive discrepancy reduction, distinguishing between "intra-space transformations" that operate within existing frames and "generative transformations" that modify the frame itself. To advance open-ended AI, the authors suggest developing objectives that reward useful representational change, implementing persistent memory architectures for invented primitives, and creating adaptive verification mechanisms that evolve alongside new representations. This work offers a theoretical foundation for building more autonomously innovative AI.