Bootstrapping Developmental AIs: From Simple Competences to Intelligent Human-Compatible AIs

Abstract

Developmental AI creates embodied AIs that develop human-like abilities. The AIs start with innate competences and learn more by interacting with the world including people. Developmental AIs have been demonstrated, but their abilities so far do not surpass those of pre-toddler children. In contrast, mainstream approaches have led to impressive feats and commercially valuable AI systems. The approaches include deep learning and generative AI (e.g., large language models) and manually constructed symbolic modeling. However, manually constructed AIs tend to be brittle even in circumscribed domains. Generative AIs are helpful on average, but they can make strange mistakes and not notice them. Not learning from their experience in the world, they can lack common sense and social alignment. This position paper lays out prospects, gaps, and challenges for a bootstrapping approach to developmental AI that follows a bio-inspired trajectory. The approach creates experiential foundation models for human-compatible AIs. A virtuous multidisciplinary research cycle has led to developmental AIs with capabilities for multimodal perception, object recognition, and manipulation. Computational models for hierarchical planning, abstraction discovery, curiosity, and language acquisition exist but need to be adapted to an embodied learning approach. The remaining gaps include nonverbal communication, speech, reading, and writing. These competences enable people to acquire socially developed competences. Aspirationally, developmental AIs would learn, share what they learn, and collaborate to achieve high standards. They would learn to communicate, establish common ground, read critically, consider the provenance of information, test hypotheses, and collaborate. The approach would make the training of AIs more democratic.