The Missing Piece: Combining Foundation Models and Open-Endedness for Artificial Superhuman Intelligence ASI

According to the researchers, foundation models play a crucial role in achieving artificial superhuman intelligence (ASI). They provide a powerful base capability that, when combined with open-ended algorithms, can enable the kind of continual, experiential learning process required for true open-endedness. Foundation models have the ability to process vast amounts of data and capture human notions of interestingness, which is essential for developing open-ended systems that can make novel, human-relevant discoveries. The researchers outline four overlapping paths towards developing open-ended foundation models, drawing inspiration from the scientific method of forming hypotheses, experimentation, and codifying new knowledge. By leveraging foundation models and open-ended algorithms, researchers believe that it may be feasible to design a truly general open-ended learning system, which is a promising path towards achieving ASI.

DeepMind researchers define "open-endedness" in AI systems as the ability to continuously generate novel yet learnable artifacts. It is crucial for the development of artificial superhuman intelligence (ASI) because it allows an autonomous system to self-improve towards increasingly creative and diverse discoveries without end. This open-ended invention is how humans and society accumulate new knowledge and technology, making it essential for artificial superhuman intelligence. By combining foundation models with open-ended algorithms, AI systems can potentially achieve ASI and continuously expand their capabilities and knowledge in a way that can be utilized by humanity.