Few-shot or One-shot Learning
Unsupervised Learning with large world models
These components are tied together by meta-learning loops and memory systems that simulate cognitive development. AGI must incorporate continual learning and adapt across tasks without retraining from scratch (Schmidhuber, 2015).
3. Keeping AGI in Control
Control in AGI means ensuring its goals and actions remain aligned with human values. Researchers have proposed several approaches:
a. AI Alignment
Aligning AGI's internal objectives with ethical human behavior.
Stuart Russell (2019) emphasized the need to design AGI systems that know they don't know what humans truly want, leading to systems that defer to human judgment.
b. Corrigibility
Systems must accept human corrections and overrides (Soares et al., 2015).
Includes designing AGI to seek assistance or defer control when uncertain.
