3 6 Ke·2026-01-09 02:42Core Insights - The article discusses the breakthrough of Nested Learning by Google's DeepMind, which may address the long-standing issue of "catastrophic forgetting" in AI, potentially leading to advancements towards Artificial General Intelligence (AGI) [1][52] - Nested Learning is positioned as a successor to the Transformer architecture, suggesting a shift from passive training to active evolution in AI systems [1][2] Group 1: Nested Learning and AGI - Nested Learning is highlighted as a significant research focus for DeepMind, with predictions that it could lead to minimal AGI by 2028 with a 50% confidence level [7][9] - The concept of Nested Learning is described as a framework that allows AI to build associative memory, enabling continuous learning without the need for retraining [1][19] - Shane Legg, co-founder of DeepMind, emphasizes that there are no current blockers to achieving continual learning, indicating progress in this area [5][7] Group 2: Technical Aspects of Nested Learning - The HOPE architecture is introduced as a mechanism for implementing Nested Learning, which combines fast self-updating systems with slow, multi-timescale memory [6][8] - The article outlines the importance of memory architecture, attentional bias, retention mechanisms, and learning rules in designing effective AI models [20][21] - The Nested Learning framework is said to unify various existing attention mechanisms and optimizers, allowing for a more dynamic understanding of memory in AI [21][24] Group 3: Performance and Implications - The HOPE architecture has shown superior performance in tasks requiring long context and continual learning compared to existing models, indicating its potential effectiveness [33][47] - The article raises concerns about the implications of AI systems that can learn continuously, suggesting that they may develop preferences based on past experiences, which could lead to ethical considerations [52]