📄 中文摘要
现代基于语言模型的人工智能系统功能强大，但其能力在三个关键方面受到人类创造者的限制。首先，尽管模型的权重可以通过微调进行更新，但在预训练后从小型专业语料库中获取新知识的效率极低。其次，这些系统的训练严重依赖于有限的历史人类生成数据。第三，训练人工智能模型所使用的流程受到人类研究者能够发现和探索的算法的限制。该研究提出了三个章节，旨在打破这些依赖关系，以创建持续自我改进的人工智能。首先，...
持续自我改进的人工智能
出处: Continually self-improving AI
发布: 2026年3月20日
📄 English Summary
Continually self-improving AI
Modern language model-based AI systems are incredibly powerful, yet their capabilities are fundamentally limited by their human creators in three significant ways. Firstly, while a model's weights can be updated through fine-tuning, acquiring new knowledge from small, specialized corpora post-pretraining remains highly data-inefficient. Secondly, the training of these systems heavily relies on finite, human-generated data from history. Thirdly, the pipelines used to train AI models are constrained by the algorithms that human researchers can discover and explore. This research presents three chapters aimed at overcoming these limitations to create continually self-improving AI. Firstly, ...