今天，我们将一起探索如何让AI更强大也更“像人”：比如，让AI的记忆不再是短暂的，而是像组织一样层层沉淀；让两个聪明的AI合作不再犯傻，甚至通过“自我反思”拥有稳定的性格。更有趣的是，我们还会看到，教AI谱曲就像教它一门外语，而训练它不出错的最好方法，竟是给它找个专门“抬杠”的对手。让我们马上进入今天的前沿探索之旅！

00:00:32 AI的记忆黑洞：为什么我们看到的深度学习只是冰山一角

00:06:40 两个聪明的AI，为何凑在一起就犯傻？

00:12:05 让AI学会“谱曲”，只需教它一门新外语

00:17:17 打造有“个性”的AI助手

00:23:36 如何训练一个更聪明的AI？给它找个“抬杠”的对手

本文介绍的几篇论文：

[LG] Nested Learning: The Illusion of Deep Learning Architectures  

[Google Research]  

https://abehrouz.github.io/files/NL.pdf

---

[LG] The Collaboration Gap  

[Microsoft Research & EPFL]  

https://arxiv.org/abs/2511.02687

---

[AS] MIDI-LLM: Adapting Large Language Models for Text-to-MIDI Music Generation  

[MIT]  

https://arxiv.org/abs/2511.03942

---

[CL] Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI  

[University of Cambridge & MATS & Allen Institute for AI & Anthropic]  

https://arxiv.org/abs/2511.01689

---

[LG] RLAC: Reinforcement Learning with Adversarial Critic for Free-Form Generation Tasks  

[Shanghai Jiao Tong University & UC Berkeley]  

https://arxiv.org/abs/2511.01758