[人人能懂AI前沿] 从加密记忆、沙盘推演到主动遗忘

今天我们要聊的五篇最新论文，将彻底颠覆你对机器思维的固有认知。你将看到AI如何带上“加密日记本”彻底治愈健忘症，又如何凭借脑内“沙盘推演”绕开昂贵的现实试错陷阱。更反直觉的是，前沿AI不仅学会了主动遗忘杂音、靠撕掉“满分答案”来打破舒适区，甚至在面对过长工时时学会了原地挂机“摸鱼”。准备好升级思维，让我们一起看看这些前沿机制究竟能给我们的生活与认知带来什么启发！

00:00:27 把“日记本”还给人工智能，当AI拥有了真正的长期记忆，我们的生活会有什么改变？

00:04:46 别再盲目试错了，顶级高手和前沿AI都在用的“内化推演”心法

00:09:42 为什么“主动遗忘”才是真正的高级智慧？

00:13:43 为什么扔掉“好经验”，反而能让人更聪明？

00:17:17 给你的人工智能派个活儿，它竟然也会“摸鱼”？

本期介绍的几篇论文：

[CL] Real Long-Term Memory for AI: A 50-Million-Token Window That Is Faster and Cheaper Than Recompute

[Corbenic AI]

https://arxiv.org/abs/2610.10845

---

[CL] RSI-Anything: Unlocking Recursive Self-Improvement with World Models

[PhAI Labs & University of Macau & Shanghai Jiao Tong University]

https://arxiv.org/abs/2610.13912

---

[AI] Agent-Controlled Forgetting for Tool-Using Agents: Reversible Context Curation in Practice

[Münster]

https://arxiv.org/abs/2610.10590

---

[LG] GRPODropout: Less is More for Online Reinforcement Learning Rollouts

[Harbin Institute of Technology]

https://arxiv.org/abs/2610.11854

---

[AI] AgentTime: Can Agents Estimate and Control Their Own Runtime?

[MATS & Max Planck Institute for Intelligent Systems]

https://arxiv.org/abs/2610.09944