今天我们要拆解五篇带来颠覆性认知的最新论文：看AI如何学会把“做事与管事”彻底分开，如何靠一把“记忆橡皮擦”主动清理过载大脑；又是怎样用“窗口缓存”少折腾提速、在作画半路“实时纠偏”，甚至敢于拒绝“正确废话”只听关键反馈。这些AI前沿技术的破局之道，其实也是我们在复杂世界里最高级的成事密码，让我们马上开启探索！

00:00:30 为什么给AI更多时间，它反而搞砸了？揭示“做事”与“管事”的本质区别

00:03:57 你以为大模型缺的是脑容量，其实它缺的是一块“橡皮擦”

00:07:26 让AI跑得快又不出错，秘密竟然是“少折腾”？

00:13:40 做事的终极心法，别等交卷才审题，要在半路就纠偏

00:18:08 为什么“全盘接受建议”反而会让人变平庸？

本期介绍的几篇论文：

[AI] Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning

[Meta Superintelligence Labs]

https://arxiv.org/abs/2609.38147

---

[AI] Context Language Models

[University of Washington & MIT]

https://arxiv.org/abs/2609.37725

---

[LG] LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization

[UC Berkeley & University of Washington & MIT]

https://arxiv.org/abs/2609.38166

---

[CV] PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

[Google Cloud AI Research & MIT CSAIL]

https://arxiv.org/abs/2609.36199

---

[AI] AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation

[Google]

https://arxiv.org/abs/2609.38142