![[人人能懂AI前沿] AI偷懒、复盘与泄密：那些藏在效率背后的秘密](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，AI在看似随机的打字节奏里，可能正在泄露自己的核心机密？或者，一个看似无害的几十兆“小补丁”文件，竟然能装下你全部的私人日记？本期节目，我们将一起揭开AI光鲜外表下的“隐藏设定”：从指导AI修炼更强“内功心法”的最新论文，到让AI学会“开小差”反而效率更高的反直觉策略，再到教会AI像顶尖棋手一样精准“复盘”自己的错误。准备好了吗？让我们一起潜入AI的后台，看看那些不为人知的智慧与博弈。
00:00:36 AI训练的内功心法，为什么有的模型学得又快又好
00:06:02 大模型加速的秘密，为什么“开小差”反而效率更高？
00:12:10 让AI学会“复盘”，从哪儿跌倒，从哪儿爬起
00:18:17 AI的小补丁，藏着多大的世界？
00:25:25 AI的秘密，藏在打字的速度里
本期介绍的几篇论文：
[LG] SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales
[NVIDIA]
https://arxiv.org/abs/2607.20548
---
[LG] Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
[NVIDIA]
https://arxiv.org/abs/2607.21535
---
[LG] Test-Time Scaling via Error Localization
[Google DeepMind]
https://arxiv.org/abs/2607.21453
---
[LG] How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning
[CMU & Columbia University]
https://arxiv.org/abs/2607.21351
---
[LG] Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing
[Purdue University & CMU]
https://arxiv.org/abs/2607.20723
00:00:36 AI训练的内功心法，为什么有的模型学得又快又好
00:06:02 大模型加速的秘密，为什么“开小差”反而效率更高？
00:12:10 让AI学会“复盘”，从哪儿跌倒，从哪儿爬起
00:18:17 AI的小补丁，藏着多大的世界？
00:25:25 AI的秘密，藏在打字的速度里
本期介绍的几篇论文：
[LG] SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales
[NVIDIA]
https://arxiv.org/abs/2607.20548
---
[LG] Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
[NVIDIA]
https://arxiv.org/abs/2607.21535
---
[LG] Test-Time Scaling via Error Localization
[Google DeepMind]
https://arxiv.org/abs/2607.21453
---
[LG] How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning
[CMU & Columbia University]
https://arxiv.org/abs/2607.21351
---
[LG] Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing
[Purdue University & CMU]
https://arxiv.org/abs/2607.20723
外观
