今天我们将通过五篇硬核又有趣的最新论文,带你拆解AI前沿技术背后的现实隐喻。我们将探讨AI如何通过反事实模型分清“运气与实力”,揭开算法单一种植带来的系统性排斥陷阱,并用热力学视角看懂AI写代码时80%的算力“废热”。此外,我们还会看到大模型如何化身物理学家拯救不完美的光计算硬件,以及全新的TOASTER架构如何巧妙解开记忆机制中效率与深度的死结。准备好给大脑来一场认知升级了吗,我们马上开始!
00:00:38 搞砸了一件事,到底是运气不好还是能力不行?
00:04:41 为什么你自以为的10次机会,其实只有1次?揭开AI时代的隐形陷阱
00:09:55 AI写代码就像烧锅炉,为什么80%的算力都变成了“废热”?
00:14:14 真实世界的不完美,AI是如何看懂的?
00:18:23 效率与深度只能二选一?如何解开思维死结
本期介绍的几篇论文:
[LG] Counterfactual World Models: JEPAs That Reason About "What If"
[Brown University & New York University]
---
[LG] Monoculture Robust Learning
[Stanford University & Cornell University]
---
[AI] Where the Tokens Go: Heat, Entropy and Error in Agentic Coding
[University of Toronto]
---
[AI] In Situ Training of Optical Neural Networks by Coding Agents
[Stanford University & Ecole polytechnique fédérale de Lausanne]
---
[LG] Toaster: Nonlinear Memory from Cascaded Linear Memories and Target Propagation
[Google Research]
![[人人能懂AI前沿] 从反事实复盘、算法单一种植到AI的物理直觉](https://image.xyzcdn.net/FqWpK8fpivLboaqBbRHUe_BCOvxu.png@small)