你有没有想过,一个学得更久的AI“尖子生”,为什么反而忘得更快?或者,想让AI更懂英语,最好的方法竟然是教它别的语言?本期节目,我们将一口气解锁五篇最新论文带来的“反常识”洞见。我们会发现,决定AI效率的瓶颈可能不是算力而是“管理”,与AI对话的成本可以靠一本“字典”轻松打个二折,而一个好的AI模拟世界,追求的不是“长得像”,而是“反应像”。
00:00:32 大模型训练的悖论,为什么学得越久,忘得越快?
00:06:02 AI的效率瓶颈,不是算力,是“管理”
00:12:33 想让AI更懂英语?那就别只喂它英语
00:18:46 跟AI对话,如何省下80%的话费?
00:24:39 你的“差不多”不是我的“差不多”,如何让AI的模拟世界更靠谱?
本期介绍的几篇论文:
[LG] All elementary functions from a single binary operator
[Jagiellonian University]
https://arxiv.org/abs/2603.21852
---
[LG] Sample Complexity of Autoregressive Reasoning: Chain-of-Thought vs. End-to-End
[Purdue University & The Hebrew University & Technion and Google Research]
https://arxiv.org/abs/2604.12013
---
[CL] Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature
[Central University of Finance and Economics & Beijing Institute of Technology & TsingyuAI]
https://arxiv.org/abs/2604.12243
---
[CL] LoSA: Locality Aware Sparse Attention for Block-Wise Diffusion Language Models
[UC Berkeley]
https://arxiv.org/abs/2604.12056
---
[LG] The Linear Centroids Hypothesis: How Deep Network Features Represent Data
[Rice University & Google Research & Brown University]
https://arxiv.org/abs/2604.11962