从“战略指纹”到“解耦赋分”：AI的非技术性进化

00:01:29 AI的“人设”：谷歌“腹黑”，OpenAI“傻白甜”？ 

00:06:09 给AI“减肥餐”：为什么数据越多，模型可能越笨？ 

00:10:32 AI训练场上的新策略：先当“神算子”，再做“阅读家”

00:14:53 功劳簿到底该怎么写？ 

今天介绍的四篇论文：

[LG] Strategic Intelligence in Large Language Models: Evidence from evolutionary Game Theory

K Payne, B Alloui-Cros  

[King’s College London & University of Oxford]  

arxiv.org

---

[LG] Data Uniformity Improves Training Efficiency and More, with a Convergence Framework Beyond the NTK Regime

Y Wang, S Gu  

[Johns Hopkins University & UC Berkeley]  

arxiv.org

---

[CL] Should We Still Pretrain Encoders with Masked Language Modeling?

H Gisserot-Boukhlef, N Boizard, M Faysse, D M. Alves...  

[Artefact Research Center & Diabolocom & Illuin Technology]  

arxiv.org

---

[LG] Disentangled Feature Importance

J Du, K Roeder, L Wasserman  

[CMU]  

arxiv.org