如果AI学会了“偷懒”和“作弊”，我们是该高兴还是该担心？今天，我们就来聊聊AI正在觉醒的几种“新智慧”：它不仅开始用“看图”的方式读完一整本书，还学会了像我们一样把精力花在刀刃上。我们还会探讨，如何用一把“尺子”去精确测量它的能力短板，以及它如何像武林高手一样，通过“左右互搏”实现自我进化。准备好了吗？让我们一起揭开这些最新论文背后，AI正在发生的深刻变革。00:00:34 给AI一双眼，让它读完一整本书00:06:06 给AI一把尺子，量量它离我们有多远？00:11:37 AI的左右互搏：如何不花钱，让AI自己把自己逼成高手？00:17:05 AI的“精力管理”智慧00:21:55 AI学会了“耍滑头”，我们该怎么办？本期介绍的几篇论文：[CL] Glyph: Scaling Context Windows via Visual-Text Compression [Tsinghua University & Zhipu AI] https://arxiv.org/abs/2510.17800 ---[CL] A Definition of AGI [Center for AI Safety & University of California, Berkeley & Morph Labs] https://arxiv.org/abs/2510.18212 ---[CL] Search Self-play: Pushing the Frontier of Agent Capability without Supervision [Quark LLM Team, Alibaba Group] https://arxiv.org/abs/2510.18821 ---[CV] Accelerating Vision Transformers with Adaptive Patch Sizes [CMU & KAIST] https://arxiv.org/abs/2510.18091 ---[CL] ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases [CMU & Anthropic] https://arxiv.org/abs/2510.20270

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related