[人人能懂AI前沿] 当AI开始说“电报”，做“清醒梦”，解“人性方程”

你有没有想过，AI之间开始说我们听不懂的“悄悄话”是为了什么？一个AI要学会新技能，最好的方法竟然是扔掉我们给它的“拐杖”？这期节目，我们就来聊聊几篇有趣的最新论文：看AI如何自创“电报文”实现高效沟通，用“清醒的梦”来检验学习成果，甚至尝试解开说服你的“人性方程”。准备好了吗？让我们一起探索AI正在发生的、超乎你想象的进化！00:00:31 当机器开始说“电报”，AI沟通的下一次进化00:05:06 AI学习的“断舍离”，扔掉“拐杖”，它能走得更远？00:10:55 小模型的大道理，为什么30亿参数能挑战万亿巨头？00:16:03 如何让机器人做一个“清醒的梦”？00:21:51 想说服我？先解开这道“人性方程”本期介绍的几篇论文：[CL] Large Language Models Do Not Always Need Readable Language [Shanghai Jiao Town University & The University of Sydney & Hefei University of Technology] https://arxiv.org/abs/2606.19857 ---[CV] You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences [UIUC & New York University] https://arxiv.org/abs/2606.15956 ---[CL] VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models [Sina Weibo Inc] https://arxiv.org/abs/2606.16140 ---[RO] SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation [University of Toronto & Physical Intelligence & NVIDIA] https://arxiv.org/abs/2606.18610 ---[AI] Using Cognitive Models to Improve Language Model Simulation of Human Persuasion Games [Princeton University] https://arxiv.org/abs/2606.17657 在小宇宙查看该单集文稿