EPISODE · May 11, 2026 · 33 MIN
[人人能懂AI前沿] 从聪明花钱、英雄涌现到AI的“偏科”报告
from AI可可AI生活
今天我们要聊点特别有意思的话题:AI是怎么“思考”和“成长”的?我们会从几篇最新的论文出发,看看AI如何学会聪明地“花钱”,如何在学习中分清“英雄”与“集体”;然后,我们会揭秘它那套先打“草稿”再复核的高效工作法。最后,我们会用一把全新的“尺子”去度量它成长的极限,并给它做一份“智商测试”,看看这个“天才”到底偏科有多严重。准备好了吗?让我们一起潜入AI的大脑深处。00:00:33 如何打造一个“更划算”的虚拟世界?00:06:10 大模型的“缩放法则”里,藏着什么秘密?00:13:40 快与慢,AI世界里的“草稿式”工作法00:19:55 你的数据值多少钱?一个新尺子,看透AI的增长极限00:26:42 AI的“智商”报告,一个偏科天才的养成本期介绍的几篇论文:[LG] On Training in Imagination [Weizmann Institute of Science & New York University & Columbia University] https://arxiv.org/abs/2605.06732 ---[LG] Spectral Dynamics in Deep Networks: Feature Learning, Outlier Escape, and Learning Rate Transfer [Harvard University] https://arxiv.org/abs/2605.07870 ---[CL] Fast Byte Latent Transformer [FAIR at Meta] https://arxiv.org/abs/2605.08044 ---[LG] On the Invariance and Generality of Neural Scaling Laws [Johns Hopkins University & MIT] https://arxiv.org/abs/2605.07546 ---[AI] Uneven Evolution of Cognition Across Generations of Generative AI Models [Google DeepMind & Google Research] https://arxiv.org/abs/2605.06815 在小宇宙查看该单集文稿
Embed this episode
Ready to play
[人人能懂AI前沿] 从聪明花钱、英雄涌现到AI的“偏科”报告
No transcript for this episode yet
Similar Episodes
No similar episodes found.