Menu
Sign In Search Podcasts Charts People & Topics Add Podcast API Pricing
Podcast Image

AI可可AI生活

[人人能懂] 从数据纯度、反馈标尺到心智公理

27 Sep 2025

Description

你是否想过,AI变聪明的速度,竟取决于数据里有多少“废话”?我们一句模糊的好评,又如何能变成让AI精准执行的指令?本期节目,我们将看到AI如何跳出经验的牢笼、自己悟出近道,并学会看人下菜碟,进化出因事而异的“情商”。我们甚至会揭示,洞察AI心思的终极难题,如何被巧妙地拆解成一道简单的计算题。准备好,和我一起探索这些最新论文背后的深刻智慧吧!00:00:35 AI变聪明的秘密:不是模型有多神,而是数据里有多少“废话”00:06:32 AI训练的两难困境:要么说不清,要么管太窄00:12:11 AI导航升级:如何用“笨”数据,教出“聪明”的活地图?00:18:03 AI的“情商”进化:怎么做到该一样时一样,该不同时不同?00:23:45 猜心思的最高境界:把它变成一道简单计算题本期介绍的几篇论文:[LG] Scaling Laws are Redundancy Laws  [Georgia Institute of Technology]  https://arxiv.org/abs/2509.20721 ---[CL] RLBFF: Binary Flexible Feedback to bridge between Human Feedback & Verifiable Rewards  [NVIDIA]  https://arxiv.org/abs/2509.21319 ---[LG] Offline Goal-conditioned Reinforcement Learning with Quasimetric Representations  [UC Berkeley & Princeton University]  https://arxiv.org/abs/2509.20478 ---[CL] LLM Output Homogenization is Task Dependent  [FAIR at Meta]  https://arxiv.org/abs/2509.21267 ---[LG] Inverse Reinforcement Learning Using Just Classification and a Few Regressions  [University of Washington & Netflix]  https://arxiv.org/abs/2509.21172 

Audio
Featured in this Episode

No persons identified in this episode.

Transcription

This episode hasn't been transcribed yet

Help us prioritize this episode for transcription by upvoting it.

0 upvotes
🗳️ Sign in to Upvote

Popular episodes get transcribed faster

Comments

There are no comments yet.

Please log in to write the first comment.