PickVidsTraining AI Without Writing A Reward Function, with Reward Modelling
推荐理由
I find this video by Rob Miles still the best presentation of the ingenious original 2017(!) paper
强烈推荐cubefox · 原文
说明
这是2019年由Robert Miles AI Safety 发布的一部讲解，主题归在AI safety, AGI & impact。它讲的是reward modelling and rlhf这一支里的具体问题，而不是把整个领域扫一遍。适合已经能看懂基本术语、想继续往下走的人。内容上可以概括成：Robert Miles explains reward modelling, in which an AI learns what humans want from their preference comparisons instead of a hand-written reward function. He walks through the 2017 paper on deep reinforcement learning from human preferences.这些读者给出的理由，核心是：Seen as the clearest introduction to the idea behind RLHF, and also borrowed by commenters as a metaphor for human motivation.讲法不绑某一条新闻，过几年仍然可以当材料看。从2019年到2023年，有3位不同的读者在3个不同月份把它推荐出来，其中明确叫别人去看的有2次。适合用一集把一个概念补上的人。抓住它用的那个图像或类比，比背公式重要。如果已经会了基本定义，就把注意力放在它怎么把reward modelling and rlhf变成能跟着复述的例子。
视频信息
- 频道
- Robert Miles AI Safety
- 时长
- 17 分 52 秒
- 发布年份
- 2019
同主题更多
- What Happened to Horses Is Happening to Us
- Why Would AI Want to do Bad Things? Instrumental Convergence
- Veritasium: What Everyone Gets Wrong About AI and Learning – Derek Muller Explains
- Intelligence and Stupidity: The Orthogonality Thesis
- We're building a dystopia just to make people click on ads | Zeynep Tufekci
- Intro to AI Safety, Remastered