application的话应该还来得及 
rl发展到今天感觉能蹭上的好蹭的大家已经蹭完了 又开始emo自己生的太晚了
要是早生十年的话 天不生我ssinz7 万古如长夜 
survey无可厚非，本来就需要人来总结过去的工作给未来人做导航，问题是一个方向一两年survey一次够了，如果是迭代快的最多一年一次，你一个月survey一次就太水了…
actually I think之前RL有so many unresolved/ongoing problems that have been stopped due to
it’s good (or I want others ) to resume
因为说白了现在还是卷算力base model (+ engineer/scientist experience) + somewhat elegant and small algorithm components for downstream tasks — everything just existed (but maybe with a fancier new name 
)
很难说survey和现在无穷无尽almost useless and so biased benchmarks孰好孰坏
1 个赞
哪位网红老师啊？感觉现在的AP全在小红书上当网红
来蹭蹭大佬带，刚毕业统计PhD在养老厂摸鱼，人不帅写码快
Learning theory应该去stoc, colt, focs吧，蹭什么neurips啊
坑太少了，nips一年几千呢 
GPT 有 learning theory吗
这不就是experiment driven 好用了再来找个theory backup
一起加油，我也想冲菲尔兹。