Doj Chubuk
① 他说过的话
7 条
所以它不是知道科学的状态，而是知道如何做科学。
指向原始笔记的链接
So rather than knowing the state of science, it knows how to do science.
—— Doj Chubuk · [03:09]
我们不认为以今天这种训练方式训练出来的系统，会思考十亿或万亿个 token 后突然直观地领悟到：这里有一个新的高温超导体。
指向原始笔记的链接
And we don’t think that a system trained in the way that they’re trained today will think for a billion or a trillion tokens and suddenly intuit, here’s a new high-temp superconductor.
—— Doj Chubuk · [05:16]
数据根本就不存在，而这才是你真正构建新知识的方式，而不只是重复现有的知识。
指向原始笔记的链接
The data simply just doesn’t exist, and that’s how you’re going to actually build new knowledge, not just rehash the existing knowledge.
—— Doj Chubuk · [05:40]
更令人意外的是，我们内部已经有 GPT-4，而它在公司内部并没有成为一个爆火的产品。
指向原始笔记的链接
It was further surprising because we had GPT-4 internally, and it wasn’t this hugely viral product within the company.
—— Doj Chubuk · [23:11]
但就是一连串的获取反馈、改进任务分布、改进数据质量，正是这些看似无聊的细节造就了今天的系统。
指向原始笔记的链接
But just a series of getting feedback, improving the task distribution, improving the data quality, just all these kind of boring details is what led to the system as of today.
—— Doj Chubuk · [23:50]
你的预训练模型已经知道答案，而你针对一个已经存在于预训练语料库中的答案做强化学习时，系统基本上可以伪造一些工作、得到正确答案，而你会强化这些策略。
指向原始笔记的链接
Your pre-trained model already knows the answer, and you do reinforcement learning derived against an answer that was already in the pre-training corpus, that the system can basically fake some work, get to the right answer, and you reinforce those strategies.
—— Doj Chubuk · [39:47]
我的意思是，从软件工程这方面来说，我们今天完全可以停下来，去复现每一个现有的诺贝尔奖，至少物理学的。
指向原始笔记的链接
I mean, I think from the software engineering side, we could literally stop today and find every existing Nobel Prize, at least from the physicals.
—— Doj Chubuk · [51:40]
② 出现在这些集
1 集
③ 他谈到的
点进去有真内容 —— 本页主要出口
Liam Fedus · Mario · Periodic Labs · OpenAI · Google DeepMind · ChatGPT · 大型语言模型 · 超导体 · 合成超级智能 · 空想主义不够