Jason Ma
① 他说过的话
12 条
这有点像你试图只靠 Reddit 或者只靠 Facebook 来构建一个互联网。
指向原始笔记的链接
It’s kind of like you’re trying to build an internet by just having Reddit or just having Facebook.
—— Jason Ma · [09:18]
我认为机器人领域的数据质量是一个非平稳问题,意思是对你的模型来说什么是高质量数据是随时间变化的。
指向原始笔记的链接
I think data quality in robotics is a non-stationary problem in the sense that what is high quality data for your model changes over time.
—— Jason Ma · [09:41]
对我们来说,我们仍然认为仅仅一百万小时的数据是不够的。这只是个开始。
指向原始笔记的链接
So for us, it’s like we still think just a million hours of data is not enough. It’s just the beginning.
—— Jason Ma · [12:50]
让小模型变得非常聪明的最好方法是蒸馏。你训练最大的模型,然后蒸馏。
指向原始笔记的链接
The best way to make a small model very smart is by doing distillation. You train the biggest model, then you distill.
—— Jason Ma · [19:00]
因为最终,是硬件承载智能。
指向原始笔记的链接
because eventually, it’s hardware that carries intelligence.
—— Jason Ma · [21:44]
现在很多讨论都集中在机器人基础模型、模型、数据上,但我认为硬件,以及造出好硬件有多难,这一点被严重低估了,但我认为那会极其重要。
指向原始笔记的链接
a lot of the discussion centers around a robot foundation model, model, data, but I think hardware and just how hard it is to build good hardware. It’s very under-emphasized, but I think That’s going to be extremely important.
—— Jason Ma · [23:14]
我把这看作是机器人版的基于人类反馈的强化学习。
指向原始笔记的链接
I think of this as the robotics version of reinforcement learning from human feedback.
—— Jason Ma · [31:31]
那就是,说到底,扩展是行得通的。但在机器人学中,扩展以及如何扩展要难得多,因为没有互联网。
指向原始笔记的链接
which is like, at the end of the day, scaling will work. But in robotics, to scale and how to scale just much harder because there is no internet.
—— Jason Ma · [42:38]
但真正的价值在于能无人监督地连续运行数小时的任务或者说工作流。
指向原始笔记的链接
But really the value is about tasks or like workflows that go on like unsupervised for hours.
—— Jason Ma · [48:42]
我觉得机器人的公司坟场很大。而且我觉得它们每家都有非常闪亮、非常酷的机器人。
指向原始笔记的链接
I think there’s a whole graveyard of robotics company. And I think all of them had very shiny, very cool robots.
—— Jason Ma · [53:49]
但每当你预测对世界的抽象表征时,你就在把自己关于什么值得预测的偏见编码进去。而 bitter lesson 告诉你,让我们去掉人类偏见。
指向原始笔记的链接
But whenever you’re predicting abstract representation of the world, you’re encoding your own bias into what’s important to predict. And the bitter lesson tells you, let’s remove human biases.
—— Jason Ma · [55:48]
但明天有人推出了一个视觉加触觉的世界模型,数据效率高出 100 倍。
指向原始笔记的链接
But tomorrow someone else comes out with a vision plus tactile world model that’s just like 100x more data efficient.
—— Jason Ma · [57:13]
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
Dyna Robotics · Dyna2 · 世界动作模型 · 世界模型 · 扩散 Transformer · 流匹配 · 思维链 · 第一人称数据 · 微调 · 蒸馏
