分布外数据 (out of distribution)
集里怎么说它
- 《每家公司都该有自己的AI模型吗？》(08:24起):本集的核心判据:数据越在模型训练分布之外,后训练越可能带来真实能力提升;企业真正的分布外数据是内部产生的判断轨迹。
① 提到它的金句
7 条
对模型进行红队测试的本质就是去找到那些对该模型来说天然就是分布外的东西,这样你就可以绕过它的正常行为。
指向原始笔记的链接
the nature of a red-timing a model is to find things that are inherently out of distribution for that model so as you can bypass its normal behavior.
—— Zico Kolter · [11:47]
你可以有一个在我们基准测试上表现非常好的模型，但它们在某种程度上过拟合了，然后你把它放进一个更通用的助手类产品里，当你超出分布范围时它就有点崩溃了。
指向原始笔记的链接
you can have a model that does very well on our benchmarks but they’ve overfit to some degree and then you put it in a more generic assistant like product and it kind of falls apart for when you go out of distribution
—— Gabe Pereyra · [22:29]
我认为我们经常看到，你越偏离分布，就越会想要定制 harness。
指向原始笔记的链接
I think we often see that the more out of distribution you get, the more you’re going to want to customize the harness.
—— Harrison Chase · [21:51]
我实际上认为那个较少讨论的分布外话题是，如果我们不够乐观会怎样？
指向原始笔记的链接
I think actually the out of distribution topic that’s less discussed is what if we’re insufficiently optimistic?
—— Anish Acharya · [04:21]
循环会帮你爬到局部最优，然后进入平台期——你需要某种分布之外的思考，需要人类直觉，需要有人真正帮你降落到下一座山的山脚下。
指向原始笔记的链接
The loop will help you climb to the local maxima, but then it plateaus and you need some sort of out of distribution thinking, you need human intuition, you need somebody to actually help you land at the base of the next hill.
—— Anish Acharya · [15:15]
模型进行新思考的能力——分布外的思考——仍然非常有限。
指向原始笔记的链接
The ability for models to do new thinking out of distribution thinking is still really limited.
—— Anish Acharya · [13:19]
我们的总体看法是，你的数据越偏离分布，后训练就越有可能真正给你带来那种能力提升。
指向原始笔记的链接
Our general take has been like the more out of distribution your data is, the more likely that post training is actually going to give you that capabilities lift.
—— Yash · [08:30]
② 出现在这些集
1 集
- 《每家公司都该有自己的AI模型吗？》 — 作为概念
③ 关联
点进去有真内容 —— 本页主要出口
Yash · Applied Compute · OpenAI · Base 10 · Harvey · Cursor · Windsurf · 后训练 · RL · 评估