Latent Space
🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science
00:00
00:00
Are you talking about introducing explicit priors that, you know, she based upon some human intuition,
你是在说引入 explicit priors，就是那种，你知道，她基于某些人类直觉的，
00:05
or maybe in this case, LOM intuition? When you talk about multiple hypothesis testing, right? There's
或者也许在这种情况下，是 LOM intuition？当你说到 multiple hypothesis testing 的时候，对吧？有
00:12
predictive models and there's descriptive models. A predictive model is like, let's say you just have
predictive models，还有 descriptive models。一个 predictive model 就像，比如说你只有
00:17
a, you have some inputs and you have some outputs. And you just, I just want to build a piece of code
嗯，你有一些 inputs，你有一些 outputs。然后你只是，我只想写一段代码
00:22
that tries to just have the lowest error rate on some data set. This is a physical model. A descriptive
试图在某个 data set 上达到最低的 error rate。这是一个 physical model。一个 descriptive
00:28
model is actually what's fine as it's trying to get to, which is, okay, it should be able to
model 实际上就是 fine 试图达到的东西，也就是说，好吧，它应该能够
00:33
extrapolate because it has sort of the physics or the actual, some description of reality that's
extrapolate，因为它有某种 physics，或者实际的、对现实的一些描述，这些
00:39
captured within it. And then you can use it to extrapolate. Yes, Newton thought of apples and
被捕捉在其中。然后你可以用它来 extrapolate。是的，Newton 想到了苹果和
00:43
gravity, but gravity isn't actually about apple right. If you take into 17th century machine learning
gravity，但 gravity 其实并不真的是关于苹果，对吧。如果你拿一个 17 世纪的 machine learning
00:47
model, like, oh, apples will fall, but how about planets? You know, I don't know. I have no data
model，比如说，哦，苹果会掉下来，但行星呢？你知道，我不知道。我没有 data
00:51
about planets. So who knows what they do, right? The distinction between those is a little bit blurry,
关于行星的。所以谁知道它们会做什么，对吧？这两者之间的区别有点模糊，
00:57
right? Because when a physicist or scientist comes, they use their intuition, or maybe even more
对吧？因为当一个物理学家或科学家来的时候，他们会用自己的直觉，或者甚至不止
01:02
than intuition. Like essentially, there's maybe a solid pile of facts that they know about the world.
是直觉。就像本质上，也许有一大堆他们知道的关于世界的坚实事实。
01:07
And then they make sure that whatever model they build is sort of consistent with what's known.
然后他们会确保自己构建的任何 model 都跟已知的东西大致一致。
01:14
Welcome to Latenspace Science. I'm Brandon, I'm joined by my cohost, RJ. It's a pleasure to have
欢迎来到 Latenspace Science。我是 Brandon，我的联合主持人 RJ 也和我在一起。很高兴能邀请到
01:20
John Platt, you know, with us today. John is a Google fellow in head of applied science
John Platt，你知道，今天和我们在一起。John 是 Google Fellow，也是应用科学负责人
01:25
at Google Research. He has really a fun, like background tea. I guess you described yourself when
在Google Research。他真的有一个很有趣的背景故事。我觉得你几分钟前描述自己的时候说你是个beginner。不，beginner，beginner，beginner。我很兴奋，完全什么都兴奋。这真的，真的能看出来，是的，以前，坦白说，如果我说的这些有任何不对的地方，但你14岁开始上大学，18岁在Caltech开始读PhD。你的，你的导师或者联合导师是John Hopfield，对吧？哦，对。对对。他刚拿了诺贝尔奖，你知道，两年前？是的。John创造了好几个，他负责了好几个教科书级算法，一个叫Platt scaling。另一个叫sequential minimal optimization，这是训练SVM的教科书级算法。即使今天，它仍然是，如果你用SK Learn，它就是
01:30
you're talking a few minutes ago as a beginner. No, beginner, beginner, beginner. I'm excited and
你几分钟前还说自己是新手呢。不，新手，新手，新手。我很兴奋，而且
01:35
absolutely everything. And it really, it really shows, yeah, it used, it was, frankly, if I'm wrong
对所有的一切都兴奋。而且这真的，真的能看出来，对，它以前，它确实，坦白说，如果我哪里说错了
01:40
about any of this stuff, but so you started college at 14 and started your PhD at 18th at Caltech.
关于这些事的话，但话说回来，你 14 岁上大学，18 岁就在 Caltech 开始读 PhD。
01:47
Your, your advice or co-advised by John Hopfield, right? Oh, yeah. Yeah, yeah. Who just won a Nobel
你的，你的导师，或者说是联合指导老师，是 John Hopfield，对吧？哦，对。对，对。他刚拿了诺贝尔
01:52
prize in, you know, two or two years ago? Yes. John created several, it was responsible for several
奖，你知道，两年前还是两年前？是的。John 创造了好几个，他负责过好几个
01:58
textbook algorithms, one known as Platt scaling. Another one, sequential minimal optimization,
教科书级算法，其中一个叫 Platt scaling。另一个叫 sequential minimal optimization，
02:03
which is the textbook algorithm for training SVMs. Even today, it's still, if you use SK Learn, it's
也就是 training SVMs 的教科书级算法。即使今天，它仍然，如果你用 SK Learn，它还是
02:09
new. John is discovered and named two asteroids has a Oscar for technical developments from 2006.
新的。John 还发现并命名了两颗
02:18
So if you've ever watched a, a Pixar movie, you've seen John's algorithms and work. John has an
所以，如果你看过 Pixar 的电影，你就见过 John 的算法和作品。John 有一个
02:23
Erdoz baking number of six or three, three and three from either side. And I'm going to skip over
Erdoz baking number 是六或三，三和三，从两边算。接下来我要跳过
02:29
like 20 years, but then jumping into Google at, you know, working at Google Sciences, you've
大概 20 年，但后来你加入 Google，你知道，在 Google Sciences 工作，你
02:36
worked on fusion quantum computing, climate modeling, and many other topics. Is that more or less right?
做过 fusion quantum computing、climate modeling，还有很多其他课题。差不多是这样吧？
02:43
That's right. Yeah. Okay. Okay. Cool. Well, I miss anything important for today. Oh, no. I mean,
没错。对。好。好。酷。那我今天有没有漏掉什么重要的？哦，没有。我是说，
02:48
I've also done, you know, lots of applied math and signal processing and all sorts of fun things
我也做过，你知道，很多 applied math 和 signal processing，还有各种好玩的东西
02:52
like that. Yeah. Yeah. Yeah. I think you also, uh, you're, you're, if you have a fun story about
之类的。对。对。对。我觉得你也，呃，你，你要是有个好玩的故事，关于
02:57
patents and, um, uh, the iPhone too. The iPod. Yeah. Yeah. Yeah. Welcome. Thank you. Yeah.
专利，还有，嗯，呃，iPhone 也是。iPod。对。对。对。欢迎。谢谢。嗯。
03:04
Thank you for having me. Can you tell us about the era is the, I think the way that the acronym is
谢谢你邀请我。你能给我们讲讲 ERA 吗？我想这个缩写词是
03:10
pronounced, um, and, and I know that there's a lot of different semi-related stuff out there,
这么读的，嗯，而且，而且我知道外面有很多不同的、有点相关的东西，
03:17
both within and outside of Google. So what can you tell us a little bit about the details of
Google 内部和外部都有。所以你能稍微给我们讲讲一些细节吗，关于
03:22
era and, and what makes it special? Well, we've been doing, um, sort of AI for science in Google
ERA 以及，以及它特别在哪里？嗯，我们在 Google research 做 AI for science 已经
03:28
research for more than 10 years now. Yeah. And, uh, around 10 years ago, it was very much, um,
超过 10 年了。对。而且，呃，大概 10 年前，它很大程度上，嗯，
03:36
using, I don't know what you call it now, maybe classical, uh, classical machine learning models,
用的是，我不知道你们现在怎么叫，也许是 classical，呃，classical machine learning models，
03:40
you know, things like, you know, convolutional ads or whatever. And they were specific models to
你知道，比如，你知道，convolutional ads 之类的。而且它们是特定的 models，用来
03:44
build to, um, you know, solve specific, um, science problems. About two years ago, we got very
构建出来，嗯，你知道，去解决具体的，呃，科学问题。大概两年前，我们变得非常
03:51
excited about these more general, uh, LM's that have popped up in the last few years.
兴奋，对这些更通用的，呃，过去几年冒出来的 LM。
03:57
And we were wondering what can be done with them. And, of course, a lot of people have been, uh,
然后我们在想，能用它们做什么。而且，当然，很多人一直，呃，
04:00
playing and trying to figure out what the right, what the right thing to do is. And, uh,
在玩，试着弄清楚什么才是对的，什么才是正确的做法。而且，呃，
04:05
we kind of, um, stumbled into this, uh, mapping, uh, and there's found that many different scientific
我们有点，嗯，误打误撞进入了这个，呃，mapping，呃，然后发现很多不同的科学
04:12
problems, uh, can be mapped into something we call scoreable tasks. So you can often phrase the
问题，呃，可以被映射成我们称之为 scoreable tasks 的东西。所以你经常可以把
04:18
scientific problem as a gosh, I really would like to have a piece of code that, you know, maximizes
科学问题表述成，哎呀，我真的很想有一段 code，你知道，能最大化
04:24
some score. And it's surprising the number of, um, different sort of scientific problems,
某个 score。而且令人惊讶的是，嗯，不同种类的科学问题，
04:30
you can make a lot of progress on, uh, by mapping into that framework. Well, one thing is, uh,
你可以，呃，通过映射到那个 framework 里，取得很多进展。嗯，有一点是，呃，
04:35
a lot of scientists spend a lot of time sort of building models. They might be statistical
很多科学家会花很多时间，差不多就是在构建 models。它们可能是 statistical
04:38
models or they might be, you know, physically based models. Um, and it was a statistical model,
models，也可能是，你知道，physically based models。嗯，而那是一个 statistical model，
04:44
like in machine learning, your scoring function is, well, I have some data set and I'd like to have
就像在 machine learning 里，你的 scoring function 是，嗯，我有一些 data set，我希望
04:48
the fit the model on the data set go up. Uh, and we can talk about overfitting in a minute. But, um,
模型在 data set 上的 fit 能上升。呃，我们可以待会儿再聊 overfitting。但是，嗯，
04:54
uh, that was one of our questions. That's actually very, that's actually very interesting.
呃，那是我们的问题之一。那其实非常，那其实非常有意思。
04:59
Maybe so machine learning is kind of a subset of this sort of scoreable task, right? But you
所以也许 machine learning 算是这种 scoreable task 的一个 subset，对吧？但你
05:04
could do other kind of things like, uh, especially Michael Brenner, who's, um, uh, the lead author
还可以做其他类型的事情，比如，呃，尤其是 Michael Brenner，他是，嗯，呃，第一作者
05:09
on the era paper, he's very, very skilled because he, he likes to, to sort of knock out a scientific
关于 era paper，他非常非常熟练，因为他，他喜欢现在用一个工具，一个晚上就差不多搞定一篇科学论文。嗯，所以，嗯，applied math 里有个东西叫 asymptotic expansions。然后，呃，也就是你，你在问，这个，这个 ordinary differential equation 或者 partial differential equation——比如说 ordinary differential equation——会怎么表现？里面有个参数带着 epsilon，你想说，当 epsilon 趋于零时它会怎么表现？然后你可以把它变成一个 empirical task，本质上就是让它，呃，它，它提出一些 asymptotically correct 的解。然后你检查一下，呃，这个 asymptotic solution 对不对，比如 epsilon 等于 1e-4 之类的。
05:15
paper in an evening now with a tool. Um, so, um, there's something in applied math called
现在用一个工具，一晚上就能写出 paper。嗯，所以，嗯，applied math 里有个东西叫
05:21
asymptotic expansions. And, uh, which is you, you're asking, how does this, how does the ordinary,
asymptotic expansions。然后，呃，也就是说，你，你是在问，这个，这个 ordinary
05:26
differential or partial differential equation say ordinary differential equation behave,
differential 或 partial differential equation，比如说 ordinary differential equation，会怎么表现，
05:30
there's some parameter has an epsilon in it and you're trying to say, how does behave is epsilon
里面有个参数带了个 epsilon，然后你想说，当 epsilon
05:34
goes to zero? And you could turn, you can turn that into a empirical task by essentially asking
趋近于零的时候它会怎么表现？然后你可以把，你可以把它变成一个 empirical task，基本上就是问
05:42
that it, uh, it, it proposes some solutions that are asymptotically correct. And you check to
它，呃，它，它提出一些 asymptotically correct 的解。然后你去检查
05:48
see if the, uh, asymptotic solution is correct for like epsilon equals one e minus four or something.
看看这个，呃，asymptotic solution 在比如 epsilon 等于 1e-4 或者差不多的时候是不是对的。
05:54
And then you, you check that fit, but then then you ask, uh, uh, you ask Gemini, which is the
然后你，你检查那个 fit，但然后，然后你就问，呃，呃，你问 Gemini，也就是它那个
05:59
core AI underneath it to, to do the mathematical reasoning, try to solve the problem while,
底层的 core AI，去，去做 mathematical reasoning，试着解决这个问题，同时，
06:05
while also maximizing the fit to the, to the data. So you can actually, there's a lot of sort of
同时也最大化对，对数据的 fit。所以其实你可以，有很多种
06:10
tricks you can do because it's not that the underlying thing that's, that's altering the code or the
tricks 可以用，因为底层那个，那个在改代码的东西，或者那个
06:15
underlying thing that's sort of making the decisions is not a random process. It's an AI itself
在做决策的底层东西，并不是一个 random process。它本身就是一个 AI，
06:21
that is smart and knows about things and knows a lot about the world. You can get a lot of,
很聪明，懂很多事情，也了解很多关于这个世界的东西。你可以得到很多，
06:26
it's all a lot of interesting problems because that sort of core inner loop is an AI that has huge
这都是很多有趣的问题，因为那个所谓的 core inner loop 是一个拥有巨量
06:30
amounts of, um, prior knowledge. So that's sort of the, the, the trick. So we, we've been, uh,
呃，prior knowledge 的 AI。所以这就是那个，那个，那个 trick。所以我们，我们一直，呃，
06:37
running around trying to map lots of scientific problems into scoreable tasks and trying to solve
到处跑，试着把很多科学问题映射成 scoreable tasks，并试着解决
06:42
them. And it's, it's really been kind of fun. I'm happy to talk about the ones that I've been involved
它们。而且这，这真的还挺好玩的。我很乐意聊聊那些我参与
06:46
in at least. Yeah, I would love to hear about some of the more, so it's a statistical one is the way
过的，至少我参与过的那些。是啊，我很想听一些更……所以那个 statistical one 就是那种
06:50
everyone listening will probably know about what you just mentioned makes science. What are some of
所有在听的人大概都会知道你刚才提到的 makes science。还有哪些
06:55
the other interesting ones? Uh, let's see that, that are not statistical. Let's see because we have
其他有趣的？呃，让我想想，那些，那些不是 statistical 的。让我想想，因为我们有
07:00
interesting ones like, uh, one that we just, uh, put a paper up on archive is, um, or actually,
一些有趣的，比如，呃，我们刚刚，呃，在 archive 上挂了一篇论文的，嗯，或者说实际上，
07:07
I think it might be on GitHub. It's, um, you often run into this in remote sensing because there's
我觉得它可能在 GitHub 上。这个，嗯，你在 remote sensing 里经常会碰到，因为总是
07:12
always a trade-off. There's satellites flying above the earth. And there's a trade-off between how
有一种 trade-off。有卫星在地球上空飞。而有一种 trade-off 是在于如何
07:18
frequently they can revisit a spot in the earth, what their spatial resolution is, how big the
经常他们可以重新访问地球上的某个点，他们的 spatial resolution 是多少，以及
07:23
pixels are, and their spectral resolution, so how many bands they have. And ideally, you'd like to
pixels 有多大，还有他们的 spectral resolution，也就是他们有多少 bands。而且理想情况下，你希望
07:29
have a monitoring of the earth that's constant and, you know, frame every five minutes at, at
有一个对地球的监测，是持续不断的，而且，你知道，每五分钟 frame 一次，在，在
07:35
high-perspectra resolution at whatever 10 centimeters. You can't get that. But, for example,
high-perspectra resolution，在随便 10 厘米。你得不到那个。但是，比如说，
07:41
to monitor CO2, the atmospheric concentration of CO2, you can take data from one satellite.
要监测 CO2，CO2 的 atmospheric concentration，你可以从一颗卫星获取数据。
07:49
That's, for example, it's OCO2 or OCO3. OCO3 is actually attached to the International Space Station,
那，比如说，是 OCO2 或 OCO3。OCO3 实际上是附着在 International Space Station 上的，
07:56
but so it gets you like a little strip of CO2 measurements that are highly accurate and pretty
但它会给你像一小条 CO2 测量数据，非常准确而且相当
08:03
high-resolution. You can actually try to do, uh, there, you could, from, because a lot of it is
high-resolution。你其实可以试着去做，呃，在那里，你可以，从，因为很多都是
08:08
in the infrared, uh, weather satellites, like, goes, has some infrared bands, and it, and it takes a
在 infrared 里，呃，weather satellites，比如 goes，有一些 infrared bands，而且它，它要拍一张
08:14
picture every essentially five minutes, but the pixels are very large, and it doesn't have such great
照片，基本上每五分钟一张，但 pixels 非常大，而且它没有这么好的
08:20
spectral resolution in terms of it wasn't designed to find CO2. So you just ask one to estimate the
spectral resolution，就它并不是为了找 CO2 设计的。所以你就让一个去估计
08:25
other, and you shovel other data in like, what's the current weather, uh, what, what's sort of the long
另一个，然后把其他 data 塞进去，比如，现在的天气怎么样，呃，什么，什么算是长期
08:30
term, um, you know, albedo, uh, and, and so you, so era came up with this very nice, um,
趋势，嗯，你知道，albedo，呃，然后，然后，所以 era 做出了这个非常不错的，嗯，
08:37
this very nice model that can sort of, can do almost like super-resolution and informed super-resolution
这个非常不错的 model，它能有点，能做得几乎像 super-resolution，还有 informed super-resolution
08:42
of, of, uh, uh, ones had like to another. So that's like one example. Yeah, yeah. Okay. So any,
把，把，呃，呃，一个到另一个。所以这就像一个例子。是，是。好。所以任何，
08:49
any scientific problem that you can map into this framework. So the, the input is to, to era,
任何你能 map 到这个 framework 里的科学问题。所以，这个，这个 input 是给，给 era，
08:57
is sort of this mapping, and the output is code, is that, well, sort of, I mean, the input is you,
差不多就是这种 mapping，输出是 code，是不是，嗯，差不多，我是说，输入就是你，
09:03
the way we've got it set up in, in the, the, uh, product, is you just start talking, right? And so
我们把它设置好的方式，在，在，那个，呃，产品里，就是，你直接开始说就行，对吧？然后
09:10
because a lot of times it's non-obvious to how do this mapping, although, you know, experts like
因为很多时候，怎么做这种 mapping 并不那么显然，尽管，你知道，像
09:14
Michael Brenner know how to do it. So we actually, um, a routine agent that actually helps you,
Michael Brenner 这样的专家知道怎么做。所以我们其实，嗯，有一个 routine agent，它其实会帮你，
09:19
it sort of talks to you, or to try to help you define what your scoreable tasks should be. So
它有点像是在跟你聊天，或者说试着帮你定义你的 scoreable tasks 应该是什么。所以
09:24
there's already, there's sort of an instance of Gemini sitting there trying to help you write a, uh,
已经，已经有一个 Gemini 的 instance 坐在那儿，试着帮你写一个，呃，
09:29
code. So that's actually sort of almost like an intermediate result. You start talking to it about
code。所以那其实差不多算是一个 intermediate result。你开始跟它聊
09:34
your problem, and it tries to produce essentially up a, uh, a Python notebook underneath that,
你的问题，然后它基本上会试着在那下面生成一个，呃，Python notebook，
09:40
that has, uh, that has a score, essentially a function with a, with a scoreable, uh, which
那个有，呃，那个有一个 score，本质上是一个 function，带有一个，带有一个 scoreable，呃，这个
09:44
essentially produces a score. And then it starts to mutate that notebook, uh, in a clever way,
本质上产生一个 score。然后它开始 mutate 那个 notebook，呃，用一种聪明的方式，
09:50
because again, it's Gemini. And it will try to sort of, uh, keep proposing code that tries to
因为再说一次，它是 Gemini。它会尝试，呃，不断提出 code，试图
09:55
maximize the, um, uh, the score. So what is different about this in just a generally a, a general
最大化那个，嗯，呃，那个 score。那么，这跟一个一般的，一个一般的
10:03
agentic system that can sort of optimize notebooks? Uh, right now it's a, uh, it's essentially,
agentic system 能够 sort of optimize notebook 有什么不同？呃，现在它是一个，呃，它本质上是，
10:10
it's own in modern 2026 parlance. We actually worked on this in 24 and 25, but the
它自己的，用 2026 年现代的说法。我们实际上在 24 和 25 年就研究过这个，但是
10:15
modern, modern parlance is kind of a specialized harness that, uh, runs an algorithm, uh, which in,
现代，现代的说法是一种 specialized harness，呃，运行一个 algorithm，呃，这在，
10:22
for era was Monte Carlo tree search. So essentially it's keeping, um, hundreds of thousands of
对于那个时代来说就是 Monte Carlo tree search。所以本质上它保持着，嗯，成千上万的
10:27
possible instances of notebooks. And then it selects one, uh, I can explain how it selects one,
notebooks 的可能实例。然后它会选一个，呃，我可以解释一下它是怎么选的，
10:34
and it decides, well, okay, what can I do? Gemini asked itself, what can I do to make that, uh,
然后它决定，嗯，好吧，我能做什么？Gemini 问自己，我能做什么来让那个，呃，
10:39
notebook be better? And then it will make a new one and test it and then put it back into the
notebook 变得更好？然后它会做一个新的，测试它，然后把它放回
10:44
candidate pool. So you can imagine the Kennedy pool is actually tree structured because it's, um,
candidate pool。所以你可以想象，Kennedy pool 实际上是树状结构的，因为它，嗯，
10:50
you know, every candidate possibly has some children. And what you do is you pick based on
你知道，每个候选者可能都有一些 children。而你要做的就是根据
10:54
something called the, uh, uh, it's actually, uh, uh, uh, fairly standing algorithm in, uh,
某个叫做，呃，呃，其实它是，呃，呃，呃，fairly standing algorithm 的东西，在，呃，
11:00
from reinforcement running called upper confidence bound UCB. So you essentially pick,
来自 reinforcement running，叫做 upper confidence bound UCB。所以你本质上是在选，
11:05
it's an optimistic algorithm. So it tries to estimate, say, what's the, what's the 95th percentile
它是一个 optimistic algorithm。所以它会尝试估计，比如说，那个，那个 95th percentile 是什么
11:10
outcome of mutation? It tries to estimate that and it picks the one with the highest bound,
mutation 的结果？它会试着估算这一点，然后挑 bound 最高的那个，
11:15
the highest optimistic bound. So no, there's a, it doesn't always pick the best performing notebook.
最高的 optimistic bound。所以，不，有个，它并不总是挑表现最好的 notebook。
11:20
It tries to predict like, what's the current performance plus two same of, of, of its
它会试着预测，比如，当前 performance 加上两个一样的，它的、它的、它的
11:25
guessed? And so it's always trying to, so it hunts around. So there's a, yeah, I recall basically,
猜测？所以它总是试着，所以它会到处搜寻。所以，有个，对，我记得基本上，
11:30
I recall, it's trying to make it's bet so that it, it most efficiently tries to make progress,
我记得，它是在下注，好让它，它尽可能高效地取得进展，
11:36
which isn't always greedily doing the best candidate. Sometimes it's the fifth best.
这并不总是 greedy 地选最好的 candidate。有时候是第五好的。
11:42
We're also play. We've also played around where, where it kind of recombines. It sort of takes
我们也在玩。我们还试过，就是，它会有点像 recombine。它会有点像把
11:46
ideas from two candidates and, and smashes them together and, um, tries to make a third candidate
两个 candidate 的想法，然后，然后把它们砸在一起，嗯，试着做出第三个 candidate
11:53
out of that. How does it see the initial candidate pool? Well, that's the amazing thing is that
从这当中出来。它怎么看待初始的 candidate pool？嗯，神奇的地方就在于
11:59
underneath Gemini is actually good at writing code. I mean, you just ask it, write me a thing,
底层上，Gemini 其实很擅长写 code。我的意思是，你只要让它，给我写个东西，
12:04
because you have, you have a textual description of the problem. It isn't just, oh, here's your
因为你有一个，你有一个问题的文本描述。它并不是那种，哦，这是你的
12:09
scoring function start. You say a textual description of the function, and you might give it,
scoring function，开始吧。你说一段关于这个 function 的文本描述，而且你可能还会给它，
12:13
in fact, we have, under some things, like, here are five papers that people tried to solve this
事实上，我们有些情况下会这样，比如，这是五篇人们试图解决这个
12:19
problem with. And it's kind of smart. It actually goes and reads the papers and will actually take a
问题的论文。它还挺聪明的。它真的会去读那些论文，然后真的会先
12:25
first stab at code. It might not be great, or it might, you know, sometimes it has bugs and it
试着写 code。可能写得不太好，也可能，你知道，有时候它有 bug，然后它
12:30
returns essentially minus infinity, but it will then try to mutate the code and say, oh,
本质上返回 minus infinity，但接着它会试着 mutate 这个 code，然后说，哦，
12:36
it'll try to make it be better. So it's, it's pretty cool. You don't actually have to give it,
它会试着把它变得更好。所以，这、这挺酷的。其实你并不一定要给它，
12:41
I mean, you can, if you want, give it a, some code starter code, but you don't have to. How many
我是说，你可以，如果你想的话，给它一个、一些 starter code，但你不一定非得给。你会启动多少个
12:46
agents are you spinning up? I guess maybe not agents, or how many different, you know, tree branches
agents？我猜也许不是 agents，或者说多少个不同的，你知道，tree branches
12:52
are you spinning up at each iteration? Oh, and at every iteration. Well, there's a trade off.
是你在每次 iteration 启动的？哦，而且是在每次 iteration。嗯，这里有个 trade-off。
12:56
You'd like to do a lot of parallel work, but if you do too much parallel work, you can't learn
你会想做很多 parallel work，但如果你做太多 parallel work，你就没法从
13:02
from previous things. So right now, we use about, the default is 10 parents, so you, so you, you try
之前的东西里学习。所以现在，我们大概用，default 是 10 个 parents，所以你会，所以你会，你会试着
13:08
to grow 10 leaves at a time. That seems to be about the right trade off. When you say you can't
一次长出 10 个 leaves。这似乎差不多是合适的 trade-off。当你说你不能
13:15
learn from previous iterations, that means that the orchestrator, or is there some sort of, yeah,
从之前的 iterations 里学习时，那是指 orchestrator，还是有某种，对，
13:21
what's the, that you said that there is some step, which is able to like,
那个什么，你刚说有个 step，它能，像是，
13:26
recombine or make decisions beyond just like the score. I mean, yeah, I guess maybe one of
recombine 或者做决策，而不只是根据 score。我是说，对啊，我猜也许其中一个
13:31
the other questions is, as a human, when you are doing some sort of a mail project, you don't just
其他问题是，作为人类，当你在做某种 mail project 时，你不会只是
13:36
look at, like, oh, there's this one metric that we're trying to optimize. Oftentimes, there's like
看着，哦，我们想优化的就这一个 metric。很多时候，还会有像
13:40
orthogonal metrics, sometimes even insights such as like just watching, you know, training curves
orthogonal metrics，有时甚至是一些 insights，比如只是看，你知道，training curves
13:46
can sometimes give you intuition about what's going on, or like looking into specific examples.
有时能让你对到底发生了什么有直觉，或者像是去看具体的例子。
13:51
Does it do any sort of intersection like this? Is there, well, it has, it has the history of,
它会做这种 intersection 吗？有没有，嗯，它有，它是有 history 的，
13:57
but why I meant why you can't do too many things in parallel is, is if you, if you have 10 parallel
但我的意思是，为什么你不能同时做太多 parallel 的事情，是因为，如果你，如果你有 10 个 parallel
14:04
searches at once, then what the number one can't actually see what numbers two through 10 or doing.
同时做多个 search，那第一个其实看不到第二到第十个在干什么。
14:11
So if you do a thousand at once, then you're using a huge amount of computation without a lot of
所以如果你一次做一千个，那你就会用掉大量 computation，却没有多少
14:16
cross-learning, whereas once you finish a little batch, you get the history. It's sort of, obviously
cross-learning；而一旦你跑完一个小 batch，你就能拿到 history。这有点，显然
14:22
you have to prove it so it doesn't blow up the context, but you get the history of what it was
你得证明它，这样不会把 context 撑爆，但你能拿到它当时在想什么的
14:27
thinking about as it was kind of writing the code and the results of the code. So it can learn
history，在它一边写 code 时，以及 code 的结果。所以它能从
14:33
from its previous attempts. Okay, and does it learn across? Oh, yes. Yes. Essentially, it's
之前的尝试里学习。好，那它会跨着学吗？哦，会。会。本质上，它
14:40
like one essentially shared context. So it is sort of thinking as it goes along. It's not like it's
就像是一个本质上共享的 context。所以它有点像边做边想。它不是像
14:47
a thousand different, completely independent branches. You're really pushing Gemini's like long
一千个不同的、完全独立的分支。你真的是在推动 Gemini 的 long
14:51
context ability. And you have to do, you have to do the right management and stuff. Yeah, yeah.
context 能力。而且你得做，你得把管理做对之类的。对，对。
14:56
Okay. Oh, that's cool. Going back to our day's question, the key point here being, the first
好。哦，那挺酷的。回到我们今天的问题，这里的关键点是，第一个
15:05
the first key goal is to, I guess, identify what specific score that you are trying to
第一个关键目标是，我猜，找出你具体想要
15:12
highlight. Sometimes that is, I agree, kind of the hardest part of the problem. And so I find it
突出的是哪个 score。有时候，我同意，这算是问题里最难的部分。所以我觉得
15:17
interesting that I'm not sure I'd always trust my agent to do that part. That part seems like the
有意思的是，我不确定我总会信任我的 agent 去做那部分。那部分看起来像是
15:24
more human task in the loop. Yeah. And often you have to be careful. In fact, a lot of what you do,
loop 里更偏人类的任务。对。而且经常你得小心。事实上，你做的很多事，
15:31
it's kind of, it's very meta. I guess everything we do is very sort of high level. You have to make
它有点算是，它非常 meta。我猜我们做的每件事都非常 high level。你得
15:36
sure that there's no one common thing as you come up with a scoring function or the agent does,
确保当你提出一个 scoring function，或者 agent 提出时，不存在某一个通用的东西，
15:41
or you do it together. And then the iteration finds a way to cheat or hold. Like, oh, no, I didn't
或者你们一起做。然后 iteration 会找到一种办法来作弊或者 hold。就像，哦，不，我不是
15:47
mean that. And so you have to go through and often sort of play and have a loop around it,
那个意思。所以你得反复过一遍，经常有点像在玩，围绕它形成一个 loop，
15:54
where you kind of iterate. Like, no, no, no, I didn't mean that. Or you have to tell it in its
在这个 loop 里你得稍微 iterate 一下。就像，不不不，我不是那个意思。或者你得在它的
15:58
instructions, okay, well, you know, don't, don't do this. So yeah, there's often iterations. And so
instructions 里说，好吧，嗯，你知道，别，别这么做。所以是的，经常会有 iterations。然后
16:03
you, even with a Gentek help, you don't necessarily get the, the right scoring function from day one.
你，即使有 Gentek 帮忙，也不一定能从第一天就得到，那个对的 scoring function。
16:08
And in fact, it's really neat because, I mean, in the old days, IE 2024 or something, you know, a lot,
而且实际上，这真的很妙，因为，我是说，以前，也就是 2024 年之类的，你知道，很多，
16:15
a lot of grad students would spend a lot of time like doing scientific software. And it's just so
很多研究生会花很多时间做 scientific software。而且它就是那么
16:20
much effort to write code at all that you kind of try maybe a few things or a few things that are
费劲，光是写 code 就要花很大力气，所以你大概只能试几件事，或者试几件……
16:27
very related. And then you sort of stop because you have to write your paper, you have to do your
非常相关。然后你差不多就停下来了，因为你得写论文，你得去做你的
16:31
next experiment. This thing is kind of underneath kind of relentless because it keeps trying,
下一个实验。这东西在底下有点不罢休，因为它一直试，
16:36
keeps trying, keeps trying. And so the people who use it are now spending all their time almost at
一直试，一直试。所以用它的人现在几乎把所有时间都花在
16:42
the right level, almost at the scientific creativity level. What does it mean to have a cost function?
对的层面上，几乎到了科学创造力的层面。有一个 cost function 到底意味着什么？
16:47
You know what I mean? And so that's almost like the essence of the scientific problem. You're not
你懂我意思吧？所以这几乎就是科学问题的本质。你现在不再
16:51
so much now in the details of, oh, oh, I have to import the CSV file or I have to get this database
是那么多地陷在细节里，哦，哦，我得 import 这个 CSV 文件，或者我得让这个 database
16:57
to work or whatever. It's you're now sort of thinking almost like deeply philosophically about your
能跑起来，或者随便什么。而是你现在有点像是在非常哲学地深入思考你的
17:04
actual scientific problem, not down in the grungy group of worrying about, you know, databases.
真正的科学问题，而不是陷在那种脏兮兮的琐事里，操心着，你知道，databases。
17:10
In fact, one cool thing the agent can do is actually suggest data, data sets to you like, oh,
其实 agent 能做的一件很酷的事，就是它居然会给你建议 data、data sets，比如说，“哦，我有个开心的念头，也许可以把这份 data set 拉进来做个 join。”
17:15
happy thought about maybe pulling in this data set and doing a join. And so it'll, it'll make
然后它就会，它就会给出建议，比如说，你知道的，你可以跟哪些 data sets 做 join，这还挺酷的。
17:21
suggestions about like, you know, data sets you can join with, which is kind of cool. Going back
回到你刚才说的，就是 agents 喜欢 hack 东西或者 reward hack 这件事。
17:25
to what you said a second ago, in terms of agents love to hack things or reward hack.
你有没有什么好玩的故事或者有意思的故事，你知道的，就是事情以很搞笑的方式彻底跑偏了？
17:31
Do you have any fun stories or interesting stories about, you know, where things were
天哪，我脑子一片空白。我知道其他人写过这个。我不确定自己有没有足够的细节，能把它那种好笑劲儿讲出来。
17:36
comedically blown off the rails? Boy, I'm blanking. I know other folks have writing to it. I don't know
但确实会。你会有点被惊到。我不知道自己有没有真的很具体的例子，抱歉，我脑子一片空白。
17:43
if I have enough details to sort of say to express the comedy. But it does. You kind of get
没事，没关系。
17:50
surprised. I don't know if I have any really concrete, sorry, I'm blanking. No, it's fine.
惊讶。我不知道我有没有任何真的很具体的，抱歉，我脑子一片空白。不，没事。
17:57
Yeah, I would like to think of machine learning. It's sort of like the old genie stories.
对，我会把 machine learning 想象成……它有点像那种古老的灯神故事。
18:01
That's right. That's right. That's why you wish more because you're going to get it.
没错。没错。所以你会许更多愿望，因为你会得到它。
18:06
That's right. And and you have that. And that happens very much with this. You have to be careful.
没错。而且，而且你确实有这种情况。而且这种情况在这里非常常见。你得很小心。
18:13
But on the other hand, it has some knowledge. The nice thing is that sort of Gemini knows a lot
但另一方面，它也有一些知识。好的一点是，Gemini 这种东西大概知道很多
18:21
about many things, sort of more than anyone I'm person can do. So it at least knows, especially
关于很多事情，大概比任何一个人能做的都多。所以它至少知道，尤其是
18:26
if you point papers point, you know, like here, they're here, five papers, then try to do this in
如果你指出 paper，指出，你知道，就像这里，它们在这儿，五篇 paper，然后试着用
18:31
some way. So to some extent, it does have that genie feel, but to some extent, it also sort of
某种方式做这个。所以某种程度上，它确实有那种灯神的感觉，但某种程度上，它也多少有点
18:37
does say things. This is why I remember the whole idea of sort of evolutionary coding. It's
会说出一些东西。这就是为什么我记得所谓 evolutionary coding 的整个想法。它
18:44
been around since the 70s. Everyone's loved to do that. Like, oh, let's mutate list code or whatever
从 70 年代起就有了。大家一直都很喜欢这么干。比如，哦，我们来 mutate 一下 list code 或者别的什么
18:51
to do things. But the reason why it just hasn't taken off is that random mutation in code space is
来做各种事情。但它一直没真正起飞的原因是，在 code space 里做 random mutation 基本上
18:57
pretty much worthless. I mean, just like, well, just like DNA, it's sort of like, you know, most
几乎毫无价值。我是说，就像，嗯，就像 DNA 一样，有点像，你知道，大多数
19:01
most things are harmful. So here, it's like, oh, no, no, we can actually find, it sort of knows,
大多数东西都是有害的。所以在这里，就变成，哦，不不不，我们其实能找到，它有点知道，
19:07
sort of underneath and knows interesting gradients to try, which is why the thing works that the
有点像在底层，知道有哪些有意思的 gradients 可以试，这就是为什么这东西能 work，因为那个
19:12
underlying loop itself is an AI. So yes, it can, it can maybe overfit and have funny sort of genie
底层的 loop 本身就是一个 AI。所以是的，它可能会 overfit，会有那种好玩的 genie
19:18
problems like you, but it also has some a lot of sanity because it sort of, it knows about the world
problems，就像你一样，但它也有很多 sanity，因为它有点，它了解这个世界，
19:24
and it has world knowledge in it. So the paper, though, you were doing Gemini 2.5. And I think,
而且它里面有 world knowledge。不过那篇 paper，你当时做的是 Gemini 2.5。而且我觉得，
19:30
you know, Gemini has advanced quite a bit. Do you have metrics or have you, you know, this is a
你知道，Gemini 进步挺大的。你有没有 metrics，或者你有没有，你知道，这是一个
19:35
tool that you're continuously using and it sounds like you're improving? Yeah. And I'm wondering,
你一直在用的工具，而且听起来你在不断改进？对。然后我在想，
19:39
like, do you internally have you seen, like, almost like a phase transition in how effective this
就是，你们内部有没有看到，就像，几乎像是一种 phase transition，在这个
19:45
tooling has been? How dramatic has the improvement been over the last, like, I guess year or two?
tooling 的有效性上？过去，大概，我猜一两年里，改进有多剧烈？
19:51
Oh, well, I mean, year or two. Yeah. Amazing. You know, there's every, even every half version of,
哦，嗯，我是说，一两年。对。太惊人了。你知道，每一个，甚至每一个半个版本，
19:58
of, I mean, essentially, I think it would have been impossible under Gemini 2.5.
我是说，基本上，我觉得在 Gemini 2.5 下这是不可能的。
20:02
And yeah, so I think so and others wouldn't want to work. So you started at 2.5. And that was like,
而且对，所以我也这么觉得，其他人也不会想用。所以你是从 2.5 开始的。那就像是，
20:07
just, oh, no, we've been trying to experiment with these things actually for a while. Okay. And
只是，哦，不，其实我们已经试着实验这些东西有一阵子了。好。然后
20:12
things just weren't working. And then they started to work. And then now they're just amazing. So
一开始就是不管用。然后它们开始能用了。然后现在它们简直太棒了。所以
20:16
the progress on Gemini major versions has just been stunningly amazing. Yeah. Yeah, I think this
Gemini 大版本上的进展简直惊艳得不可思议。是啊。是啊，我觉得这
20:22
is an experience a lot of people have been having worse things were just seemed impossible. Whatever
是很多人一直在经历的一种体验，更糟的是，那些事情曾经看起来根本不可能。不管什么
20:27
are suddenly becoming magically useful. Yes, really quickly. Yeah. So, so if people are, I even say
突然就变得神奇地有用了。是的，真的很快。是啊。所以，所以如果人们，我甚至会说
20:35
this to scientists, because there's some people like, oh, I tried whatever 2.0. And I didn't like,
这话对科学家说，因为有些人会说，哦，我试过某某 2.0。然后我不喜欢，
20:39
oh, yeah, that was a long time. That was a year ago. That was like a long journey. It was a
哦，是啊，那是很久以前了。那是一年前了。那就像一段漫长的旅程。那是一次
20:43
journey ago. Right. Yes. And in fact, all the, all the, we even have one of the preprints where we've
很久以前的旅程。对。是的。而且事实上，所有这些，所有这些，我们甚至有一篇 preprint，里面我们
20:49
sort of combined a hero with anti-gravity. And that, you know, the whole, that whole harness of
算是把 hero 和 anti-gravity 结合在了一起。而且，你知道，整个，整个 harness 的
20:55
anti-gravity is pretty amazing too. And it's that, that's the one where you can sort of pull in lots
anti-gravity 也挺神奇的。就是那个，那个你可以拉进来很多
21:00
of papers and it can write lots of code for you. And so, yeah. Is that publicly available? Or is that
论文，而且它还能帮你写很多代码。所以，是啊。那个是公开可用的吗？还是说
21:06
the anti-gravity? Yeah, yeah. Well, anti-gravity certainly publicly. Oh, sorry, sorry. The
是 anti-gravity？对，对。嗯，anti-gravity 肯定是公开的。哦，抱歉，抱歉。那个
21:11
era plus, you can take your time. Not yet. Okay. Okay. I find this area really fascinating,
era plus，你可以慢慢来。还没。好。好。我觉得这个领域真的特别有意思，
21:16
because like you said, it's been, there's been some form of code mutation out there since the
因为就像你说的，从那时候起，一直都有某种形式的 code mutation，从
21:22
dawn of computer science basically. The canonical problem is sort of the overfitting or multiple
计算机科学诞生之初就基本如此了。那个典型问题差不多就是 overfitting 或者 multiple
21:29
hypothesis testing problem. I think which is maybe a little bit better matched to the problem where
hypothesis testing 问题。我觉得这可能更贴合那种问题，就是
21:34
you're basically my hypothesis now that this algorithm work. Now my hypothesis is, and so that
你基本上就是，我的假设现在是这个 algorithm 能行。现在我的假设是，然后所以那个
21:40
you run the risk that sort of it has exponentially exploded, right? Because now suddenly,
你就会有这种风险，就是它某种程度上已经 exponentially 爆炸了，对吧？因为现在突然之间，
21:45
I have like these, it's like hyper hyper parameters that I'm optimizing. And so that you have this
我就像有这些，有点像我在 optimizing 的 hyper hyper parameters。于是你就会有这种
21:51
explosion of state space that you're exploring. And so that it seems much easier to,
state space 的爆炸，你在探索的那个 state space。于是看起来就容易得多，
21:57
to sort of overfit to a problem. What are your thoughts about that? Because on the other hand,
有点像对一个问题 overfit。你对此怎么看？因为另一方面，
22:02
empirically, my experience, I even tried the sort of open source version of era. I kind of
empirically，以我的经验，我甚至试过那种 open source 版本的 era。我有点
22:08
strapped it into cloud. And it's running right now, so I can't tell you how well it's working.
把它接进了 cloud。它现在正在跑，所以我没法告诉你它效果到底怎么样。
22:12
I'm curious. Yeah, I'll let you know. But I'm just curious to know, this is a question that's
我很好奇。嗯，我会告诉你的。但我就是好奇，这是一个一直
22:18
been in my mind about just general AI for science. And what are your experiences with this sort of on
在我脑子里关于 general AI for science 的问题。那你在这种 on 上有什么经验
22:24
the sort of on the ground? I guess there's two questions sort of embedded in your question,
那种实际落地层面的？我猜你的问题里其实藏着两个问题，
22:29
I think, right? Because when you talk about multiple hypothesis testing, right? There's predictive
我觉得，对吧？因为当你谈到 multiple hypothesis testing 的时候，对吧？有 predictive
22:35
models and there's descriptive models, right? You know, you know, I've been doing machine learning
models，也有 descriptive models，对吧？你知道，你知道，我做 machine learning
22:40
for a long time, statistics. That's actually a really good point though. Do you mind explaining
已经很久了，statistics。不过这其实是个非常好的点。你介意解释一下
22:44
that? I'm not sure that's something that everyone with general audience would be familiar with.
吗？我不确定这是不是普通受众都熟悉的东西。
22:48
Right. Especially in modern days, I think people are trying to sort of obscure it too.
对。尤其是现在，我觉得人们也在试图把它搞得有点模糊。
22:53
If you started with LLNs, I'm not sure that distinction would be meaningful. That's right.
如果你是从 LLNs 开始，我不确定这个区分还有没有意义。没错。
22:57
Yeah. So a predictive model is like, let's say you just have you have some inputs and you have
是的。所以 predictive model 就像，假设你只是有一些 inputs，然后你有
23:02
some outputs and you just, I just want to build a piece of code that tries to just have the lowest
一些 outputs，然后你只是，我只是想写一段代码，让它尽量在某个 data set 上有最低的
23:08
error rate on some data set. That's just a statistical model, right? A descriptive model is actually
error rate。那只是一个 statistical model，对吧？而 descriptive model 其实才是
23:14
what science is trying to get to, which is, okay, it should be able to extrapolate because it has
科学想要达到的东西，也就是，好吧，它应该能 extrapolate，因为它有
23:20
sort of the physics or the actual some description of reality that's captured within it.
某种 physics，或者说某种对现实的描述，被包含在它里面。
23:26
And then you can use it to extrapolate because it's sort of like, you know, yes, Newton thought
然后你就可以用它来 extrapolate，因为它有点像是，你知道，对，Newton 想的是
23:31
apples and gravity, but gravity isn't actually about apple right if you just fit
苹果和 gravity，但 gravity 其实并不是关于苹果的，对吧？如果你只是 fit
23:35
if you take into the machine, you know, the 17th century machine learning model, like,
如果你把它放进 machine 里，你知道，17 世纪的 machine learning model，就像，
23:38
oh, apples will fall, but how about planets? You know, I don't know. I have no data about planets.
哦，苹果会掉下来，但行星呢？你知道，我不知道。我没有关于行星的 data。
23:43
So, who knows what they do, right? So when you say extrapolate, okay, I didn't realize we're
所以，谁知道它们会做什么呢，对吧？所以当你说 extrapolate 的时候，好吧，我都没意识到我们
23:47
kind of going to go ahead. No, no, no, no, no, no, I'm curious. Okay, so when you say
有点要直接往下讲了。不不不不不不，我很好奇。好吧，所以当你说
23:50
extrapolative, so there's different ways I can think about this. One of them is you said a model of
extrapolative 的时候，所以我可以从不同角度来想这件事。其中一个是你说的 model of
23:54
physics or a model of the world. Are you talking about introducing explicit priors that, you know,
physics，还是世界的 model。你是在说引入 explicit priors，那种，你知道，
24:00
she based upon some human intuition, or maybe in this case, LLM intuition, or are you talking about
她基于一些人类直觉，或者也许在这种情况下，是 LLM 直觉，还是你在说
24:05
this is the physics is actually learned by the model or the underlying process of the world is under
这是说 physics 实际上是由 model 学到的，还是说世界的 underlying process 是 under
24:12
well, the distinction between those is a little bit blurry, right? Because when a physicist or
好吧，这两者之间的区别有点模糊，对吧？因为当一个物理学家或者
24:18
scientist comes, they use their intuition and they were about, or maybe even more than intuition,
科学家来的时候，他们会用他们的直觉，而且他们大概……或者也许甚至不只是直觉，
24:23
like essentially there's maybe a solid pile of facts that they know about the world.
就像，本质上，也许有一大堆他们了解的关于世界的扎实事实。
24:28
And then they make sure that whatever model they build is sort of consistent with what's known.
然后他们确保，不管他们构建什么 model，都和已知的东西大致一致。
24:35
Currently, in era, it is, it's sort of it is LLM intuitions. And that's what I was trying to say
现在这个时代，它差不多就是，它就是 LLM 直觉。这就是我想说的
24:40
about having a good gradient underneath that, that especially if you point it at existing papers,
关于在那底下有一个好的 gradient，尤其是如果你把它指向现有的论文，
24:47
it will try to build models that are kind of sane underneath because it's, again,
它会试着在底下构建出某种程度上还算靠谱的 model，因为它，还是那句话，
24:52
especially if you give it guidance, like, oh, be sure to incorporate this in this or look at
尤其是如果你给它指导，比如，哦，一定要把这个纳入这个，或者看看
24:56
these papers and get these things. So it will, there is a, you can introduce a bias towards
这些论文，把这些东西搞明白。所以它会，你可以引入一种 bias，针对
25:04
certain model choices and it will have a bias because it just, it's own little sort of world
某些 model 选择，而它会有 bias，因为它就是，它自己那个小小的世界
25:09
knowledge is accumulated inside of inside of itself in, in sort of pre-training.
知识是在它自身内部、在自身内部，在某种程度上，在 pre-training 里积累起来的。
25:15
Can you give us an example of what that might look like? Is it modeling something in a way where
你能给我们举个例子，说明那可能是什么样子吗？它是不是以某种方式对某个东西建模，其中
25:19
there's different, for example, if you're doing something with partial differential equations,
会有不同的，比如，如果你在处理 partial differential equations，
25:23
there's these formalisms people have like neural operators, for example, or where you can embed a,
会有这些人们有的 formalisms，比如 neural operators，或者你可以嵌入一个，
25:29
you can encame an encoded differential equation in some sense, or I think that's called physics
你可以在某种意义上把一个编码过的 differential equation 封装进去，或者我觉得那叫 physics-informed
25:35
and form neural networks. The thing is like one for fluid and kind of PDE type modeling systems,
neural networks。这东西就像一种用于 fluid 和某种 PDE type modeling systems 的，
25:41
or for, let's see molecular systems, there's oftentimes this idea about equivariants.
或者对于，让我想想，molecular systems，经常会有这种关于 equivariants 的想法。
25:46
I mean, the model is like picking up on these, like, you know, tricks which have been developed
我的意思是，模型就像是在捕捉这些，比如说，你知道的，已经被开发出来的技巧
25:50
in literature, or are they adding some weights for some physical prior or something?
在文献里，还是说他们在为某种 physical prior 之类的加一些 weights？
25:55
Yeah, kind of what does that look like? When it introduces, like, a physics,
对，那大概是什么样子？当它引入，比如说，一个 physics，
25:59
when it introduces some sort of, you know, knowledge like that.
当它引入某种，你知道，那样的 knowledge 时。
26:03
I don't know if I have enough data to sort of say, oh, you know, 73% of the time it does this.
我不知道我有没有足够的 data 来大概说，哦，你知道，73% 的情况下它会这么做。
26:08
Okay.
好。
26:08
But especially when you point it at existing papers, it will try to, in fact, it will do very well at
但尤其是当你把它指向现有的论文时，它会试着，事实上，它会非常擅长
26:16
adapting the methods that are described in the papers for the problem. In fact, it will do
把论文里描述的方法适配到这个问题上。事实上，它会做得
26:23
an amazing job. You can actually often just recreate or reverse engineer paper. That's again,
非常出色。你实际上经常可以直接复现或者 reverse engineer 论文。这又是，
26:29
what Michael Brinter actually likes to do this. He'll say, oh, that sounds like an interesting paper.
Michael Brinter 其实就喜欢这么做。他会说，哦，这听起来像一篇挺有意思的 paper。
26:34
Oh, we did this actually for the, we had this thing. It was actually kind of a hack. I suggested
哦，这个我们其实做过，我们有过这么个东西。那其实算是个 hack。我提议
26:39
this to Michael where there's this one MIT professor who he came up with some code to do essentially,
把这个跟 Michael 说，当时有个 MIT 教授，他写了一些 code，基本上能做这么件事，
26:49
if you have a rooftop, if you have a rooftop with a fixed area, and you want to maximize the
如果你有一个屋顶，如果你有一个面积固定的屋顶，而你想最大化
26:54
amount of solar power you capture over a day, sort of solar energy, you can build up, which is,
一天里你捕获的 solar power 量，类似 solar energy，你可以往上建，这当然
26:59
of course, captures more sunlight and you can sort of build. You can have it design a widget,
会捕获更多阳光，而且你可以大概建。你可以让它设计一个 widget，
27:03
you know, with involving mirrors or struts or solar panels to sort of stock it, whatever
你知道，涉及 mirrors 或 struts 或 solar panels，来 sort of stock it，随便
27:10
you know, out and goals or sizes you like. And usually with a maximum height,
你知道，out and goals 或你喜欢的尺寸。而且通常还有个最大高度，
27:16
and then try to figure, let it sort of explore that design space. And I believe, Michael,
然后试着弄清楚，让它去探索一下那个 design space。而且我相信，Michael，
27:21
we can ask him, I believe it actually just, I don't think he actually installed the simulator.
我们可以问他，我觉得它其实只是，我不觉得他真的装了这个 simulator。
27:26
I think, I think the code just reproduced the code because it has like a coding agent inside of it,
我觉得，我觉得它就是直接把代码复现出来了，因为它里面像是有个 coding agent，
27:33
just reproduced the code with the paper and sort of figured it all out. So yes, it's very good,
就靠着那篇论文把代码复现出来，然后把一切都搞明白了。所以是的，它非常厉害，
27:39
especially if given a pointer to what other people have done, it's very good at kind of like,
尤其是如果有人给它一个 pointer，告诉它别人做过什么，它就特别擅长那种，
27:45
oh, I haven't seen it, try it, equilibrium modeling that that can get very hairy if you know
哦，这个我没见过，试试看，equilibrium modeling 那玩意儿，如果你知道
27:50
about clipped, clipped Gordon coefficients. It's pretty fun. Yes. So I don't know if it'll do
关于 clipped，clipped Gordon coefficients，那会变得非常棘手。这挺有意思的。是的。所以我不知道它会不会
27:56
the trick or varying it stuff, but it actually knows a lot. I remember actually when
搞定这一手，或者把它变来变去之类的，但它其实懂的很多。我记得其实当
28:02
Gemini 2.5 came out. I know this is not exactly about era, but I remember sort of thinking,
Gemini 2.5 出来了。我知道这并不完全是在聊 era，但我记得我当时大概在想，
28:07
oh, this is a new world when the day 2.5 came out because he said, hey, Gemini 2.5, can you write
哦，2.5 出来的那天，这是一个新世界，因为他说，嘿，Gemini 2.5，你能写
28:12
me some boost decision tree code? And it did. And it worked. It just did. Yes. And I said, you know,
给我一些 boost decision tree 代码吗？它做到了。而且能用。它就是做到了。是的。然后我说，你知道，
28:19
yeah, this is this is a new world. So yes, I think it's before to loop back to your question.
是的，这、这，这是一个新世界。所以是的，我觉得最好还是回到你的问题。
28:25
I think, yes, if you give it sort of guidance about, oh, you know, it's important to put this kind
我觉得，是的，如果你给它一些指导，比如说，哦，你知道，重要的是要把这种
28:29
of thing in, it will. And so it won't necessarily at least not that we've seen discover completely new,
东西放进去，它就会。所以它不一定会——至少我们没见过——发现完全新的东西，
28:38
like if you didn't know about clipped Gordon, go, I mean, you know, you didn't know about something.
就像如果你不知道 clipped Gordon，go，我是说，你知道，你不知道某个东西。
28:42
It won't completely discover a new kind of physical model from scratch, but it will certainly,
它不会完全从零开始发现一种新的 physical model，但它肯定会，
28:48
if you tell it about interesting constraints about the world that are known, it will certainly follow
如果你告诉它一些关于这个世界、已知而且有意思的 constraints，它肯定会遵循。
28:53
I don't know. So there's so room for humans for the next year or two. Oh, in fact, there's
我不知道。所以接下来一两年，人类还是有很大空间的。哦，其实，还有
28:58
going back, I think there's totally room for humans because I don't know, I mean, we have
往回看，我觉得人类完全还有空间，因为我不知道，我是说，我们有
29:02
co-scientist that tries to help you come up with sort of hypothesis generation. But really,
co-scientist，它会试着帮你进行某种 hypothesis generation。但说真的，
29:10
I still haven't seen sort of the creativity and the philosophy and sort of the sort of
我还是没看到那种创造力、那种哲学，还有那种、那种
29:17
the careful rigor. You totally need humans. I don't see humans going away. They can make
那种细致严谨。你绝对需要人类。我不觉得人类会消失。他们能提出
29:24
the strange suggestions and I've used co-scientists actually for an interesting problem in geochemistry
那些奇怪的建议；而且我其实用过 co-scientists 来处理 geochemistry 里一个有意思的问题
29:29
and I learned about a new kind of ion. I didn't realize happened in magma, but I, and so it'll tell
然后我了解到一种新的 ion。我之前没意识到这会在 magma 里发生，但我，所以它会告诉
29:36
you interesting things and you'll learn stuff, but I don't think it sort of substitutes for you
你能做有趣的事情，也会学到东西，但我不觉得这能在某种程度上替代你
29:40
in creativity. You know, going back, we were just talking about two years, two point, you had
在创造力方面。你知道，往回看，我们刚刚还在聊两年，two point，你有
29:45
Gemini 2.0 to 2.5 and this was like, you already saw a leap and now it's been another year or two
Gemini 2.0 到 2.5，而这就像，你已经看到了一次跃升，现在又过了一两年
29:51
and now it's 3.5 and, you know, you're saying this is working much better. I mean, whenever you look
而现在到了 3.5，而且你知道，你说它现在好用多了。我的意思是，每当你看
29:56
at a graph, you know, you can, if something looks like an exponential, it can either, you can either
一张图的时候，你知道，你可以，如果某个东西看起来像 exponential，它可能要么，你可能要么
30:00
be in a sigmoid or you can be at the beginning of a takeoff, right? I guess every exponential turns
处在 sigmoid 里，要么处在 takeoff 的开端，对吧？我猜每个 exponential 最终都会
30:04
into a sigmoid eventually. Every exponential, yeah. But the question is like, where are we on that?
变成 sigmoid。每个 exponential，对。但问题是，我们现在处在那个过程的哪里？
30:09
I mean, I guess I'm a big believer in sort of the whole jagged.
我的意思是，我想我是那种很相信整个 jagged 这一套的人。
30:13
I'm a frontier. Yeah, the jagged frontier. And so certainly, at least what I see, I mean,
我在前沿。对，jagged frontier。所以当然，至少我看到的是，我是说，
30:17
I don't know what's going to happen a couple of years, but yes, there's some big spikes out in
我不知道几年后会发生什么，但没错，在编程能力、获取知识和寻找相关内容方面，确实存在一些大的尖峰和不均衡。这很巨大也很棒，我觉得这对科学家来说是好事。到目前为止，在严谨性方面可能还差一些，我们可以聊聊像国际甲基苯丙胺之类的，还有数学总体上，但就哲学和创造力而言，我觉得还是不太行。也许它会，也许一切都会，有些人在说什么都会膨胀然后过去，但我仍然看到很多非常强烈的不均衡。所以我能看到，嗯，也许它会变成一种品味的变体，但品味就像是另一边，对吧？就像是严谨性。就像是
30:22
jaggedness in terms of coding ability and just gathering knowledge and finding related things.
在 coding 能力、收集知识和找相关信息方面的 jaggedness。
30:28
And that's huge and wonderful, which I think is great for scientists. So far, it's kind of less in
这非常巨大又很美妙，我觉得这对科学家来说很棒。到目前为止，它在
30:33
terms of rigor and, we can talk about things like the international methamphetamine, but
严谨性方面还有点不够，而且，我们可以聊像 international methamphetamine 这样的东西，但
30:41
and math in general, but in terms of sort of philosophy and creativity, I think it's still kind
以及一般意义上的 math，但就哲学和创造力来说，我觉得它还是有点
30:46
of not. And maybe it'll, maybe everything will, some people are saying everything to an inflate
不行。也许它会，也许一切都会，有些人说一切都会膨胀
30:52
and pass, but I'm still seeing a lot of very strong jaggedness. So I can see, well, maybe it'll get
然后通过，但我还是看到很多非常强的 jaggedness。所以我能看出，好吧，也许它会变得
30:57
to be extra good at coding and extra good at fitting models and extra good at making suggestions
特别擅长 coding，特别擅长 fitting models，还特别擅长提建议
31:02
and things. But I don't know so far, not so far you need the humans. Yeah, I want to get back to the
以及诸如此类的事情。但到目前为止我不知道，还没到需要人类的地步。对，我想回到那个
31:07
question about the multiple hypothesis test. Sorry, sorry. Totally, totally love the tangent.
关于 multiple hypothesis test 的问题。抱歉，抱歉。真的，真的超喜欢这个跑题。
31:15
Multi hypothesis testing is when you have a descriptive model, and you're saying this is the way
Multi hypothesis testing 就是当你有一个 descriptive model，然后你说这就是
31:19
the world works, and you have a bunch of data, and you take a billion darts and you throw, and so
世界运作的方式，然后你有一堆数据，你拿十亿支飞镖扔出去，于是
31:25
you have to be careful, there's something called false discovery rate. And so the question is,
你得小心，有个东西叫 false discovery rate。那么问题就是，
31:30
is this finding descriptive models, or is this finding sort of predictive models? And fundamentally,
这是在发现 descriptive models，还是在发现某种 predictive models？而从根本上说，
31:37
the scientist is there to make sure that whatever is saying is descriptive. We haven't been able
科学家在那儿就是要确保无论说的是什么，都是 descriptive 的。我们还没能
31:45
to make a system so far out of these pieces that can really sort of discover completely new
用这些部件造出一个系统，让它真的能在某种意义上发现全新的
31:52
physics or completely new science, but this is sort of a power tool to help you discover completely
物理，或者全新的科学，但这有点像一种电动工具，帮你发现全新的
31:57
new science. So, so maybe I'm trying to unask your question of a multiple. But I think this gets
科学。所以，所以也许我是在尝试取消你那个 multiple 的问题。但我觉得这正好触及
32:03
at the heart of it. Yes, yes. But then you're saying, what about just pure over to, okay,
它的核心。对，对。但接着你是在说，那纯粹的 over to 呢，好吧，
32:06
let's set aside, it's not trying to figure out a descriptive model of the world that's still up
先放一边，它不是要试着找出一个描述世界的模型，那仍然
32:11
to the scientist. But what about just plain old overfitting? Yes, you have to be very careful,
得由科学家来做。但老掉牙的 overfitting 呢？对，你必须非常小心，
32:17
because it's a power tool, it can, I shouldn't probably say, it can slice your fingers off.
因为它是个电动工具，它可能，我大概不该这么说，它可能把你的手指切掉。
32:21
You know what I mean? It's a very careful, and you have to be very rigorous. In fact, now,
你懂我意思吧？这得非常小心，而且你必须非常严谨。事实上，现在，
32:25
you have to be more careful, more rigorous to not fool yourself. You really, really need to be
你必须更加小心、更加严谨，才能不欺骗自己。你真的、真的需要
32:30
just excruciatingly careful about having very hidden, hold out sets you don't look at. You have
极其小心地保留那些非常隐蔽的、你不去看的 hold out sets。你必须
32:37
to be just super, super rigorous to make sure that you don't completely, because it is a total
超级、超级严谨，确保你不会完全——因为它完全是一个强大的工具。所以问题在于我怎样才能不切到自己的手指，答案是
32:43
power tool. So the question to how do I not slice my fingers off is you need to use the same
你需要使用同样的技术，但要对它们非常小心。是的。这是一个非常清晰的回答。是的，我其实不
32:49
techniques, but be very careful with them. Yes. That's a very clear answer. Yeah, I actually don't
觉得我听过任何嘉宾说过这个。是的，这就像，我觉得这是一项非常重要的技能，
32:57
think I've heard any guests say that. Yeah, it's like it is, I think, a very important skill,
也许是新领域里最重要的技能之一。人们经常谈论taste，但也许这
33:01
maybe one of the most important skills in the new. People talk a lot about taste, but maybe this
是taste的一个变体，但taste像是另一面，对吧？就像是严谨。就像是
33:07
is a variation of taste, but the taste is like the other side, right? It's like the rigor. It's like
是品味的一种变体，但品味像是另一边，对吧？它就像严谨性。它就像
33:12
the yes, yes, in fact, if anything, yeah, taste or rigor, which one's more important? Well,
那个，对，对，其实，如果非要说的话，是啊，taste 还是 rigor，哪个更重要？嗯，
33:17
I don't know. I think people, at least the way I am viewing it is, I mean, the aspect of the
我不知道。我觉得，人们，至少从我目前看这件事的方式来说，我是说，从……这个方面来说，
33:25
researchers are software developers, which there's a lot of overlap in a lot of fields.
研究者就是 software developers，这在很多领域都有很多重叠。
33:30
I am seeing that software engineers are, it's almost like, obviously, there's a lot of concern,
我看到的是，software engineers 几乎就像，显然，有很多担忧，
33:37
like, oh, no, what am I going to do? Coding seems to be getting automatic. So I think there's
比如，哦不，我该怎么办？coding 似乎正在变得自动化。所以我觉得这里面有
33:43
sort of both. I think there's a lot of people get pulled into. Well, I'll be the creative source.
某种两者兼有。我觉得很多人会被卷进这种想法：嗯，那我来当创意来源。
33:49
So I'll try to figure out new science. I'll try to figure out new products. I'll try to
所以我会试着搞清楚新的科学。我会试着搞清楚新的产品。我会试着
33:53
sort of really be very, very creative. And again, I'm strong believe that I don't think that's
真正地、非常非常地有创造力。而且再说一遍，我坚信，我不觉得那是
33:58
going to go away. There's also people sort of pull towards rigor. Like, oh, I want to make sure
会消失的。也有那种倾向于严谨的人。就像，哦，我想确保这个不会崩溃。我想确保这个能扩展。我想确保这个没有错。我觉得你两者都需要。而且我觉得你需要在这两方面都非常擅长的人，但他们不一定非得是同一批人。但没错，我觉得你需要，我觉得这甚至比科学更广泛，就像软件工程演变一样。是的，会是，你知道的，带来创造力的人和带来严谨的人。我觉得这些会是某种锚点。我看到的其他一些事情是外面的相关工作。有一个非常酷的 leaderboard，你知道的，就像 call leaderboard、agent leaderboard，用于科学问题的，来自 Stanford。
34:02
this doesn't, this doesn't crash. I want to make sure this scales. I want to make sure this isn't
这个不会，这个不会 crash。我想确保它能 scale。我想确保这个不是
34:06
wrong. I think you need both. And I think you need people who are really good at both, but they
错的。我觉得你两者都需要。而且我觉得你需要那种两边都很擅长的人，但他们
34:12
don't necessarily have to be the same people. But yes, I think you need, I think this is even broader
不一定是同一批人。但没错，我觉得你需要，我觉得这甚至比
34:16
than science, just as sort of software engineering evolves. Yeah, it'll be, you know, the people who
science 还更广，就像 software engineering 演化那样。对，它会，你知道，是那些
34:23
bring the creativity and the people who bring the rigor. And I think those will be sort of anchors.
带来创造力的人，以及那些带来严谨的人。而且我觉得那些人会像是某种锚点。
34:29
Some other things that I've seen are related work out there. There's a really cool leaderboard
我看到的另外一些东西是外面的 related work。有一个特别酷的 leaderboard
34:37
for, you know, like call leaderboard, agent leaderboard for scientific problems from Stanford.
用来，你知道，比如 call leaderboard，还有来自 Stanford 的 scientific problems agent leaderboard。
34:43
I don't know if you're familiar with it. It seems like a really interesting idea to me to have,
我不知道你熟不熟悉这个。对我来说，能有这么个东西真的挺有意思，
34:49
you know, sort of different agents kind of competing on the leader. So it seems like
你懂的，就是不同的 agents 在 leader 上竞争之类的。所以看起来
34:55
if you squandle a little bit, what AI is doing is kind of a leaderboard that's internal
如果你稍微眯眼看一下，AI 在做的其实有点像一种内部的 leaderboard
35:00
and it's recombining ideas. Whereas, what are your thoughts about this? And do you,
而且它在重新组合各种想法。那你怎么看这个？还有你，
35:04
is that like a thing that you guys are working on? And is there problems with that or advantages
这是你们正在做的东西吗？这有什么问题或者优势
35:10
to that? Ironically, you know, the whole era project actually started because people
吗？讽刺的是，你懂的，整个 era project 其实一开始是因为人们
35:16
and I realized Kaggle is actually part of Google. Oh, yeah. And so it was called the Auto-Kaggle
和我意识到 Kaggle 其实是 Google 的一部分。哦，对。所以它叫 Auto-Kaggle
35:21
Problem. So it was actually like, that's what it was. Let's try to have a system that can sort
Problem。所以其实就像，事情就是这样。我们试着搞一个能 sort 的系统
35:29
of, you know, win at Kaggle competitions. So that's sort of why it sort of has the shape that
就是，你知道，在 Kaggle 比赛里赢。所以这差不多就是为什么它大概会有那种形态
35:34
that's sort of how the project started. And it goes back to sort of overfitting, right? If you've
这差不多就是这个项目开始的方式。而且它又回到差不多 overfitting，对吧？如果你
35:39
ever actually competed in a Kaggle competition. I have done Kaggle competitions and
真的参加过 Kaggle 比赛。我做过 Kaggle 比赛，而且
35:44
or I've done one. It is a really interesting phenomenon because there's this overfitting is like
或者说我做过一次。这真是个很有意思的现象，因为这种 overfitting 就像
35:51
rampant. Yeah, yeah. And it's really impressive how people can overfit
泛滥。是啊，是啊。而且真的很让人印象深刻，人们能 overfit
35:55
to certain data sets in a way that is, yeah. Or even we had a, we have a fun project
到某些 data sets 上，以一种，是的，方式。或者甚至我们有一个，我们有一个好玩的项目
36:00
thing to talk about. Or if you like, that tries to mitigate Contrails. Check out Trails. Yeah,
要聊的东西。或者如果你愿意，它试图缓解 Contrails。去看看 Trails。是啊，
36:05
about that if you want. And we had a Contrail Kaggle competition and people actually
如果你想的话，可以聊聊那个。而且我们办过一个 Contrail Kaggle 比赛，人们真的
36:10
beat us. But they found that we had a half pixel error in our labels. And they had to do with
打败了我们。但他们发现我们的 labels 里有一个 half pixel error。而且他们得处理
36:19
the center versus the lower left. Like where is zero zero is it in the lower left of the
中心和左下角的对比。就像 zero zero 在哪里，是在……的左下角吗
36:24
pixel or is it in the center? You know what I mean? Yeah. So they found that and exploited that
pixel，还是说它在中心？你懂我意思吧？对。所以他们发现了这一点，并利用了这一点
36:29
and squeezed to whatever a little bit extra stuff. Cause it turns out when you make artificial
然后榨出一点额外的东西。因为事实证明，当你做 artificial
36:34
data and you rotated, you have to make sure that that you take into account that half pixel
data 并旋转它时，你必须确保考虑到那个 half pixel
36:39
of set. So yes, people or people themselves will be act like these LMSs and try to sort of reward
offset。所以是的，人们，或者说人们自己，会表现得像这些 LLMs 一样，并试图有点 reward
36:47
hack on facing. So it sort of goes back to what is a good heart's law, you know, good heart's
hack on facing。所以这有点回到什么是 Goodhart's law，你懂的，Goodhart's
36:52
law. Yeah, yeah. Any that see, let's say any metric that becomes a target is no longer good
law。对，对。而且你看，比如说任何指标一旦变成目标，就不再是好
36:59
as a metric. Yeah. And so that's the, I mean, it's good. And it just, you have to be, you have to be
作为一个 metric。对。所以这就是，我是说，这挺好的。而且它就是，你得，你得
37:06
very, very careful and you have to, again, you have to have like layers of rigor. It's like,
非常非常小心，而且你得，再说一次，你得有那种层层严谨。就像，
37:10
okay, but we'll do this and we'll optimize for this. But you have to realize, okay, that's just
好吧，但我们会做这个，我们会针对这个去 optimize。但你必须意识到，好吧，那只是
37:13
now good heart's law applies and you have to be careful. And so that's a lot of reasons why
现在 Goodhart's law 适用了，而且你得小心。所以这就是很多原因，为什么
37:19
the whole AI field has been kind of constantly exhausting these, these things because again,
整个 AI 领域一直在不断耗尽这些、这些东西，因为再说一次，
37:24
good heart's law applies individually to every, to every leaderboard you make. So again,
Goodhart's law 会单独适用于你做的每一个、每一个 leaderboard。所以再说一次，
37:29
it's sort of, you just have to step back and be very, very careful. I, yeah, that's not, I don't
这有点像是，你就得退一步，非常非常小心。我，对，那不是，我不
37:36
know. That's really useful to my thinking is, you know, as we've had guests on it, it's been
知道。这对我的思考真的很有用，就是，你知道，我们请过嘉宾来聊它，它一直
37:42
a recurrent theme of how do you manage all this, the complexity that's introduced by
一个反复出现的主题是，你怎么管理这一切，也就是由
37:48
elements in agentic science? I think my follow-up question was about overfitting in gaggle.
agentic science 里的元素引入的复杂度？我想我接下来的问题是关于 gaggle 里的 overfitting。
37:53
Yeah, it is, if you, you had an auto, auto-caggle problem and then the question is given auto-caggle,
对，是这样，如果你，你有一个 auto，auto-caggle 问题，然后问题是，给定 auto-caggle，
38:00
how often was it successful? I mean, assuming you probably just read this on like all your
它成功的频率有多高？我是说，假设你可能只是在你的所有
38:04
gaggle competitions or something, well, we tried it on various, like what they call Playground
gaggle 竞赛之类的地方读到这个，嗯，我们在各种，像是他们叫 Playground
38:08
competitions and it did very, very well in Playground competitions. We've entered into different
竞赛上试过，它在 Playground 竞赛里表现得非常非常好。我们参加过不同的
38:14
competitions. Some of them, it turns out there's in the last few years, just the number of leaderboards
竞赛。其中一些，结果发现，在过去几年里，光是排行榜的数量
38:20
and competitions whatnot have just exploded far beyond gaggle. So we've done very well and some
和竞赛之类的，已经爆炸式增长，远远超出了 gaggle。所以我们做得很好，而且有些
38:26
of them, like one thing we're super proud of is the whole CDC set up this competition where you
其中，有一件我们特别自豪的事，就是整个 CDC 搞了这个竞赛，让你
38:33
try to predict next weeks the number of COVID in flu cases that will happen in every state and
去预测接下来几周，全美每个州和
38:40
territory in the US, and you try to predict a weekend advance. And Ira did super well on that.
领地会出现多少 COVID 和流感病例，而且你要提前一周预测。Ira 在那方面做得特别好。
38:46
It's funny because in some sense, Google invented the concept of using data to track disease
有意思的是，从某种意义上说，Google 发明了用数据追踪疾病
38:52
progression with Google flu. So it's kind of funny that you were sort of going full circle 20 years
进展的概念，靠的是 Google flu。所以有点好笑的是，你差不多是在 20 年
38:58
years later. So that did very well. Other ones where we've entered, we worked quite as good
后兜了个完整的圈。那个做得非常好。我们参加的其他一些比赛，我们做得也相当好
39:04
often because people are very, again, you know, you have to sometimes, sometimes how well you do
往往是因为人们非常——还是那句话，你知道，你有时候必须，有时候你做得有多好
39:11
in these competitions is a measure of how much sort of TLC you put into it and how much you
在这些比赛里，其实是衡量你投入了多少那种 TLC，以及你投入了多少
39:16
willing to squeeze the last point 001. And so so it was, I mean, Ira did well. It got you pretty
愿意把最后的 point 001 也挤出来。所以事情就是这样，我是说，Ira 做得很好。它让你很接近了，但我们没有，我们没能在最后那大概 30 个位置上把差距补上，因为没人去把最后那点，你知道，point 001 从这些东西上削掉。这个过程是不是很迭代？就像你做到那一步，你知道，它有点，我在论文里看到那些图表，你知道，你会发现它发现某个东西时会出现这些阶梯式变化，然后，然后再把它飞起来。然后，所以这很大程度是 human in a loop 吗？就像，好吧，你在问题上卡住了，试试这种。是的。好。是的。在 outer loop，我觉得那几乎就是更好玩、更有创造性的那部分。所以是的，哦，这里得看看这篇论文。哦，你在做坏事，或者你不
39:22
close, but we didn't, we didn't close the jump in the last and whatever 30 places or whatever
接近了，但我们没有，我们没有缩小最后那个 jump，不管是什么 30 个位置还是什么
39:26
because no one was there to shave the last, you know, point 001 off the things. Is it very
因为没人去把最后那，你知道，0.001 从这些东西上削掉。它非常
39:34
iterative? Like you get to it, you know, it sort of, I saw the charts in the paper and you know,
iterative 吗？就像你推进到那儿，你知道，它有点像，我在论文里看到那些图表，你知道，
39:38
you sort of get these step changes as it discovers something and then, and then fly it. And then,
你大概会看到这些阶梯式变化，当它发现点什么，然后，然后再让它飞起来。然后，
39:43
so is it very much human in a loop? Like, okay, you've stalled on the problem, like try this kind of
所以这很大程度上是 human in a loop 吗？就像，好吧，你在问题上卡住了，那试试这种
39:50
thing. Yes. Okay. Yes. At the outer loop, which is I think that's almost like the more fun creative
东西。是的。好吧。是的。在 outer loop，我觉得那几乎就像更有趣、更有创意的
39:56
part. So yes, oh, here have to look at this paper. Oh, you're doing something bad or you don't
部分。所以是的，哦，这里得看看这篇论文。哦，你在做坏事儿，或者你不
40:02
know what I mean. So it's almost like having a hyper eager grad student or something who doesn't
懂我意思吧。所以这几乎就像有一个超级积极、特别想干活的研究生之类的，他不用
40:07
sleep and you sort of tell things and you, and you, and you sort of guided around how often does
睡觉，然后你大概跟他说点东西，你、你、你大概会引导他，比如多久
40:14
someone intervene versus like, what is the outer loop look like actually? It might run for a few
会有人介入，或者，这个 outer loop 到底是什么样？它可能会跑几个
40:20
hours and and come back and give you some examples. And then you would, you know, you can do it
小时，然后回来给你一些例子。然后你，你知道，你可以做这件事，
40:26
as far as you like. You can, you can sort of keep trying and keep poking at it. So that's,
想做到什么程度都行。你可以，你可以一直试，一直戳它。所以这，
40:31
it's very much designed to be human in the loop then. Yes. Yes. Interesting. Because a lot of the
那它就是非常明确地被设计成 human in the loop。对。对。有意思。因为很多
40:36
other tools that I've tried tend to be very one shot. Well, I guess it depends on your
我试过的其他工具往往都非常 one shot。嗯，我觉得这取决于你
40:44
definition, right? I mean, it's obviously you talk to it and you start it and go for some number
的定义，对吧？我是说，很明显你跟它说话，你启动它，然后跑某个数量
40:50
of hours and come back. And then, but of course, then you say it, but then that's what the human
过几个小时，再回来。
40:54
creativity kicks in. And then you're sort of doing the outer loop where you sort of every,
然后，当然，接着你说出来，但这时候人类的创造力就登场了。
40:57
you know, depends if you want to sleep. But you know, every few hours you go and you give it
然后你大概是在做 outer loop，你大概每……你知道，取决于你想不想睡觉。
41:01
and try and you, what kind of budget are you giving me just like you blew through a million dollars
但你知道，每隔几个小时你就过去，交给它、试一下，然后你……你到底给了我什么样的 budget？就像你不小心一下子烧掉了一百万美元那种？
41:08
accidentally kind of thing? I don't actually know because we're using sort of, you know, internal
我其实不知道，因为我们用的是，你知道，对 Gemini 的 internal calls。
41:13
calls to Gemini. So actually, I don't actually know. So, but okay, so there's token budget,
所以实际上，我真的不知道。
41:19
but then there's also like I'm solving a problem that is computationally expensive. Oh yes,
所以，但好吧，所以有 token budget，但还有一点是，我是在解决一个 computationally expensive 的问题。
41:24
that also essentially underneath it because the scoring function itself might have the
哦对，那本质上也在它底下，因为 scoring function 本身可能会有那个……
41:29
Omani Carlo estimation or whatever. Yes. So you actually end up, you can actually end up
Omani Carlo estimation 还是什么的。对。所以你实际上最后会，你实际上最后会
41:34
using a lot of compute to just even do or simulation, like if you have a simulator inside,
用很多 compute，就为了甚至只是做 simulation，就像如果你里面有一个 simulator，
41:40
as to run a simulation. So yeah, you can, you can, you can spend a fair amount of just CPU or
来跑一个 simulation。所以对，你可以，你可以，你可以花不少 CPU 或者
41:46
GPU. So my little experiment with era era and cod is, is to build a neural network for some
GPU。所以我用 era era 和 cod 做的小实验，就是，就是搭一个 neural network，用来解决一些
41:54
classification problems. And so they obviously like, if you have enough data, then, you know,
classification problems。然后它们显然就像，如果你有足够的 data，那么，你知道，
41:59
larger networks work better, but they're more expensive to train and you get to, you start to
更大的 networks 效果更好，但训练起来更贵，然后你就会，你就开始
42:03
run into a question of how do I manage my budget? If I have a fixed budget so that I'm spending
遇到一个问题：我该怎么管理我的 budget？如果我有一个固定的 budget，这样我就在花
42:10
my, my, my dollars on the most effective solutions. That's right. And I think that's still
我的，我的，我的钱在最有效的 solutions 上。没错。而且我觉得那仍然
42:17
something we need to figure out, but it's of course, it's no different than if you have a grad student
这是我们需要搞清楚的事情，但当然了，这和一个研究生的情况没什么不同
42:22
and they're trying to train a very, very large neural network or very, very large data set.
而他们正试着训练一个非常非常大的 neural network，或者非常非常大的 data set。
42:28
They themselves have to, there's something like, oh, is there a scaling law can extrapolate? So
他们自己得，就好像，哦，有没有一个 scaling law 可以 extrapolate？所以
42:33
it's not, I guess that you, it's in the same problem, but maybe more urgent because it just runs
这不是，我猜你，它面对的是同一个问题，但也许更紧迫，因为它就这样不断
42:39
into this problem because it's so relentless, it runs into the problem much quicker than a grad
撞上这个问题，因为它太不停歇了，它撞上这个问题的速度比一个研究生
42:42
student could. One of the things that you optimized was, um, Contrails, um, can you talk a little bit
能撞上的速度还要快得多。你优化过的东西之一，嗯，是 Contrails，嗯，你能稍微谈谈
42:50
about that? Uh, well, uh, let me maybe spend a minute or two talking about the Contrails problem.
这个吗？呃，好，呃，也许让我花一两分钟谈谈 Contrails 这个问题。
42:56
Yes, it was for context. Contrails, not Kim trails, which is a conspiracy theory. Uh, yes.
对，先交代一下背景。Contrails，不是 Kim trails，那是个阴谋论。呃，对。
43:01
Although you should also dislike Contrails, but maybe not for the same reason.
虽然你也应该讨厌 Contrails，但可能不是因为同样的原因。
43:05
So Contrails are, uh, if you've ever seen, uh, those white clouds, um, form behind jets,
所以 Contrails 就是，呃，如果你见过，呃，那些在喷气式飞机后面形成的白色云，嗯，
43:11
those are called condensation trails or Contrails. And it turns out they add at least
它们叫 condensation trails 或者 Contrails。而且事实证明，它们至少
43:17
according to estimates that people have, uh, about 1% of all anthropogenic, uh, global warming
根据人们的一些估计，呃，所有 anthropogenic，呃，global warming 中大约 1%
43:23
is caused by Contrails. Why is that? I can just talk about the, maybe the physics of that. So it
是由 Contrails 造成的。为什么会这样？我可以就讲讲这个，也许是背后的 physics。所以
43:27
turns out that, uh, there's actually two, uh, countervailing effects. Contrails are, are, uh, well,
事实证明，呃，其实有两种，呃，countervailing effects。Contrails 是，是，呃，怎么说，
43:33
sometimes if you've ever seen them, they, they're streaking and they kind of go away. Those,
有时候如果你见过它们，它们，它们会拉成长条，然后渐渐消失。那些，
43:37
don't, they do anything, but sometimes they last for a long time. You'll just see in the cloud,
不会，它们会做点什么，但有时候它们会持续很长时间。你会在云里看到，
43:41
in the sky, just like almost like a waffle of, of, of, of just persistent Contrails, they're called.
在天上，几乎就像，呃，呃，呃，就像华夫饼格一样的 persistent Contrails，它们就是这么叫的。
43:48
And, uh, there's two effects that they have. Those are thin white clouds, so they reflect sunlight,
而且，呃，它们有两个效应。它们是薄薄的白云，所以会反射阳光，
43:52
but that only of course happens during the day. It turns out, uh, all, uh, all objects
但这当然只在白天发生。结果发现，呃，所有，呃，所有物体
43:57
emit something called black body radiation. And the earth does at whatever the temperature is about
都会发出一种叫 black body radiation 的东西。而地球也会，在温度大概是
44:01
300 Kelvin. It, it's in the far infrared around 10 microns. And at those, uh, wavelengths, uh,
300 Kelvin。它，它是在 far infrared，大概 10 microns 左右。而在那些，呃，波长下，呃，
44:08
Contrails have very low albedo. They're almost essentially black. And so they'll absorb a little bit
Contrails 的 albedo 非常低。它们基本上几乎是黑的。所以它们会吸收一点点
44:14
of the outgoing, uh, infrared radiation and then re-emitted both directions. So essentially,
向外散发的，呃，infrared radiation，然后再向两个方向重新发射出去。所以本质上，
44:18
they'll reflect some of the outgoing heat. So it'll trap you like a blanket. Um, and so because
它们会反射一些向外散发的热量。所以它会像毯子一样把你裹住。嗯，所以因为
44:23
that happens 24 hours a day, uh, they tend to be, uh, warming. And it turns out it's a surprising,
这种情况一天 24 小时都在发生，呃，它们往往会，呃，有 warming 效应。而且结果发现，这还挺令人惊讶的，
44:30
again, there's, uh, some uncertainty about it, but you know, uh, Contrails seris, seris,
再者，呃，这方面还有一些不确定性，但你知道，呃，Contrails seris，seris，
44:35
this sort of comes from, uh, from, from Contrails, um, might, might cover, especially in places like
这算是来自，呃，来自，来自 Contrails，嗯，可能，可能会覆盖，特别是在像
44:40
Europe, which has a lot of air, like traffic, a few percent of the actual skies covered by sort of
Europe 这种地方，那里有很多空中交通，实际天空的百分之几被某种
44:45
additional, uh, Contrails, which so it adds, it, in those places like Europe, it adds about
额外的，呃，Contrails 覆盖，所以它增加了，它，在像 Europe 这样的地方，它大约增加了
44:51
one watt per square meter of, uh, forcing, uh, it locally, at least, uh, which means that, uh,
一 watt per square meter 的，呃，forcing，呃，至少在当地是这样，呃，这意味着，呃，
44:58
just to give you a sense, all of anthropogenic, uh, uh, warming, sort of average across the whole
让你有个概念，所有 anthropogenic，呃，呃，warming，大致在整个
45:03
globe is about three watts per square meter. So in places of high, uh, airplane traffic, uh, it can
地球上的平均大约是 three watts per square meter。所以在，呃，飞机流量高的地方，呃，它可以
45:09
be a lot of warming locally. Uh, so what can you do? Well, um, it turns out, uh, Contrails are
会在局部造成大量变暖。呃，那你能做什么呢？嗯，原来，呃，contrails 是
45:15
caused by areas in the atmosphere that are ice supersaturated. They're a little bit like rock candy.
由大气中 ice supersaturated 的区域造成的。它们有点像冰糖。
45:21
So like, when you have, uh, rock candy, you get, uh, water solution that has too much sugar in it
所以就像，呃，当你做冰糖时，你会得到，呃，一种糖分过多的水溶液
45:25
and any little, um, you know, a little bit of sugar in it will just crystallize all the, the,
而里面任何一点点，嗯，你知道，一点点糖，就会把所有的、所有的、
45:31
the sugar out, just like in this Contra, there are these regions they tend to be kind of pancake
糖都结晶出来，就像在这个 contrail 里，有这些区域，它们往往有点像煎饼
45:35
shaped, only a few hundred meters tall. And if you fly through it, the jet exhaust has a little bit
形状的，只有几百米高。如果你从里面飞过，jet exhaust 里有一点
45:40
of, uh, moisture in it, which, uh, will turn into droplets and then freeze. And then for every
呃，水分，呃，会变成水滴，然后冻结。然后每
45:46
gram, if you're in this bad region, for every gram of water, uh, ice or so, you put out, it's about
一克，如果你在这个糟糕的区域里，每排放一克水，呃，冰之类的，大约是
45:53
10 kilograms of, whoa, water gets sucked out. So there's this enormous 10,001 curing ratio.
10 公斤的，哇，水被吸出来。所以有一个巨大的 10,001 curing ratio。
45:59
Um, so it's a big problem. So what, what you can do is you can figure out where these regions
嗯，所以这是个很大的问题。所以，你，你能做的是，你可以弄清楚这些区域
46:03
that are invisible, of course, uh, these regions of ice supersaturated are and then tell the plane
那些是看不见的，当然，呃，这些 ice supersaturated 的区域在哪里，然后告诉飞机
46:09
to go underneath and you only have to drop essentially, uh, what they call two flight levels. So it,
从下面飞过去，而你基本上只需要下降，呃，他们所说的两个 flight levels。所以它，
46:14
it actually, it costs a little bit of fuel, but not very much to kind of avoid these sort of bad
它其实，它会多花一点燃料，但不会很多，就能大概避开这种糟糕的
46:19
regions. So we built a system that sort of looks at satellite images and tries to detect where
区域。所以我们建了一个系统，它会看 satellite images，并试着检测
46:26
controls are. So we have essentially a continuous monitoring system and then try to build a model
controls 在哪里。所以我们基本上有一个 continuous monitoring system，然后试着建一个 model
46:31
of, of, because it turns out the weather models are not quite accurate enough to, uh, to find these,
的，的，因为事实证明 weather models 还不够准确，呃，找不到这些，
46:37
uh, places of ice supersaturation. So we build a custom model. Again, like a convolutional net or
呃，ice supersaturation 的地方。所以我们构建了一个 custom model。还是说，就像一个 convolutional net 或者
46:43
a unit or something, uh, to essentially, to predict where they're going to happen. So that, uh,
一个 unit 之类的，呃，本质上是为了预测它们会在哪里发生。然后，呃，
46:48
then we give, uh, maps to a flight planning software so that they can dodge it and, and inexpensively
我们把，呃，地图给到 flight planning software，这样他们就能避开它，并且，并且低成本地
46:54
reduce the, the, um, the climate impact of, of aviation by a lot. What's the physics behind
大幅减少，那个，呃，航空业的气候影响。这背后的物理原理是什么，
47:00
why you can predict that? Right. Is it just, I see it in the satellite and then tomorrow,
为什么你能预测这个？对吧。是不是只是，我在卫星上看到它，然后明天，
47:05
I think it'll be there because planes go to the same place or, oh, no, it's because you're trying
我觉得它还会在那儿，因为飞机去同一个地方，还是，哦不，是因为你在试图
47:10
to detect these, uh, regions of ice supersaturation because they're very, very persistent.
检测这些，呃，ice supersaturation 区域，因为它们非常非常持久。
47:15
Essentially, they're, oh, they're persistent. Oh, yeah, yeah. I mean, no one knows exactly,
本质上，它们是，哦，它们会持续存在。哦，对，对。我是说，没人确切知道，
47:19
but they could last for days. Essentially, they're caused by, they think, sort of warm, moist air
不过它们可能会持续好几天。基本上，他们觉得，是由某种温暖、潮湿的空气
47:23
being injected just at the boundary of the tropopause between the, uh, just at the bottom of the
被注入到 tropopause 的边界处，在，呃，就在
47:29
stratosphere. And then when humidity gets up there, it sort of sticks there for a long time,
stratosphere 的底部。然后当 humidity 到上面时，它就会在那儿停留很长时间，
47:33
and then gradually dissipates. Got it. So it's, so it's, there's just a sort of, they're like
然后慢慢消散。明白了。所以就是，所以就是，那里只是有种，它们就像
47:37
bad spots in the atmosphere you don't fly through. Right. Okay. And so once you've established that,
大气里你不会飞过去的坏点。对。好。所以一旦你确定了那个，
47:42
that's probably good for a couple of days, at least. Uh, well, you have to keep predicting
那可能至少几天内都管用。呃，嗯，你得一直预测
47:46
where that is. Yeah. And the, the models you're using are, you mentioned like CNN's or something.
那个在哪里。对。而且你用的，那些模型，你提到像 CNN 之类的。
47:51
That's right. And we haven't replaced those with era level models yet, but there was a very
没错。而且我们还没用 era level models 替换掉那些，但有一个非常
47:57
interesting problem that came up, which is you sort of want to know, well, just how much warming
有个很有意思的问题冒出来了，就是你有点想知道，嗯，到底有多少 warming
48:03
did this contrail make and how much did add to global warming? Uh, because, for example, you might
是这条 contrail 造成的，又给 global warming 增加了多少？呃，因为比如说，你可能会
48:09
want to find the biggest ones because there's some fuel cost and maybe a cost a bit of money
想找出最大的那些，因为会有一些燃油成本，可能还要花一点钱
48:15
for the airplanes to avoid it. So you say, well, Jay, I'd like to kind of know how much it did.
让这些飞机去避开它。所以你会说，嗯，Jay，我想大概知道它到底造成了多少。
48:19
But that's actually what they call a counterfactual problem. Like, okay, you made a contrail and certain
但这其实正是他们所说的 counterfactual 问题。就像，好吧，你制造了一条 contrail，并让一定
48:24
amount of infrared radiation happen. So we can measure that if you're careful. But would have happened
量的 infrared radiation 发生了。所以如果你够仔细，我们是可以测量这个的。但如果没有
48:31
if there hadn't been a contrail there. That's a very difficult thing to estimate because you can't
那条 contrail，又会发生什么？这是非常难估计的事情，因为你没法
48:35
access the universe where that didn't happen. So you have to make these things called counterfactual
进入那个没发生这件事的 universe。所以你不得不弄出这些叫做 counterfactual 的东西
48:41
models. And those are, if I don't know if you're listening or snow, counterfactual models are
models。而那些，如果我不知道你是在听还是 snow，counterfactual models 其实
48:45
actually pretty tricky to fit and make. Uh, and we were, there was, remember there were two,
真的挺难 fit 和做的。呃，而且我们当时，有，记得有两个，
48:51
there were two, um, the reflecting the sunlight and then there's the infrared. It turns out the
有两个，嗯，反射阳光的那个，然后还有 infrared。结果发现，
48:58
measuring what the effective reflection sunlight is actually more difficult. And we were actually
测量有效的反射阳光其实更难。而且我们实际上
49:03
stuck on it for two years. We had a model that worked okay, a counterfactual model for the outgoing
在这个问题上卡了两年。我们有一个 model 还算能用，一个针对 outgoing
49:07
longwave radiation, but not further reflected sunlight. Era actually helped us find a, uh, model
longwave radiation 的 counterfactual model，但不是针对进一步反射的阳光。Era 实际上帮我们找到了一个，呃，model
49:15
is sort of searched all the confounders and sort of figured out like, oh, how can we estimate it?
它有点像是搜索了所有 confounders，然后有点搞明白了，哦，我们该怎么估计它？
49:20
Because we, again, we even had like test, test code on, on sort of artificial, because you can kind of
因为我们，再说，我们甚至还有像 test，test code，基于，基于某种 artificial 的，因为你可以有点
49:25
inject artificial, uh, make artificial data sets where they're sort of injected, um, contrails and
注入人工的，呃，制作 artificial data sets，把 contrails 算是注入进去，然后
49:30
sort of figure out, oh, well, we know how much it was because we, we injected it. And so again,
算是搞清楚，哦，好吧，我们知道它有多少，因为，是我们把它注入进去的。所以，再一次，
49:36
our own attempts to even pass our own tests, but the era's thing actually did and sort of unstuck
我们自己那些甚至想通过自己测试的尝试，但 era 的那个东西确实做到了，算是解开了
49:42
this problem. So yeah, we're in the middle of writing up a paper. Uh, we have a paper about the
这个问题。所以是啊，我们正在写一篇论文。呃，我们有一篇关于
49:46
outgoing longwave radiation, but we have a paper, um, that, uh, it's not submitted yet, but we've
outgoing longwave radiation 的论文，但我们有一篇论文，嗯，那个，呃，还没提交，但我们
49:52
talked about it at, uh, at EG, I think, um, where, um, we actually solved this problem. So,
在，呃，在 EG 聊过，我想，嗯，在那里，嗯，我们实际上解决了这个问题。所以，
50:00
and, and the models the era comes up with, are they just like a big monstrosity of, of, of, of
而且，而且 era 想出来的那些 models，是不是就像一个巨大的，的，的，的
50:07
code? Or are they like pretty basic and is just, you needed the intuition to develop?
code？还是说它们其实挺基础的，只是，你需要那种直觉才能开发出来？
50:12
Yeah, it is actually, in this particular case, it was actually more of the ladder that it
对，确实是这样，在这个具体案例里，其实更多是那个 ladder，它
50:15
essentially sort of helped identify what the, it was a very simple model with them, just, uh,
基本上算是帮着识别出了那个什么，那是一个非常简单的 model，跟他们一起，只是，呃，
50:20
some number of confounders that we just hadn't, we just hadn't tried that combination before
有一些 confounders，我们之前就是没有，我们之前就是没试过那个组合，
50:24
and it worked very, very well. So yeah, it actually sort of came up with the, the, uh,
而且效果非常非常好。所以对，它其实算是得出了那个，那个，呃，
50:28
and it was seen in, in retrospect. So that was, I think, a, a big win. Yeah, that's interesting.
而且回头看，这也看出来了。所以我觉得那是一次，一次很大的胜利。对，这挺有意思的。
50:33
I know you've done a lot of work in climate. What other, other, um, stuff have you done?
我知道你在气候方面做过很多工作。你还做过哪些，哪些，嗯，事情？
50:38
I think I talked about this, right? The, the, I talked about the CO2 thing. That was pretty fun
我想我聊过这个，对吧？那个，那个，我聊过 CO2 那件事。那挺好玩
50:42
because it's this, it's still quite a, uh, the reason why estimating CO2 in the
因为就是这个，它还是挺，呃，估算 CO2 在……里为什么……的原因
50:48
atmosphere is an interesting problem. Is we actually don't know what the carbon flux is into
atmosphere 是个挺有意思的问题。我们其实不知道 carbon flux 到底有多少进入
50:54
the bio, in and out of the biosphere? I mean, we do, we know that the biosphere captures,
bio，进出 biosphere 又有多少？我是说，我们确实知道，我们知道 biosphere 会捕获，
51:00
right? We emit a bunch of CO2 out into the atmosphere and some of it gets absorbed into the ocean
对吧？我们往 atmosphere 里排放一堆 CO2，其中一些会被海洋吸收，
51:05
with sort of mostly an organic chemistry, some, some phytoplankton and a lot of it gets absorbed
大概主要是靠 organic chemistry，有些、有些 phytoplankton，还有很多会被
51:10
on land. Uh, but we put the air about bars about what happens is our, moderately large,
陆地吸收。呃，但关于会发生什么，我们给出的 error bars 算是中等偏大，
51:17
and the air bars 50 years from now are very large, like the models in 2100. We don't know how,
而 50 年后的 error bars 就非常大，就像 2100 年的 models 一样。我们不知道，
51:25
how the biosphere will react to the ever increasing temperatures and CO2. So we don't actually know how
biosphere 会怎么、会怎么应对不断升高的温度和 CO2。所以其实我们不知道
51:31
much the CO2 will absorb and the, the, the, the, the air bars are 300 ppm of CO2 just from the
CO2 会被吸收多少，而且那个、那个、那个、那个、那个 error bars 是 300 ppm 的 CO2，仅仅来自
51:40
uncertainty of how it gets absorbed. And just to point out, you know, right now there's about
它会如何被吸收的不确定性。而且我想指出，你知道，现在大概有
51:44
what? 40, 40, 40, 50 ppm. So it's huge. I mean, it could be seriously amazingly awful or not
多少？40、40、40、50 ppm。所以这很大。我是说，它可能糟糕得惊人，或者不
51:52
great, but you know, it, the, the, the 300 ppm is like enormous, uh, uncertainty. So it'd be
太好，但你知道，它、那、那、那 300 ppm 就是巨大的，呃，不确定性。所以要是能
51:57
really nice to figure out, you know, can we, can we reduce that? So this is like the CO2
弄清楚，你知道，我们能不能、能不能减少它，那就太好了。所以这就像 CO2
52:03
concentration is like one step towards that, um, solving that. And I know that Google has made
浓度是朝着那个，嗯，解决那个问题迈出的一步。而且我知道 Google 已经做出了
52:11
some really big, um, improvements in climate modeling and, and whether prediction as well, right?
一些非常大的，嗯，在 climate modeling 和，还有 weather prediction 方面的进步，对吧？
52:17
Is it, I was at NURPS this year, there's last NURPS. Yeah. And, um, and I stopped by the climate
是不是，我今年去了 NURPS，就是上一届 NURPS。对。而且，嗯，我还去了 climate
52:25
track. And, you know, I maybe only had a chance to listen to two talks or something, but I was,
track。而且，你知道，我可能只有机会听两场 talk 之类的，但我当时，
52:30
it really blew my mind. The sort of step change, I think this happened in the past. I don't know
这真的让我很震撼。那种 step change，我觉得过去也发生过。我不知道
52:36
what it is, maybe 10 years or whatever in terms of climate modeling. And I know a lot of that happened
是什么，也许是 10 年或者随便多久，在 climate modeling 方面。而且我知道很多都发生在
52:40
at Google. Can you talk a little bit about what has happened in Google and other places it has
Google。你能稍微谈谈在 Google 和其他地方发生了什么，它已经
52:46
made that allowed that really big transition and, and climate and weather modeling?
使得那种真正巨大的转变成为可能，以及 climate 和 weather modeling？
52:51
Okay. So let me, people often sort of collapse climate and weather together,
好的。那么让我说，人们经常有点把 climate 和 weather 混为一谈，
52:56
well, because they're fundamentally the same physics. Yeah. Although, at least for the atmospheric
嗯，因为它们在根本上是一样的 physics。是的。不过，至少对于 atmospheric
53:02
physics, they're obviously when you start having ice and land, you know, climate is long-term weather.
physics，它们显然，当你开始有冰和陆地时，你知道，climate 就是长期的 weather。
53:08
And so, the complexity of a full earth system model, which is a climate model, is much, much bigger
所以，一个完整的 earth system model 的复杂度，也就是 climate model，要大得多得多
53:14
than an atmospheric model. Like you have to actually measure what's the water flux and the CO2
比一个 atmospheric model。就像你得实际测量 water flux 是多少，以及 CO2
53:19
fluxing out of the land or will happen with ice. And so, there has been a step change with
从陆地 flux 出去，或者冰那边会发生什么。所以，已经出现了一个 step change，伴随着
53:26
weather models. Sorry, sorry, I want to make this, so weather is up to 15 days. Okay.
weather models。抱歉，抱歉，我想把这个说清楚，就是天气能预报到 15 天。好。
53:30
Approximately. Yeah. Because, you know, whether itself or the atmosphere appears to be chaotic,
大概吧。对。因为，你知道，天气本身或者大气似乎是 chaotic 的，
53:37
I'm hoping, I don't know if I should explain chaos. Essentially, it's the butterfly effect,
我希望，我不知道我该不该解释 chaos。本质上，这就是 butterfly effect，
53:42
right? That small perturbations, like a butterfly flaps its wings and the weather will be completely
对吧？就是那些小的 perturbations，比如一只蝴蝶扇动翅膀，然后天气会完全
53:47
different in, you know, two or three weeks. So, weather is trying to predict the actual
不同，你知道，在两三周后。所以，天气预测是在试图预测实际的
53:53
trajectory of the atmosphere over, say, two weeks. And that's now, that has been a huge step change.
大气的 trajectory，比如说，两周内的。而现在，那已经是一个巨大的 step change。
54:00
And that's because that's been a lot of not even the new LLM stuff that was based on, you know,
而那是因为，那很多甚至不是新的 LLM 东西，而是基于，你知道，
54:05
the 2018 era machine learning stuff and just a large amount of data and a large amount of compute.
2018 时代的 machine learning 东西，以及大量的数据和大量的 compute。
54:12
So, it's been, there's been a lot of sort of very clever work and a lot of it from Google,
所以，这一直，有很多那种非常聪明的工作，而且很多都来自 Google，
54:16
making new, new weather models. And it's been, it's been great. And in fact, we had a,
在做出新的、新的 weather models。而且这，这很棒。事实上，我们有一个，
54:22
a really neat breakthrough because now we can apparently predict tracks of cyclones, tropical
一个非常棒的突破，因为现在显然我们可以预测 cyclones、tropical
54:28
cyclones, much more accurately, many days in advance. And so places like Jamaica had, you got
cyclones 的路径，准确得多，提前很多天。所以像 Jamaica 这样的地方，你
54:35
hammered by a terrible hurricane. And a lot of the classic models didn't actually predict it,
被一场可怕的 hurricane 狠狠打击了。而很多 classic models 实际上并没有预测到它，
54:42
partially because it's often that the, especially the intensification, it's all being driven by
部分原因是，往往那个，尤其是 intensification，它都是由
54:48
what the surface temperature is because hurricanes are people might not realize their heat engines.
surface temperature 是多少，因为飓风——人们可能没意识到——是 heat engines。
54:52
Essentially, they convert sort of heat in the ocean to big atmospheric motions. So,
本质上，它们会把海洋里那种热量转化成大规模的 atmospheric motions。所以，
54:58
weather has been great. Climate is much more difficult because you don't actually care about,
天气一直很不错。气候要难得多，因为你其实并不关心，
55:02
you're not trying to predict whether it's going to rain in Seattle in 2070. You're trying to get
你不是想预测 2070 年西雅图会不会下雨。你是想得到
55:07
kind of like averages. And what makes it difficult is that it's what they call non-stationary.
有点像平均值的东西。而难点在于，它是所谓的 non-stationary。
55:13
So, that the, in fact, literally the, it's like the underlying physics or the underlying, like,
所以，那个，事实上，真的，它就像是底层的 physics，或者底层的，那个，
55:18
you know, plants are behaving differently and ice behaves differently. And, and so it's a,
你知道，植物表现不一样了，冰的表现也不一样了。而且，而且所以它是一个，
55:24
it's very, very difficult to use sort of classical ML on sort of true climate models. And so,
非常非常难把那种 classical ML 用在那种真正的 climate models 上。然后所以，
55:30
that's why sort of the whole discussion you guys had about, you know, what we're talking about,
所以这就是为什么，你们刚才那整段讨论，你知道，关于我们正在聊的东西，
55:35
about descriptive models and multiple hypothesis testing. That is incredibly severe in climate
关于 descriptive models 和 multiple hypothesis testing。这在气候领域里极其严重，
55:40
because we don't, we have no data from 30 years from now and we don't want to wait 30 or 50 years
因为我们没有，我们根本没有 30 年后的数据，而且我们不想等 30 年或 50 年
55:46
to find out whether we were right or that we overfit. So, whatever things we do, you have to be kind
才知道我们当时到底是对了，还是 overfit 了。所以，不管我们做什么，你都得挺
55:51
of careful and try to peel off subproblems and, and the problem of, of exactly how do you inject,
小心，试着把 subproblems 拆开，而且，还有这个问题，就是到底该怎么注入，
55:58
how do you build a big model that can predict into the future, but is still constrained by what we
怎么构建一个能预测未来、但又仍然受我们所
56:04
know. It's a fascinating problem. I think it's still unsolved, but it's a, it's a great problem to
知约束的大模型。这是个很吸引人的问题。我觉得它还没被解决，但这是个，这是个很好的问题，
56:09
have because of, again, these uncertainties, we really would like to know what will happen
能遇到，因为，还是那句话，这些不确定性，我们真的很想知道会发生什么
56:13
in, in 60 years to the climate. So, it's still, it's a thing, it's a very, very interesting problem
在，在 60 年里，climate 驱动的。所以，它仍然，它是这么回事，这是一个非常、非常有趣的问题
56:19
to work on. But, so far, AI has not revolutionized it because it's very, very resistant again because
值得研究。但是，到目前为止，AI 还没有彻底改变它，因为它非常、非常抗拒，又因为
56:25
of this data problem. It's, it's a low data problem. Does the butterfly effect chaotic
这个 data 问题。它，它是一个 low data 问题。butterfly effect 那种混沌的
56:31
nature of weather, does that also impact climate or is the time scale so large that you have a
天气的本质，这也会影响 climate 吗，还是 time scale 大到，你有一个
56:38
closed system for which, you know, maybe it's oscillating between poles or whatever, but it's,
closed system，你知道，也许它在两极之间 oscillating 之类的，但它，
56:43
it's sort of, when you will get at that time scale, it's more stationary.
它算是，当你到那个 time scale 的时候，它就更 stationary。
56:46
It's unfortunately non-stationary in a different way, but the original sort of whole chaos thing
不幸的是，它以另一种方式 non-stationary，但最初那种整个 chaos 的事
56:52
was, well, maybe many people came up with it, but I mean, meteorology, it was back to a person
是，嗯，也许很多人都提出过，但我是说，meteorology，它又回到一个人
56:58
named Lawrence, who had this sort of very model, very simple model, ODE. So, the difference between
叫 Lawrence 的，他有一个这种非常 model、非常简单的 model，ODE。
57:05
climate and weather is, is weather is where are you on the attractor? And climate is about the
所以，气候和天气的区别是，天气是你在 attractor 上的哪个位置？
57:10
statistics itself of the attractor. The problem with climate is that we're altering it so the
而气候是关于 attractor 本身的统计。
57:14
attractor itself is change and it's moving and there could be, and there could be, everyone
气候的问题在于，我们正在改变它，所以 attractor 本身在改变，它在移动，而且可能会有，而且可能会有，大家都在说 tipping points，那意味着 attractor 会突然改变。
57:20
talks about tipping points, that means that the attractor suddenly changes. And the trouble is,
麻烦的是，这非常、非常、非常难预测。
57:23
that's very, very, very difficult to predict. So, even the attractor's shape changes quickly?
所以，连 attractor 的形状也会很快改变？或者可能会。可能会。
57:30
Or could. Could. And the trouble is, when you're in a simulator, you don't know, like,
麻烦的是，当你在一个 simulator 里时，你不知道，比如，它是不是变得不稳定了，是因为我的 model 不够好，还是说这是一个真实的物理...
57:35
is it, did it go unstable because my model's not great, or is it an actual physical
是不是，它变得 unstable 是因为我的 model 不太好，还是它是一个实际的 physical
57:41
instability? Interesting, yeah. And it's extremely difficult to tell the difference.
不稳定性？有意思，是啊。而且极难分辨其中的区别。
57:45
So, what do you do, especially when, I mean, to me, it strikes me that not only do you have,
所以，你该怎么办呢，尤其是当，我的意思是，对我来说，我意识到你不仅没有，
57:48
not have future data, you really don't have much past data. You can do some measurements and
未来数据，你其实也没有多少过去数据。你可以做一些测量，以及
57:54
ice cores and lots of stuff to try to do that, but there was nobody with an instrument a hundred
ice cores 和很多其他东西来尝试做到这一点，但一百年前没有人有仪器
58:00
years ago. That's right. So, if you have annual data or whatever, maybe you have,
年前。没错。所以，如果你有年度数据或者别的什么，也许你有，
58:03
it's your lucky 50 data points in any one location, right? So, whatever we do has to be very constrained
运气好的话，你在任何一个地点能有 50 个 data points，对吧？所以，无论我们做什么，都必须非常受限于
58:11
by what we know, but it's just very different. I'm just telling you, sort of the horns
我们所知道的东西，但这非常不同。我只是在告诉你，有点像两难困境的
58:18
of the dilemma people on. So, people make these. In fact, people in applied size in general,
两端，人们所处的。所以，人们做这些。事实上，从事 applied size 的人一般来说，
58:21
I would say climate is the most extreme, make these things called process models, where
我会说，气候是最极端的，就是做这些叫做 process models 的东西，其中
58:25
what you do, and I've seen the code. Oh, well, you know, I'm going to be reductionist, and I'm going
你要做什么，而且我见过那些 code。哦，好吧，你知道，我会采取 reductionist 的做法，而且我会
58:30
to sort of take the horrible, complicated climate thing and sort of boil it down to a thousand
有点像把那个可怕又复杂的气候东西，有点像把它浓缩成一千
58:34
pieces, and then I'm going to, you know, find the expert who wrote a paper about, you know,
个部分，然后我会，你知道，找到那个写过论文的专家，关于，你知道，
58:39
piece number 763, an e-fidicubic to some data, like, for example, one thing that's very mysterious,
第 763 个部分，对某些 data 拟合一个 cubic，比如说，有一件非常神秘的事，
58:46
which is related to contrales is, how does ice behaving clouds? It turns out, you might, I get,
这件事跟 contrails 有关，就是，ice 在云里到底是怎么表现的？结果发现，你可能会，我明白，
58:52
everything is complicated once you dig into it, but it turns out that, like, when you make a
一旦你深入研究，一切都很复杂，但结果发现，就像，当你做出一个
58:57
contral, how long does it last? Well, it depends on, because the way contrales can evaporate is ice
contrail，它能持续多久？嗯，这取决于，因为 contrails 能蒸发的方式就是 ice
59:04
starts to accumulate, as I said, and then the ice crystals get big, and then they fall. But,
就像我说的，它开始累积，然后 ice crystals 变大，然后它们落下来。
59:10
of course, how quickly they fall depends on their shape, which is not known. And how much does the
但是，当然，它们落得多快取决于它们的形状，而这是未知的。
59:17
contral mix from the moist inside the contral out? Again, people have approximations, but they don't
还有，contrail 内部潮湿的部分向外混合了多少？
59:23
know. And so the uncertainties very much confound, and it's not just, oh, John, who cares about
还是那句话，人们有近似，但他们并不知道。
59:29
contrales, it turns out that the actual physics of microphysics of ice has very strong implications
所以这些不确定性非常让人摸不着头脑，而且这不只是，哦，John，谁在乎 contrails 啊，事实证明，冰的 microphysics 的实际物理机制对 climate models 的行为有非常强的影响，而且有点像，我们就是不知道。
59:36
about what climate models do, and it's sort of we just don't know. So I'm trying to say,
所以我想说的是，这非常棘手，而且不是一个已经解决的问题。
59:42
it's very gnarly, and it's not a solved problem. My hope is that with tools, maybe not like today's
我的希望是，借助工具，也许不是今天这个时代，而是明天那个时代，因为它，你记得，就像我之前说的，它不仅能够……
59:49
era, but maybe tomorrow's era, because it, you remember, as I was saying before, it not only can
era，但也许是明天的 era，因为它，你记得，就像我之前说的，它不仅能够
59:55
fit data, it can read papers, right? And the question is, it can read a lot more papers than we can.
fit data，它能读论文，对吧？问题是，它能读的论文比我们多得多。
60:01
So maybe we can integrate all the data or all the knowledge that people have carefully evaluated,
所以也许我们可以把所有 data，或者人们仔细评估过的所有知识，都整合起来，
60:09
much more than any one person writing a piece of code, and fit data. I mean, that would be,
这远远超过任何一个人写一段 code、再 fit data。我是说，那会，
60:15
that would be utterly glorious. We don't have that today, but that's sort of one of the hopes
那会简直太棒了。我们今天还没有这个，但这算是我抱有的一个希望，
60:19
that I have, even for a more amazing tool in the future is something that really can write code
甚至对于未来一个更厉害的工具来说，这个希望就是某种真的能写 code 的东西，
60:26
in a sane way, even much more sane, because it'll be constrained by all the scientific knowledge
以一种合理的方式，甚至合理得多，因为它会受到所有科学知识的约束，
60:31
that we've accumulated so far. That would be amazing. We don't have that today. I mean,
这些知识是我们到目前为止积累起来的。那会很棒。我们今天还没有这个。我是说，
60:35
that strikes me as being very similar to biology. Oh, yes. Oh, boy. Right? If you've ever played by
这让我觉得和生物学非常相似。哦，对。哦，天哪。对吧？如果你曾经玩过——
60:43
biology or even looked at biology, there's so many exceptions and so many hacks in the biological
生物学，甚至说，如果你去看生物学，biological systems 里有太多例外，太多 hacks。是的。天哪。
60:49
systems. Yes. Oh, boy. So yeah, it would be amazing if we could have a thing that could really
所以，是啊，如果我们能有一个东西，能真正把所有已知的科学知识和 data 整合起来，并尝试 synthesize 出某些新的 models 和新东西，那就太棒了。
60:55
integrate all known scientific knowledge with data and try to synthesize
我觉得你说的是，AI 在某种程度上可以在这里成为一个突破口，因为这些 models 太零碎了。
61:01
certain new models and new things. I think what you're saying is that AI can be an unlock here
它们必然是零碎的。
61:07
to some extent because the models are so piecemeal. They are necessarily piecemeal. And so they
所以，它们能够拼装这个拼图——当然不是把它们混在一起，而是真正拼装这个拼图——单靠 scale 和 capacity 其实就有很大帮助。
61:15
being able to assemble the jigsaw puzzle, not to mix them out of course, but to assemble like a
没错。
61:21
really this jigsaw puzzle, having just scale and capacity actually helps a lot. That's right.
这些 AI 拥有的一件事就是，不知怎么的，人类，甚至我基本说得没错，我想，
61:28
The one thing that these AI's have is somehow humans, even I'm pretty well right, I think,
这些 AI 拥有的一件事就是某种人类，即使我觉得我大体上是对的，我想，
61:34
but it's just difficult for me to kind of integrate across the n squared, different
但对我来说，就是很难去跨越 n squared 进行整合，不同的
61:40
n is the number of papers I've read in the life. It's pretty big. And it's just difficult for me
n 是我一生中读过的论文数量。这个数字挺大的。而且我很难
61:43
to even do that n squared thing. But somehow there's just so much data in those billions and billions
去做到那个 n squared 的事情。但不知怎么，那些数十亿乃至数十亿的
61:50
of parameters. And you can also give it access to read PDFs that it can somehow start to pull
parameters 里就是有那么多数据。而且你还可以让它读取 PDF，它就能以某种方式开始把
61:56
things together that people wouldn't do it. So that's again, I'm starting to see little indications
东西整合在一起，而人们不会这么做。所以，再次，我开始看到一些小小的迹象
62:01
of that inside of area. I'm not claiming that's what era it does today. But yeah, that's sort of my
在领域内部。我不是在声称这就是它今天时代所做的。但是的，这差不多就是我的
62:06
hope where this is going to go. I think I've heard a lot of people suggest something like the route
希望，事情会往这个方向发展。我想我听到很多人建议类似这样的路线
62:11
to intelligence is to combine elements with some form of search. So it's actually amazingly like
通往智能就是把元素和某种形式的 search 结合起来。所以它其实惊人地像
62:16
doing that right. Something which is maybe a very strong database look up with a good search algorithm
把这个做对。某种可能非常强的 database 查询，配上一个好的 search algorithm
62:24
is one way. And of course, I mean, there was the whole, I mean, people still do, I guess, the whole
是一种方式。当然，我是说，有那整套，我是说，人们现在还在做，我猜，整套
62:28
rag thing, of course. And if you think about it, Google itself, the 10 blue links things,
rag 那套东西，当然。而且你想想，Google 本身，10 blue links 那种东西，
62:35
it was, or is, a form of AI before we had allies, right? Because it was like, you could cast
它曾是，或者说现在也是，一种 AI，那时候我们还没有 allies，对吧？因为那就像，你可以把
62:42
yourself back to whatever 2010 or 2015, you could ask Google about literally anything and it will
自己带回 2010 或 2015 年那会儿，你可以问 Google 任何东西，它都会
62:48
tell you stuff, right? And so it's surprisingly, well, yeah, surprisingly well, because somebody on
告诉你一些东西，对吧？所以它出奇地，嗯，对，出奇地好，因为有人在
62:53
the internet has written about it probably. That's right. So if you can match that. In fact, that was
网上可能已经写过它了。没错。所以如果你能匹配那个。事实上，那是
62:58
one of the reasons why I wanted to come to Google. It's just that was such an amazing thing, right?
我想来 Google 的原因之一。只是那真是件太了不起的事，对吧？
63:03
I wonder how much of our audience did a search pre-Google and just know how bad that experience was.
我很好奇，我们的听众里有多少人在 Google 之前做过搜索，然后知道那段体验有多糟。
63:09
Yeah, I agree. 1998, I think, I think it's when Google, like I was using Altimista. And like, I
对，我同意。1998 年吧，我想，我觉得就是 Google 那时候，我当时在用 Altimista。然后就像，我
63:15
don't know, Google just got released and I used it. I just, sorry, sorry, a digital. I just,
不知道，Google 刚发布，我就用了。我只是，抱歉，抱歉，一个 digital。我只是，
63:21
I just dropped just like a like a hot potato or something and started immediately using Google. So
我就把它像烫手山芋什么的直接扔了，然后马上开始用 Google。所以
63:26
it's, so that is sort of a form of AI. And so yes, it might be, yeah, it could be that just having
也就是说，那算是 AI 的一种形式。所以是的，可能是，对，可能是只要拥有
63:34
access to all that and sort of keeping it in mind at the same time. That model of sort of
接触所有这些，并且同时把它记在脑子里。那种算是
63:40
scientific discovery in as much as it pans out is kind of comforting too, because it is reductionist
科学发现的模式，只要它能行得通，也挺让人安心的，因为它是 reductionist
63:46
so that you can look at the individual parts and understand them. So it found the exact things to
这样你可以看各个单独的部分并理解它们。所以它找到了确切的东西来
63:51
assemble, but there are actually maybe fundamentally things that people have invented or it's done
组装起来，但其实也许从根本上说，有些东西是人们发明出来的，或者是已经在它上面
63:58
iterations on. And so those, all those little pieces are individually understandable. And then
做过迭代的。所以那些，所有那些小部分，单独看都是可以理解的。然后
64:03
you can also put them together into a coherent picture. I think for a lot of problems like biology
你也可以把它们拼成一个连贯的整体图景。我觉得对于很多问题，比如生物学
64:08
or climate science, I don't think we would trust the answer unless it was in that shape. Yeah.
或者气候科学，我觉得除非答案长成那样，否则我们不会相信它。对。
64:15
Because if there was some giant black box model, it said, oh, this is how a cell works. It's like,
因为如果有一个巨大的 black box model，它说，哦，细胞就是这么运作的。那就像是，
64:20
do I believe it? I mean, I don't know if I believe it because I can't examine it. So
我信吗？我的意思是，我不知道我信不信它，因为我没法检查它。所以
64:24
I mean, to argue against that, though, if it works really well, but you'd have to gather,
我的意思是，不过要反驳这一点的话，如果它真的效果很好，但你得收集，
64:29
you'd have to, I mean, you have to test it, obviously, but a statistical model,
你得，我的意思是，你显然必须测试它，但一个 statistical model，
64:33
you have again, it has to extrapolate. Yeah. And it has to extrapolate to the extreme
你又得，它必须得 extrapolate。对。而且它必须 extrapolate 到极端
64:39
or the sort of the black swan events that that's right. Yeah. So it's very hard. This is why
或者那种 black swan events，没错。对。所以这非常难。这就是为什么
64:44
things like self-driving cars are very, very, it's a very difficult problem. It's all corner cases.
像 self-driving cars 这种东西非常非常，这是个非常难的问题。全都是 corner cases。
64:49
Yeah. It's kind of amazing how well they've done. That's an interesting point thinking about when
对。它们能做到这么好还挺让人惊讶的。这是个很有意思的观点，想想当你
64:54
you're coming from the world of physics, where a model was usually a single equation or a small
来自物理世界的时候，那里一个 model 通常就是一个 equation 或少数几个
65:00
number of equations, which uniquely define a system and everything about it. And you just crank,
equations，就能唯一地定义一个 system 以及关于它的一切。然后你就直接算，
65:04
you just find a solution to the system and you know, know everything you need to know. And I think
你只要找到这个 system 的一个 solution，然后你就知道，知道所有你需要知道的东西。而且我觉得
65:11
something like alpha fault was kind of a shift for a lot of people where before they thought,
像 AlphaFold 这样的东西对很多人来说算是一种转变，之前他们会觉得，
65:16
oh, protein folding is, you know, a problem where you just, if we find the right force field,
哦，protein folding 你知道，就是个问题，基本上只要我们能找到对的 force field，
65:21
and we have the right computational engine, we will solve protein folding. And the thought of even
又有对的 computational engine，我们就能解决 protein folding。而且就连
65:27
really solving it in a data-driven way was only a period of few years before, you know, alpha
真正用 data-driven 的方式解决它，也只是，你知道，在 alpha
65:33
fold, alpha fold one came out. And it's interesting that I think it's sort of forced the new AI
fold、alpha fold one 出现之前的几年而已。有意思的是，我觉得这有点像逼着新的 AI
65:42
modeling is, I think, forced people to reevaluate, oh, what is science? Because alpha fold is and
modeling，我觉得，也在逼着人们重新评估，哦，什么是科学？因为 alpha fold 以及
65:47
similar models are incredibly powerful. There's a lot of things that they've opened up as tools,
类似的 models 都强大得不可思议。它们作为工具打开了很多东西，
65:54
but at their core, they oftentimes don't give intuition in nearly the same way that, let's say,
但说到底，它们常常并不能像，比如说，
66:00
the most physicists historically would have wanted. And so, yes, it's sort of, there's this old saying,
历史上大多数物理学家本来会希望的那样，给出直觉。所以，是的，这有点像，有这么一句老话，
66:06
all models are wrong, some are useful. Yes. And yeah, box said that. Yeah, yeah. When do you find
所有模型都是错的，但有些是有用的。对。而且，是的，Box 说过这话。对，对。你什么时候会觉得
66:12
the data-driven models to be sufficient, and when do you want sort of like something which is
data-driven models 已经够用了，而什么时候你又想要某种
66:17
interpretable that humans can actually understand? I think it boils down to almost like the
interpretable、人类真的能理解的东西？我觉得这几乎可以归结为
66:21
difference between weather and climate. If you're in a data rich regime, like weather or even proteins
天气和气候之间的区别。如果你处在 data rich regime 里，比如天气，甚至 proteins
66:28
because of PDB, you can feel, oh, yes. In other words, I've got enough data to kind of cover,
因为 PDB，你会觉得，哦，对。换句话说，我有足够的数据，差不多能覆盖，
66:36
and so a statistical model like alpha fold should do. And so, in fact, a lot of people happily use
所以像 AlphaFold 这样的 statistical model 应该就够了。而事实上，很多人很乐意用
66:41
alpha, I think it's really revolutionized by understanding I'm not a biochemist, but people
Alpha，我觉得它真的带来了革命性变化，靠的是理解——我不是生物化学家，但人们
66:46
seem to love it. And one amazing thing they did is they exhaustively, they just ran it on all
好像很喜欢它。而且他们做的一件很了不起的事是，他们穷尽式地，直接把它跑在了所有
66:53
PDB and published it. It was just really, really cool. It's like 6 billion protein protein.
PDB 还把它发布了出来。那真的、真的太酷了。大概有 60 亿个 protein protein。
66:58
The vast majority are actually quite accurate. Yeah. Yeah, so that's just amazing. But it feels close
绝大多数其实都相当准确。是啊。是啊，所以那真是太厉害了。但感觉挺接近的
67:04
if you know what I mean. Yeah. Well, when it's like climate and it's open and it's non-stationary,
如果你懂我意思的话。是啊。嗯，但当它像气候那样，是开放的，而且是 non-stationary 的时候，
67:09
or you have to make these big extrapolations, you have to be much more cautious or maybe biology.
或者当你必须做这些很大的 extrapolations 时，你就得更谨慎，或者也许在生物学里也是。
67:16
Again, there's probably, there may be parts of biology like, oh, there was this virtual cell
再说，很可能，生物学里可能有些部分，比如，哦，有个 virtual cell
67:21
challenge from the art institute. That had a funny result. I know that people were, there may
来自 art institute 的 challenge。那个结果挺有意思的。我知道人们当时，可能
67:27
have been some overfitting for at least that. Yeah, we would say more, sorry. Or I guess maybe
至少在那个上面有些 overfitting。是啊，我们会说更多，抱歉。或者我猜也许
67:32
at a high level, I think simple baselines were very, very well. Yeah, yeah, yeah. Yeah, just like
从宏观层面来看，我觉得 simple baselines 非常非常好。是啊，是啊，是啊。是啊，就像
67:39
one of the classic things in whatever you do biology is always start with a simple baseline.
不管你做什么生物学，一个经典做法就是永远从一个简单的 baseline 开始。
67:43
Maybe this is probably just good email in general. It's good email. Yeah. Understand your
也许这大概就只是普遍来说好的 email。这是好的 email。对。理解你
67:47
simplest case in biology. There are many problems where they're extremely resistant to anything
在生物学里最简单的情况。有很多问题，它们极其抗拒任何
67:52
beyond the simple baseline, even if you have a lot of data. That that's right. And in fact,
超出简单 baseline 之外的东西，即使你有很多数据。对，没错。而且事实上，
67:55
I tell people the same thing. I said, always just fit linear aggression. Just do it. Just do it.
我跟人们说同样的话。我说，永远就 fit linear aggression。做就对了。做就对了。
68:01
Just do it. Yeah. Yeah. Oh, or SVMS. I mean, SVMs are just a different
做就对了。对。对。哦，或者 SVMS。我是说，SVMs 只是另一种
68:06
different form within your aggression. Yes. So when do you need the more process modeling thing?
在你的 aggression 里的不同形式。对。那么你什么时候需要更 process modeling 的东西？
68:12
I think it's just when you when you have, I mean, sort of climate is on one end and I don't know
我觉得就是当你，当你有的，我是说，有点像 climate 在一端，然后我不知道
68:18
whether maybe on the other end, maybe that may be too extreme, but I think it's where are you
或者说，也许在另一端，那可能太极端了，但我觉得这就是你在
68:23
on the on the data richness thing? When can you feel like, oh, no, I really have a closed,
data richness 这件事上的位置？什么时候你会觉得，哦，不，我真的有一个 closed，
68:27
I really have a closed problem. And I think I can actually cover it. Yeah, closed problem,
我真的有一个 closed problem。而且我觉得我其实能覆盖它。对，closed problem，
68:34
that you data fully covers is yeah, I think that that makes a lot of sense in what I've seen as well.
你的 data 能完全覆盖的，是的，我觉得这在我看到的情况里也很说得通。
68:41
Yeah. I think one thing I'm kind of curious about is when you're working on climate modeling,
对。我有点好奇的一件事是，当你在做 climate modeling 的时候，
68:46
what are the you talked about? Contrails. You've talked about CO2 predictions. What are the
你谈到的那些是什么？Contrails。你谈过 CO2 predictions。那些
68:54
the broad things you're trying to accomplish? So one of them is I was making up like making interventions
你想要达成的大方向上的事情是什么？所以其中一个是我在做类似 interventions 的事情
68:59
and the other one might be making predictions for things like insurance or like how do you help
另一个可能是为像保险这样的东西做 predictions，或者像是你怎么帮助
69:08
adjust for some sort of climate change? Or yeah, what are the like,
那是不是要针对某种气候变化做些调整？还是说，有哪些比较，
69:12
ensible goals? I guess, are you specifically or the community at large?
合理的目标？我猜，是你个人还是整个社区？
69:15
I think, you know, just like any community, there's probably many different goals. For me,
我觉得，你知道，就像任何社区一样，可能有各种不同的目标。对我来说，
69:20
I'm or my team. We're very, very interested in interventions. So like which ones are
我或者我的团队。我们对干预措施非常非常感兴趣。所以比如哪些是
69:28
possible at at relative cost? I mean, Contrails was kind of amazing because it turns out that
以相对成本来说可行的？我是说，Contrails挺神奇的，因为事实证明
69:34
the interventions quite low cost and we and also one amazing thing about Contrails is they are
这些干预措施成本相当低，而且Contrails还有一个神奇之处是它们是局部的，不像CO2那样的东西，因为如果一个国家决定解决自己上空的Contrails，而且
69:39
local unlike things like CO2 because so if a country decides to fix Contrails over itself and
确实改善了，我是说，它有全球影响，但它主要改善的是
69:45
actually improves, it's I mean, it has the global effects, but it mostly improves the the
其实会改善，我是说，它确实有全局效应，但它主要改善的是那个、那个
69:50
climate a little bit over themselves. So they like that. I guess that if you were in a
让气候在他们自己上方稍微……所以他们喜欢这样。我猜如果你在一个
69:55
cold climate and you want to warp it up, this is now your own. Oh, you could. Yeah, it turns out
寒冷的气候里，又想把它弄暖一点，那现在就归你了。哦，你可以。对，事实证明
70:00
that it's a little bit asymmetric. The warming is constant essentially and global. The cooling only
它有点不太对称。升温基本上是恒定的，而且是全球性的。降温只
70:06
happens when you're sort of at a good than the sun is at a good angle over you. So it's very rare
会在你差不多处在一个不错的位置、太阳又以不错的角度照在你上方时发生。所以这非常罕见
70:13
that the uncertainty, there are Contrails that wear our uncertainty bounds in terms of the warming.
就升温而言，有些 Contrails 是落在我们的 uncertainty bounds 之内的。
70:19
There are many, many Contrails mostly at night. Of course, where it's like a largely warming
有很多很多 Contrails，大多在晚上。当然，那些地方基本上是大范围升温
70:24
and we're very sure in terms of like two sigma. There's very not very many Contrails worries.
而且就 two sigma 来说我们非常确定。其实并没有特别多关于 Contrails 的担忧。
70:30
Oh, I know for sure that it's cooling and I want more of it. So only over sort of the poles
哦，我很确定它是在降温，而且我想要更多。所以只在类似极地的地方
70:36
in polar summer, do you know that the Contrails are cooling and therefore if you got rid of them,
在极地夏季，你知道吗，Contrails 其实是在降温的，所以如果你把它们去掉，
70:43
they would warm up, but there's essentially no flights over Antarctica and not that many over
反而会变暖，但南极洲上空基本上没有航班，夏季飞越
70:48
over the poles in the summer. So no one's going to, no one who lives in a cold climate is going to
极地上空的也不多。所以没有人会，住在寒冷气候里的人不会
70:53
use this maliciously. Well, yes. In that, in that, well, they wouldn't know for sure whether it was
恶意利用这个。嗯，是的。在那方面，在那方面，嗯，他们也没法确定到底是
70:59
warming or cooling and something would do stuff and things. So mostly we just sort of ignore
变暖还是变冷，而且各种因素会起作用之类的。所以大多数情况下我们就是有点忽略，
71:03
where we don't recommend that people fly those. You also brought up an interesting point about
我们不建议人们飞那些航线。你还提了一个关于
71:08
the economics. I mean, a lot of there, I think was a lot of resistance historically about
经济学的有趣观点。我是说，很多，我觉得历史上对于
71:15
certain, you know, to climate, certain climate change interventions which have in some sense
某些，你知道，对气候，某些气候变化干预措施有很多抵触，这些措施在某种意义上是
71:19
the market has just taken over. Like at this point, unambiguously, like renewables and batteries are
市场已经接管了。就现在这个阶段，毫无疑问，renewables 和 batteries 已经
71:25
just almost universally unambiguously just better than alternative. For for for for for for for
几乎普遍地、毫无疑问地就是比替代方案更好。对于对于对于对于对于对于对于
71:31
for non-mobile. I mean, you know, sorry, that's a really good point. Yeah, like planes we we do not
对于 non-mobile。我是说，你知道，抱歉，这真是个好观点。对，像飞机，我们我们并不
71:36
have a solution to. Correct. I mean, there are some battery part planes, but they were just small
有解决方案。没错。我是说，确实有一些 battery part 的飞机，但它们只是很小，
71:41
and have no very limited range. But they probably will never actually be a or it's hard to imagine
而且续航非常有限。但它们可能永远都不会真正成为，或者很难想象
71:46
that physics would be very, very on the unless we came up with something like nuclear batteries.
物理学会非常非常支持，除非我们能搞出像 nuclear batteries 这样的东西。
71:49
We kind of amazing, but I don't we don't know how to do that. Or even if we did, I think that the
那会挺神奇的，但我不觉得，我们不知道怎么做到。或者就算我们做到了，我觉得
71:54
risk of late people be too afraid of a nuclear battery going wrong or something. Oh yeah, yeah,
风险是，人们会太害怕 nuclear battery 出问题或者什么的。哦对，对，
71:59
we since we don't know what they are, we don't know what the risk. I guess we don't have risk.
我们，既然不知道它们是什么，就不知道风险是什么。我猜我们就没有风险。
72:03
Yeah, so we don't know. So yeah, that's that's the problem with I talk about I have given talks
对，所以我们不知道。对，所以问题就在这儿，我聊到、我做过一些演讲
72:07
about climate change and I talk about the pie pie chart of badness pie chart of sadness,
关于气候变化，我会讲那个糟糕饼图、悲伤饼图，
72:12
which is there's no one silver bullet for climate change, right? There's so many different
也就是说，气候变化没有单一的银弹，对吧？有那么多不同的
72:17
things that contribute greenhouse gases just from across our economy. So they sort of all have to
东西会排放 greenhouse gases，就来自我们整个经济的方方面面。所以它们差不多都得
72:22
be fixed or many, many of them have to be fixed. So there's no one single thing. I mean,
被解决，或者很多很多都得被解决。所以没有单一的一件事。我是说，
72:26
I worked on fusion fusion school and it might actually knock a lot of out if it's cheap enough
我研究过 fusion、fusion school，如果它足够便宜，它可能真的能干掉很多
72:31
which we don't know because we don't know if it'll work yet. I mean, fusion was interesting
这我们不知道，因为我们还不知道它能不能行。我是说，fusion 挺有意思的
72:36
things where the joke was always fusion is 30 years away. But I think it's actually no less than 30 years
这种事里，老笑话一直是，fusion 还要 30 年。但我觉得其实不少于 30 年
72:41
away, maybe. Yeah, no, I think there's a there's a definite probability that that someone will
还要，也许吧。对，不，我觉得有一种，有一种明确的可能性，有人会
72:48
make a commercially relevant fusion even by the end of this decade. So I think it's like three
哪怕在这个十年结束前做出有商业价值的 fusion。所以我觉得这大概还要三
72:53
years away now, not 30 years. It's very real. Interestingly enough, I think a lot of that I'm
年，不是 30 年。这是非常现实的。有意思的是，我觉得其中很多我
72:59
going to not just stump or advertise some of our other episodes, but a lot of it actually comes
不打算只是给我们的其他几期节目站台或打广告，但其中很多其实都
73:04
down to material science. I'm interesting enough in that. Well, I'm skeptical. Oh, super,
要归结到 material science。我对这个也挺感兴趣的。嗯，我持怀疑态度。哦，太好了，
73:08
sure, sure. Well, yeah, sorry. We can talk about fusion if you expect. Yeah, yeah.
当然，当然。嗯，对，抱歉。如果你希望的话，我们可以聊聊 fusion。对，对。
73:12
There was actually 200 two things. One is fusion. The other is better control systems,
其实有 200，两件事。一件是 fusion。另一件是更好的 control systems，
73:16
which I think is. Yes. And in fact, yeah, Google DeepMine has been working on control systems
我觉得是这样。对。而且事实上，对，Google DeepMine 一直在做 control systems
73:22
for Tokamax. Yeah. To make sure they don't essentially go and stay boned. Yeah, that's right.
给 Tokamax 用的。对。确保它们基本上不会说崩就崩、一崩到底。对，没错。
73:28
Disruptions are quite interesting to themselves. Yes. Yes. Yeah. It's basically the entire,
Disruptions 本身就挺有意思的。对。对。对。基本上就是整个，
73:34
the all energy in the Tokamax columnates into one little beam. And then it hits your
Tokamax 里所有的能量都 columnates 成一束小 beam。然后它打到你的
73:39
vacuum chamber. And you're very, very sad. Very sad. Yes. Yeah. I think people believe that either
vacuum chamber。然后你就非常非常难过。非常难过。对。对。我觉得人们相信，要么
73:45
could be could turn on after $30 billion disrupt and then basically have a $30 billion,
可能可以，可能会在 $30 billion 之后启动，然后 disrupt，结果基本上就有一个 $30 billion 的、
73:50
$30 billion brick or something. Oh, yeah. I guess you could try to patch it. I remember
$30 billion 的砖头之类的。哦，对。我猜你可以试着 patch 一下它。我记得
73:55
working. We again, before LMS, we worked with a fusion company called TAE. And I was in
工作。我们再说一次，在 LMS 之前，我们和一家叫 TAE 的 fusion 公司合作过。而我当时在
74:03
their control room. And yes, it was kind of sad. You have to be very careful. We were making
他们的控制室。而且是的，这有点让人难过。你得非常小心。我们当时在做
74:07
systems to recommend new experiments. And they were very, very skeptical and jaundice,
用来推荐新实验的 systems。而且他们非常、非常怀疑，还带着有色眼镜，
74:12
which was what they should because I've even been there, even under human control. It's like
他们确实该这样，因为连我都经历过，哪怕是在人工控制下。就像
74:18
they were doing some experiment and then you're has big bang. And it's like, oh, no. And then
他们在做某个实验，然后就砰的一声大爆炸。然后就，哦，不。然后
74:22
it's like, you know, then then the apparatus is down for two weeks as they patch. So
就像，你知道，然后、然后装置就得停两周，等他们 patch。所以
74:26
you were there during a disruption. Oh, no. This is sorry. They have field reverse configuration.
你当时正好在一次 disruption 现场。哦，不。这真遗憾。他们有 field reverse configuration。
74:30
Oh, okay. So which has its own, I mean, things, you know, there's some arc. So what is that? Sorry,
哦，好的。所以它有它自己的，我是说，东西，你知道，会有一些 arc。所以那是什么？抱歉，
74:36
I'm not familiar. Oh, oh, what's the field reverse configuration? Well, it turns out Tokamax
我不太熟悉。哦，哦，field reverse configuration 是什么？嗯，原来 Tokamax
74:40
are not all of them to perhaps the most studied form of plasma. There's many different kinds of
并不是它们全都是，也许是研究得最多的 plasma 形式。有很多不同种类的
74:44
architectures. Essentially, ways to try to stabilize and compress plasma. There was a shape.
architectures。本质上，是试图稳定并压缩 plasma 的方式。曾有一种形状。
74:51
Essentially, it's essentially a self-contained football of plasma called the field reverse
基本上，它本质上是一个自成一体的 plasma 橄榄球，叫做 field reverse
74:55
configuration, where essentially the magnetic field inside and outside are opposite. So they're
configuration，其中基本上内部和外部的 magnetic field 是相反的。所以它们
74:59
separated by something called separate tricks. And that is sort of in theory unstable, but in practice
被一种叫做 separate tricks 的东西分隔开。而那在理论上有点不稳定，但在实践中
75:06
stable. Like, for example, when you run magnitude hydrodynamics, MHD code, it's unstable under that
稳定的。比如说，当你运行 magnitude hydrodynamics，MHD code 时，它是不稳定的，在那种
75:13
assumption, but that's an assumption that's not the way the real world works. And so, yeah,
假设下，但那个假设并不是真实世界运作的方式。所以，是的，
75:18
it was kind of disfavored for many years, but TA and other people and the Kylion have
它很多年来有点不受青睐，但 TA 和其他人以及 Kylion 已经
75:25
FRCs because they are actually relatively robust. You can actually knock them against walls
FRCs，因为它们其实相对 robust。你其实可以把它们往墙上撞
75:31
and they'll still stay stable. Yes, but then you can still get discharges and things that
而且它们还是会保持 stable。是的，但那样你还是会有 discharges，还有一些会
75:36
punch holes in your vacuum chamber, which is kind of unfortunate for clarification. So you have
在你的 vacuum chamber 上打出洞的东西，这有点不幸，澄清一下。所以你有
75:41
these fusion reactors, they are, or trying to be reactors. And you create a plasma. The plasma
这些 fusion reactors，它们是，或者说正试着成为 reactors。然后你制造出 plasma。这个 plasma
75:53
is magnetically charged. Oh, you can find it. So it's confined by a magnetic field. So you have some
是带磁性的。哦，你可以找到它。所以它被 magnetic field 约束住。所以你有某种
76:01
sort of magnetic system that is tunable by a computer. And then the computer tries to kind of
算是可由计算机调谐的 magnetic system。然后计算机试着去，嗯，有点
76:08
maintain the, well, indoor confinement. Well, FRCs kind of once you make them, they're sort of
维持，嗯，indoor confinement。嗯，FRCs 一旦你造出来，它们算是
76:13
sustained. There's different ways of trying to make sure you, okay, so all of fusion boils down
sustained。有不同方式试着确保你，好吧，所以所有 fusion 归根结底都
76:20
as many call the loss in criteria. There's essentially, and it explains why fusion is hard.
正如很多人所说的，criteria 里的 loss。本质上就是，而且它解释了为什么 fusion 这么难。
76:25
Essentially, you can just very easily, on the back of an envelope, just show that the density,
基本上，你可以非常容易地，在信封背面随手一算，就证明 density、
76:30
the temperature, and essentially the energy loss. It's called the confinement time. It's one over
temperature，以及本质上就是 energy loss。它叫做 confinement time。它是 1 除以
76:34
the amount of time it takes for the energy to decay away, one over E in a plasma. So the product
能量衰减掉所需的时间，也就是 plasma 里 E 的倒数。所以这三个
76:40
of those three numbers has to be bigger than some constant. And then you can get fusion. And if you
数值的乘积必须大于某个常数。然后你就能实现 fusion。而如果
76:45
don't, then you don't. And that's explains the fact that it's a product of three numbers.
你做不到，那就做不到。这就解释了它是三个数值的乘积这个事实。
76:49
It explains why fusion is so hard, because every approach has an Achilles heel, where one of those
它解释了为什么 fusion 这么难，因为每种方法都有一个 Achilles heel，就是其中某个
76:54
numbers is not very big. And then they try to desperately make that be higher. And every
数值不够大。然后他们就拼命想把它提高。而每一个
77:01
approach is, every approach to fusion is kind of different. And a lot of, you have to be a bit
方法就是，每种实现 fusion 的方法都挺不一样的。而且很多时候，你得有点
77:05
skeptical in there. There's all these sort of breathless news things about fusion, because they'll say,
怀疑。关于 fusion，总有一堆那种一惊一乍的新闻，因为他们会说，
77:10
you know, now confinement time is starting to like, oh, stable for X minutes or whatever,
你知道，现在 confinement time 开始有点，哦，能稳定 X 分钟还是什么的，
77:15
and it's talking about like one of the three numbers. You have to have all three numbers before
而这只是在说三个数字里的一个。你得三个数字都达到，才能
77:19
it, you can get fusion. I think that the whole field is making a lot of progress, and it's very
实现 fusion。我觉得整个领域正在取得很多进展，而且非常
77:24
exciting. But you do have to, you have to be a little bit cautious about the breathless news
令人兴奋。但你还是得，你得对那些一惊一乍的新闻
77:28
articles that only talk about one number. So what is the computational part of that?
那些只谈一个数字的文章，稍微谨慎一点。那这里面 computational 的部分是什么？
77:32
Oh, and unfortunately, for better or for worse, it depends on the approach. So for Tokamax,
哦，而且很不幸，不管是好是坏，这取决于具体方法。所以对于 Tokamax，
77:38
as Brandon said, it's that there's this, it's mostly stable, except that there's occasionally
就像 Brandon 说的，就是有这么一个东西，它基本上是稳定的，但偶尔会有这种 instability，它会把所有的能量和最大值都猛地集中到一个地方。所以你得差不多把一切都控制在某种范围内。所以它是一个控制系统。FRCs 本身有非常简单的 instabilities。比如说，它们有一种所谓的 Z instability。所以它没问题。它是稳定的。它只会晃动，真的就是来回晃动，但你只需要做一个所谓的 PID controller，它就能把足球保持在 reactor 的中心，那就没事了。它是通过调整 magnetic field 来做到这一点的。是的，其实它差不多就是，我觉得 electric field 差不多就是...不，是把它来回撞。人们遇到的问题在于
77:44
this instability that takes all the energy and max and it smacks it into one place. And so you
这种不稳定性会把所有能量和 max 都卷走，然后把它猛地集中到一个地方。所以你就
77:49
have to sort of keep everything sort of under control. So it's a control system. FRCs themselves
得把一切都尽量控制住。所以这是个控制系统。FRCs 本身
77:55
have very simple instabilities. So for example, they have what they call a Z instability. So
有非常简单的不稳定性。比如说，它们有一种所谓的 Z instability。所以
78:00
it's fine. It's stable. It'll just wobble, literally wobble back and forth, but you just make
没事。它是稳定的。它只会晃，真的就是来回晃，但你只要做一个
78:04
what they call a PID controller that just keeps the football in the center of the reactor and
他们所谓的 PID controller，它只是让 football 一直待在 reactor 的中心，然后
78:09
things are fine. And it does that by adjusting the magnetic field. Yeah, it's sort of actually just,
一切就没问题了。它是通过调节 magnetic field 来做到这一点的。对，它其实差不多就是，
78:13
I think the electric field is sort of not knocks it back and forth. The issue that people have is
我觉得 electric field 有点不是把它来回敲。人们的问题是
78:19
it really depends on which sort of plasma architecture they're deciding to use. Climate is
这真的取决于他们决定采用哪种 plasma architecture。气候是
78:25
sort of political because of economics, basically, probably mostly, maybe other stuff. But the economics
有点政治性的，因为经济，基本上，可能主要是这样，也许还有别的东西。但它的经济
78:34
of it, you have to persuade people to somehow spend more or you have to have a solution that
层面来说，你得说服人们想办法花更多钱，或者你得有一个解决方案，
78:42
has this happy coincidence where it's both economically better and better for the climate.
能恰好两全其美：既在经济上更好，也对气候更好。
78:48
That's hard. Yeah. But in cases, it's not. I mean, in cases, it hasn't been hard.
这很难。是啊。但在有些情况下，并不是这样。我是说，在有些情况下，它并不难。
78:52
Yeah. I just don't think about like predicting even weather, right? You can prep and you could
是啊。我甚至不会去想比如预测天气，对吧？你可以做准备，然后你就能
78:59
see how that could be economically beneficial. So what kind of work are you doing with interventions?
看出那在经济上可能怎么有益。所以你在做哪类 interventions 的工作？
79:05
And how does that kind of interact with economics? It sounds like the Contrails one.
而那又是怎么和经济相互作用的？听起来像是 Contrails 那个。
79:13
An analysis that actually, this is great because it's very low economic impact, but high value.
有一种分析是，其实这很棒，因为它的经济影响非常低，但价值很高。
79:19
That's right. So if you're trying to think, there's sort of energy interventions. So you have
对。所以如果你在琢磨，有一些 energy interventions 之类的东西。所以你
79:22
to be, you have to sort of compete with existing forms of energy. That's not trivial. Unless
得，你得跟现有的能源形式竞争。这可不简单。除非
79:29
there's a co-benefit or there's some sort of clever, just co-benefit. This, again, is highly
有 co-benefit，或者有某种很巧妙的，就是 co-benefit。这，再强调一次，高度
79:36
speculative. It wasn't our work. There was a startup that was, I don't know if you saw the news. It
推测性。这不是我们的工作。有一个 startup，我不知道你有没有看到那条新闻。那
79:40
was last year, I think, where someone figured out if you inject mercury into a fusion reactor
是去年，我想，有人发现如果你把水银注入 fusion reactor
79:47
that the neutron flux can actually transmute the mercury into gold. Then you can sell the gold,
neutron flux 其实能把水银 transmute 成黄金。然后你可以把黄金卖掉，
79:53
which I thought was very clever. It might not work. As a physicist, the one thing I want out
我觉得这非常聪明。它可能行不通。作为一名物理学家，我唯一想从中得到的
79:57
of a fusion reactor is helium, but that's a different story. Yeah, helium is three. Well,
fusion reactor 出来的是 helium，但那是另一个故事了。对，helium 是 3。嗯，
80:03
I mean, the helium 4 is kind of boring, although it's getting because the strategic reserve has been
我是说，helium 4 有点无聊，虽然它现在变得……因为战略储备已经
80:08
shut down. There's less of it. Yes. And of course, I want helium 3, 4, not even just a fuse,
关停了。它的量更少了。对。而且当然，我想要 helium 3、4，甚至不只是一个 fuse，
80:14
just to make dilution refrigerators for corners. Or MRIs or so much technology, we think of
只是为了给 corners 做 dilution refrigerators。或者 MRIs，或者那么多技术，我们会想到
80:21
how it actually just goes up the window if we run out of helium. That's true, no one's thinking
如果 helium 用完了，它实际上就直接泡汤了。确实，没人会想
80:24
about it. That's like a complete aside though. Yes, the fact that the US had a helium, a strategic
这件事。不过那完全是个题外话。对，美国曾经有一个 helium，一个战略性的
80:30
helium was to reserve was for a very important blimp-fling. But they kept it for decades anyway,
helium 本来是要留作储备，是为了一个非常重要的 blimp-fling。但不管怎样，他们还是保留了几十年，
80:35
so that was nice. But then we stopped. We got rid of it all went up in the air. Yes, in balloons
所以那还挺好的。但后来我们停掉了。我们把它全处理掉了，全都飘到空中了。对，装在气球里。
80:41
and stuff. Or out of natural gas wells. Sorry, now we're talking about helium.
之类的。或者从 natural gas wells 里提取的。抱歉，我们现在说的是 helium。
80:46
Yeah. So what are the interventions? What are some of the most exciting, interesting ones?
是啊。那有哪些 interventions？其中最让人兴奋、最有趣的又有哪些？
80:51
Well, I'm very excited by fusion. I mean, I don't know if it's intervention. That's sort
嗯，我对 fusion 非常兴奋。我是说，我不知道这算不算 intervention。那算是某种
80:54
of a source of energy because if we can make it work and we can make it be sort of low enough
能源来源，因为如果我们能让它成功，并且能让它做到某种足够低的
81:00
capital cost, that that will actually help a lot because at least the current models are,
capital cost，那真的会帮上大忙，因为至少就目前的 models 来说，
81:06
renewables are great. Ideally, you'd like to electrify everything, right? Which has problems
renewables 很棒。理想情况下，你会想把所有东西都电气化，对吧？但这会有问题
81:12
because you can't electrify flights, but you could try to electrify a lot of stuff. You know,
因为你没法把航班电气化，但你可以试着把很多东西电气化。你知道，
81:16
there are EVs. You'd have to figure out how to electrify things like cement or steel. Those
有 EVs。你得想办法把水泥或钢铁这类东西电气化。那些
81:23
are hard, especially things like making steel reduction power anyway to essentially you're
都很难，尤其是像炼钢、reduction、power 这些事，反正本质上你就是
81:29
adding carbon and you're reducing iron ore. So there's a lot of sort of things that are difficult
加 carbon，然后把 iron ore 还原。所以有很多种事情都挺难的
81:35
about electrifying everything, but if you could electrify everything, then the amount of electricity
要把一切都 electrify 这件事，但如果你能把一切都 electrify，那么 electricity 的量
81:38
required would grow by a factor of five. And you could try to grow renewables. Renewables
所需的会增长到五倍。然后你可以试着扩大 renewables。Renewables
81:46
was battery. Again, trying to squeeze all of it out, it starts getting ever more expensive because
就是 battery。再说一遍，想把最后这点也全榨出来，成本会变得越来越高，因为
81:52
you just need ever more. You need like a huge number of batteries to recover the last few percent
你只会需要越来越多。你需要巨量的 batteries，才能把那最后几个百分点
81:58
or even 10 or 20 percent. So we do need some sort of power that can cover the last 20 percent,
甚至 10% 或 20% 也补回来。所以我们确实需要某种 power 能覆盖最后那 20%，
82:04
something that's based low. So fusion might be a thing for that. So that's super exciting. Again,
某种 based low 的东西。所以 fusion 也许就是干这个的。所以这太让人兴奋了。再说一次，
82:11
there's no one sort of silver bullet that can sort of cover all the cases. So I'm happy to
并没有什么单一的银弹能覆盖所有情况。所以我很乐意
82:16
sort of talk about any specific case, but it's sort of like the world is a very complicated place,
去聊任何具体的情况，但这就有点像，世界是一个非常复杂的地方，
82:21
and the global economy is a very complicated place. So it's super hard to sort of talk about
而全球经济也是个非常复杂的地方。所以很难去泛泛地谈
82:25
sort of interventions in general. Maybe instead of interventions, one thing I'm curious about
一般性的干预。也许与其谈干预，我更好奇的一点是
82:30
is how does this make effect decisions into, for example, like what do we build? How do we build?
这到底是如何影响决策的，比如，我们造什么？我们怎么造？
82:37
I think you're from LA, right? Or at least you... Well, I spent 11 years there, yeah.
我记得你是从 LA 来的，对吧？或者至少你……嗯，我在那儿待了 11 年，是的。
82:42
You're what? Okay. So yeah, you spend a lot of your life in LA. I mean, LA just basically,
你什么？好吧。所以是的，你人生中很大一部分都在 LA 度过。我是说，LA 基本上，
82:46
large parts of it just burned down. And maybe probably close to where you used to live.
它的大片区域都被烧毁了。而且也许，很可能离你以前住的地方很近。
82:52
So this is something that I think a lot of people kind of saw coming. Maybe partially
所以这件事，我觉得很多人多少都有点预感会来。可能一部分
82:58
due to regulatory issues, but partially due to other issues. And we were completely unprepared.
是因为监管问题，但一部分也是因为其他问题。而我们完全没准备好。
83:03
And it seems like there is a lack of preparation about what to do next or to sort of adjust for this.
而且看起来，对于接下来该做什么、或者该怎么为此调整，大家都缺乏准备。
83:10
And I mean, have you worked on basically predicting new risk assessments or suggestions
我的意思是，你有没有做过基本上是在预测新的 risk assessments 或建议的工作，
83:18
like what do we actually change to maybe harden society, even for what's coming regardless
比如我们实际该改变什么，才能让社会更能扛住，甚至是无论
83:26
of whether or not we actually do something to solve the underlying problem? That's right.
我们到底有没有真的做点什么来解决那个根本问题，都要为即将到来的事做准备？对。
83:30
So in fact, there's a big effort at Google into something called Crisis Resilience.
所以其实，Google 内部有一个很大的 effort，在做一项叫 Crisis Resilience 的事。
83:36
And so we had a very fun project called FireSat. I don't know if you know about this.
而且我们有个很好玩的项目叫 FireSat。我不知道你知不知道这个。
83:42
So it turns out that for wildfires, a lot of these wildfires, you could, if you only caught
所以结果发现，对于野火来说，很多这类野火，你其实可以，只要你能足够早地
83:48
them early enough, it's very easy to put out a wildfire the size of this room. But even if it's
发现它们，扑灭一场这个房间大小的野火是非常容易的。但即使它只是
83:55
like an acre, it gets much, much harder. And so, and of course, under certain circumstances,
像一英亩那么大，也会变得难得多得多。所以，当然了，在某些情况下，
84:00
they can grow exponentially from the size of this room up to an acre. So that might be hard to catch.
它们可以从这个房间的大小指数级地增长到一英亩。所以那可能很难抓住。
84:05
But they often sort of start small and spend a while. So we figured out that, oh, if you had
但它们往往有点像从小开始，然后持续一段时间。所以我们发现，哦，如果你有
84:12
a global constellation of lower orbit satellites that could detect in the midwave IR,
一个由 lower orbit satellites 组成的全球 constellation，能够在 midwave IR 中探测，
84:19
which would go back to the black body essentially, that's the temperature of fire,
这基本上会回到 black body，也就是火的温度，
84:23
they stand, fire stand out in the midwave IR. And so we designed a sensor that,
它们会凸显，火在 midwave IR 中会凸显出来。所以我们设计了一个 sensor，
84:32
if you built, it depends on exactly what their orbits, but roughly 50 to 80 of them,
如果你建起来的话，这取决于它们的 orbits 到底是什么，但大概 50 到 80 个，
84:37
you could actually find fires about the size of this room, about five meters on a small,
你其实就能发现像这个房间这么大的火情，大概五米见方，
84:44
maybe it's a big large in this room, five meters on a side. And anywhere on the planet,
对这个房间来说可能有点大，五米见方。而且在地球上任何地方，
84:49
and again, depending on how many satellites you had within, within like 15 to 20 minutes,
再说一遍，这取决于你在大概 15 到 20 分钟内有多少 satellites，
84:53
you'd have to put a fair number up like 80 to get them within 15 minutes. And then you could
你得部署相当多的数量上去，比如 80 颗，才能让它们在 15 分钟内到位。然后你就能
84:59
actually intervene. You could decide not to, if you wanted to, you know, I have the fire burn
真正介入了。如果你愿意，你也可以决定不介入，你知道，我让火去烧
85:03
fuel and you thought it was safe, but if it was going to blow up to something unsafe. So we worked with
燃料，而你觉得这是安全的，但如果它要演变成不安全的情况。所以我们和一个
85:08
a now a nonprofit called Earth Fire Alliance that we're part of. And so they're starting to,
现在叫 Earth Fire Alliance 的非营利组织合作，我们也是其中一员。所以他们开始，
85:13
we've launched one satellite, which is a prototype. We've worked with a company named Muon Space
我们发射了一颗 satellite，是个 prototype。我们和一家叫 Muon Space 的公司合作
85:18
to actually sort of make the satellites. So that's cool. We have wildfire boundary detection,
来真正算是把这些 satellites 造出来。所以这挺酷的。我们有 wildfire boundary detection，
85:24
and we propagate that information out through Google. So we can actually sort of figure out
然后我们把这些信息通过 Google 传播出去。所以我们其实算是能搞清楚
85:29
from existing satellites and existing data fees where the boundaries of fires are, and then we
从现有的 satellites 和现有的 data fees 里，火灾的边界在哪里，然后我们
85:35
sort of tell people through through their Android phones or through search about fires.
算是通过他们的 Android 手机或通过搜索告诉人们火灾的情况。
85:39
We've worked with the US Forest Service on making new models for how fires propagate.
我们和 US Forest Service 合作，制作新的 models 来研究火灾如何 propagate。
85:45
Because again, that goes back to these process-based models from the 70s by a person named
因为再说一次，这又要追溯到 70 年代由一位叫
85:49
Brother Mel. So we've actually made a little neural network proxy model based on a new essentially
Brother Mel 的人提出的这些 process-based models。所以我们其实做了一个小小的 neural network proxy model，本质上基于一个新的
85:54
to sort of be able to run it very, very quickly. So we worked with the Forest Service on that. So yeah,
某种程度上能非常非常快地运行它。所以我们和 Forest Service 一起做了这件事。所以，对，
86:00
yeah, we're very, very interested in trying to minimize. Because it turns out people might not realize
对，我们非常非常想尽量把它降到最低。因为结果发现，大家可能没意识到
86:06
the World Health Organization estimates that there are 300,000 excess deaths a year across the world
World Health Organization 估计，全球每年有 30 万例 excess deaths
86:12
from wildfire smoke. Yeah, I mean, I remember it's been a few years since we had a really bad fire
是由野火烟雾造成的。对，我是说，我记得，距离我们上一次经历特别糟糕的火灾
86:18
season, maybe what, four or five years ago, there was this cloud of smoke which like crossed all
季已经好几年了，可能大概四五年前吧，有一团烟雾，简直横跨了整个
86:23
of northern US and Canada and caused a lot of respiratory issues, I think. Yeah, and it's very
美国北部和加拿大，导致了很多呼吸道问题，我觉得。对，而且这非常
86:29
hard to track. I mean, you have to get these. You have to get the estimate these excess deaths from
难追踪。我是说，你得拿到这些。你得通过
86:33
statistical means. But yeah, it's a very serious public health problem, and also just very scary,
统计手段来估算这些 excess deaths。但对啊，这是个非常严重的公共卫生问题，而且也非常吓人，
86:39
and it burns people's houses down and it's terrible. Yeah, yeah. My view is that climate change
而且它会把人们的房子烧掉，太可怕了。对，对。我的看法是，气候变化
86:44
is sort of like a serious disease. Do you treat the symptoms? I do adapt or do you try to attack the
有点像一种严重的疾病。你是治症状？是去适应，还是试着解决
86:49
underlying thing? And the answer is, well, if it's serious enough, it's both. Yeah. Right. And so,
根本问题？答案是，嗯，如果它足够严重，那就两者都做。对。没错。所以，
86:53
yes, so we take sort of adaptation, especially around climate resilience very seriously at Google,
是的，所以我们在 Google 非常重视 adaptation，尤其是围绕 climate resilience，
86:59
and we try to give people informational tools to sort of help. That's part of the reason why we're
而且我们试着给人们提供信息工具来帮忙。这也是为什么我们
87:03
working on weather and then sort of cyclone prediction and things. So it all actually hangs
在做天气，然后还有 cyclone prediction 之类的。所以这一切其实都
87:08
together. So it's more than just you write, it's more than just interventions. It's climate resilience,
连在一起。所以它不只是你写的东西，不只是 interventions。它还是 climate resilience，
87:12
too. Yeah, having lived through four or five fire seasons on the west coast, they can be
也是。对，在西海岸经历过四五个火灾季之后，它们可以变得
87:18
quite, quite nasty. And it used to not be, I mean, I have a cabin up in the Sierra Nevada mountains,
相当、相当糟糕。以前可不是这样，我是说，我在 Sierra Nevada mountains 有个小屋，
87:24
and yeah, it used to be, oh, you know, summertime, it's nice. And then now it's like well, not every
然后是啊，以前是，哦，你知道，夏天的时候，挺好的。可现在呢，嗯，并不是每个
87:28
year, but yeah, there's like, you know, winter, spring, summer, and smoke. Yes. I want to stay to the
年，但是是啊，就比如说，你知道，冬天、春天、夏天，还有烟。对。我想待在
87:37
west of the fire line. Yes. And I, yes. So that is another thing that I'm interested in. And
火线的西边。对。而且我，对。所以那是另一件我感兴趣的事。而且
87:43
that Google's also very interested in his is climate resilience. Are these irons in through small
Google 也非常感兴趣的是 climate resilience。这些手头的事是不是足够小
87:49
enough that they could hitch a ride and like a micro satellite grid, like what it makes sense to
足够让它们搭个便车，比如说搭上 micro satellite grid？那怎样才说得通
87:54
would it be almost cheaper just to hire? I intend to pay someone who's watching a consolation.
那是不是几乎更便宜，直接雇人？我打算付钱给某个在看 consolation 的人。
88:00
Oh, they're not that small. They're not small. The thing is you need refrigeration.
哦，它们没那么小。它们不小。问题是，你需要 refrigeration。
88:05
Because it's midwave IR, so you have to you have to keep it cool. So these would have to be their
因为它是 midwave IR，所以你得，你得让它保持冷却。所以这些就得是它们
88:10
own satellites. They're not super large. They're not like the, you know, the satellites in
自己的 satellites。它们不是超级大。它们不像，你知道，那些 satellites 在
88:16
and choosing to store better giant monsters because of all the optics and who knows what, but
然后选择存储更好的巨型怪物，因为所有的 optics 还有谁知道什么，但是
88:22
yeah. And they have, you know, they're basically just ironsensors with a resolution of five by five.
嗯。而且它们有，你知道，基本上就只是 ironsensors，resolution 是 5x5。
88:28
No, that's the other cute thing is the resolution is about 50 by 50 meters. But you can use super
不，另一个有趣的地方是 resolution 大概是 50x50 米。但你可以用 super
88:34
resolution because it's essentially multi spectral and you sort of know where fires are. And you have,
resolution，因为它本质上是 multi spectral，而且你大概知道火在哪里。然后你有，
88:38
so yeah, there's a fair sprinkling of AI and half of them to to reach that five by five meter.
所以是的，里面撒了不少 AI，其中一半是为了达到那个 5x5 米。
88:45
When you also have, you have a convolution over the what you're reading out, right?
当你还有，你对你读出的东西做一个 convolution，对吧？
88:50
The, I forget the frame rate. The satellite is moving. I don't remember what the point spread
那个，我忘了 frame rate 是多少。卫星在移动。我不记得 point spread function 是什么了。
88:57
function is. I'm sorry. I know. But you're right. They do move. But I don't remember how fast they,
抱歉。我知道。但你说得对。它们确实在动。但我不记得它们有多快，我不记得它们有多快，这个也被叫做 broom sensor。
89:03
I don't remember how fast they, this has also been called a broom sensor. So there's this funny
所以有个挺有意思的事，就是想让你把 spectrum 往一个方向铺开。
89:07
thing of trying to you kind of spread out the spectrum one way. And there's also it's somewhat
而且还有，这其实有点复杂。它不只是一个，像 polar ride 那样的东西。
89:14
complicated thing. It isn't just a, like a polar ride. It's a complicated sensor. Yeah.
它是个复杂的 sensor。对。
89:19
Sort of switching gears a little bit. You know, you've been at the intersection of AI and science
稍微换个话题。你知道，你在 AI 和科学的交叉领域已经挺长时间了。
89:23
for quite some time. I think you've sort of wound your way into and out of it back and forth.
我觉得你有点在这个领域里进进出出、来回绕。
89:29
How do you see the field has evolved? Because I feel like it's evolving very quickly now.
你怎么看这个领域的发展？因为我觉得它现在发展得非常快。
89:34
And like, what are the sort of lessons that you've learned that you think the community has learned?
而且，比如说，有哪些经验教训是你学到的、并且觉得整个社区也学到了的？
89:39
And how do you think this should change if you were a young scientist or young practitioner?
那你觉得，如果你是一个年轻的科学家或者年轻的从业者，这应该怎么改变呢？
89:44
How should this change what, how you should approach, you know, the future?
这应该怎么改变你，嗯，面对未来的方式呢？
89:49
Well, I think there has been a phase change in the last 12 to 18 months. So I mean, a lot of
嗯，我觉得过去 12 到 18 个月里发生了一次阶段性的转变。所以我是说，很多
89:56
what we used to do, as I said, was build these specialized models to solve
我们过去做的事情，就像我说的，是构建这些 specialized models 来解决
90:04
initial problems. And if you think that's your job, it's kind of fun. You find a problem,
最初的问题。如果你觉得那就是你的工作，其实还挺有意思的。你发现一个问题，
90:08
you solve it, you find another problem, you solve it. But now we have these much more general AI
你解决它，再发现另一个问题，再解决它。但现在我们有了这些更加通用的 AI
90:14
things. And I think the whole AI for science community is kind of still feeling around.
东西。而且我觉得整个 AI for science 社区还在摸索。
90:20
The fact that they're working is so new that collectively, we're not sure, like, what's the best
它们真的能奏效这件事太新了，以至于从整体上来说，我们都还不确定，比如说，什么才是最好的
90:25
thing to do? Or maybe there's no one best. Maybe there's a tool chain. And I think we're all
做法？或者也许并不存在唯一最优解。也许存在一个 tool chain。而且我觉得我们所有人
90:29
trying to figure out, like, what should we do? And so there's a question of what should young scientists
都在试着弄清楚，比如，我们该做什么？所以就有了一个问题：年轻科学家应该
90:37
do? I think it would be, you know, I have a son who's just turned 21. And he's really into both
做什么？我觉得，你知道，我有个儿子刚满21岁。而且他真的同时很喜欢
90:46
sort of AI encoding and chemistry. And I look at him, I think he's doing the right thing because
算是 AI encoding 和化学。我看着他，我觉得他做的是对的事，因为
90:52
he's both learning a lot, trying to be a domain expert about RNA. But he's also sort of using
他既在学很多东西，努力成为 RNA 方面的 domain expert。但他也在某种程度上用
90:59
vibe coding and using all the tools. I think that's the right answers is because everyone's
vibe coding，并且在用所有工具。我觉得那就是正确答案，因为每个人都在
91:05
figuring it out. Still be a deep domain. I don't think domain expertise is going away because it
摸索清楚。仍然会是一个很深的 domain。我不觉得 domain expertise 会消失，因为它
91:11
goes back to a lot of people who said it goes back to taste and trying to figure out how people get
这又回到很多人说的：这归根结底是 taste，以及试着搞清楚人怎么能不做那些、那些申请 grant 的工作就获得 taste。这是个很有意思的开放性问题，但也要有成熟的 domain expertise，同时我会说，也要去玩、去尝试所有可用的不同工具，因为并不是说，哦对，我们知道会发生什么，那些聪明的老前辈就知道 what's the egg。不，我们也在实验。所以，对，所以我会说，一定要培养 domain expertise，试着用这些工具，尽你所能去解决重大而困难的科学问题。还有一个巨大的悬而未决的问题：你究竟该怎么处理实验室里的实际动手工作，因为它不会消失，因为实验才是 ground truth，而且是个瓶颈，还是个瓶颈。
91:15
taste without doing all the, the grant work. That's an interesting open question, but a developed
不用做所有那些，那些 grant work 就能有品味。这是个有趣的开放问题，但一个成熟的
91:20
domain expertise, but also try and play, I would say, with all the different tools that are available
domain expertise，但也要试着去玩，我会说，玩所有可用的不同工具
91:27
because it's not like, oh, yes, we know what's going to happen and the smart old people are knowing
因为并不是说，哦，对，我们知道会发生什么，那些聪明的老人们知道
91:32
what's the egg. No, we're experimenting too. And so, yeah, so I would say, definitely develop
什么是蛋。不，我们也在做实验。所以，对，所以我会说，一定要发展
91:39
domain expertise and try to use these tools and try to solve big hard scientific problems as best
domain expertise，并试着使用这些工具，试着去解决重大艰难的科学问题，尽可能
91:46
you can. There's still the huge open issue about what do you actually do about physical lab work
你能。仍然有一个巨大的开放问题：对于物理实验室工作，你实际上该怎么办
91:50
that is not going away because experiments are the ground truth and a bottleneck and a bottleneck.
它不会消失，因为实验是 ground truth，也是瓶颈，还是瓶颈。
91:58
I mean, people are talking about lab and the loop, but that's still very, very, very open because
我的意思是，人们在谈论 lab 和 loop，但那仍然非常、非常、非常开放，因为
92:04
how do you know and has, as far as I know, a general lab that does everything. There's a lot of
你怎么知道呢？而且据我所知，并没有一个什么都做的通用 lab。有很多
92:09
very specific labs that are controllable. So I think it's just we've gone through this phase change.
非常具体、可控的 lab。所以我觉得，我们只是经历了一次 phase change。
92:16
It seems super exciting. Again, I would advise people to play with whatever tools are available
这看起来超级令人兴奋。再说一次，我会建议人们去玩任何可用的工具
92:23
and to develop sort of deep domain expertise and taste to the extent you can. And I would advise
并在力所能及的范围内培养某种深度的 domain expertise 和 taste。而且我会建议
92:29
people also not to be scared and largely try stuff. You know, I'm always happy. We have student
人们也不要害怕，并且大体上多去尝试。你知道，我总是很开心。我们有学生
92:36
researchers at Google and they come and they do sort of wild and crazy things and that's always
研究员在 Google，他们会来，然后做各种疯狂的事情，而那总是
92:40
just delightful. So yeah, people should be trying sort of wild and crazy things and see what happens.
非常令人愉快。所以是的，人们应该去尝试各种疯狂的事情，看看会发生什么。
92:45
This may be a question without an answer, but when I think about how I developed expertise and how
这可能是一个没有答案的问题，但当我想到我是如何发展出专业能力的，以及
92:51
a lot of people developed expertise, it was by starting with a simple, defined problem and then
很多人是如何发展出专业能力的，都是从一个简单、定义明确的问题开始，然后
92:56
hammering it and then in that process of exploration, you learn more and you know, some ways you go
反复死磕它，然后在那个探索过程中，你学到更多，你知道，有些方向你会
93:02
broader, some ways you go deeper, but you still the process of just banging your head against the
走得更广，有些方向你会钻得更深，但你仍然有那个过程，就是拿头去撞
93:07
problem, which now would be instantly solvable, teaches you the skills you need to solve harder
那个问题，现在可能瞬间就能解决，它教会你解决更难的
93:13
problems. But what advice would you give to your son for that? You know, I don't know. Maybe it's a bit
问题所需的技能。但对此你会给你儿子什么建议？你知道，我不知道。也许这有点
93:17
like hiking, which is yes, I mean, you obviously can't drive everyone or you could you could drive up
像徒步，是的，我的意思是，你显然不能开车带所有人，或者你可以，你可以开车上
93:22
the mountain. Yeah, or you could hike up the mountain and maybe it's okay, even fun to occasionally
那座山。对，或者你可以徒步上山，也许偶尔这样也没关系，甚至挺好玩
93:27
hike up the mountain, even if you can drive up the mountain. Yeah, I mean, you're old days and
徒步上山，哪怕你能开车上去。是啊，我是说，你说的过去，
93:31
it sounds like a real thing in the old days. In the old days of six months ago. No, no, I was even
听起来像是过去真实存在的东西。就在六个月前的过去。不，不，我甚至
93:36
thinking of in the old days of the 80s and 90s. Like, you know, a lot of people take you like,
在想 80 年代和 90 年代的过去。就像，你知道，很多人会跟你说，
93:40
oh, there's open source packages. There's there's can learn. There's there's all sorts of things.
哦，有 open source packages。有，有可以学的。有，有各种各样的东西。
93:44
We didn't have that. I had to write my own numeric library. I had to write my own machine learning.
我们那时候没有这些。我得自己写 numeric library。我得自己写 machine learning。
93:48
I've written boosting probably agree in it four times, four different languages. And so now I know
我写过 boosting，大概用四种不同的语言写过四次。所以现在我知道
93:54
boosting, you know, it's like, and so maybe not taking the totally easy, but obviously, I mean,
boosting，你知道，就像，所以也许不是走最轻松的路，但显然，我的意思是，
94:00
there's this trade off like, oh, but I want to be as efficient and productive as possible. Yes,
这里有这种 trade-off，就像，哦，但我想尽可能高效、尽可能高产。是的，
94:04
but you also have to develop the muscles. So it's a little bit maybe like being an athlete, like
但你同时还得把肌肉练出来。所以这有点像当运动员，对吧？有些场合、有些时候你其实是在做 exploit，是在拼命跑快。然后还有 training time。所以也许人就是得 train。而且完全有可能，如果你花时间真的一直猛干、下苦功，哪怕在那儿进展更慢，长期来看也会给你更大层面的，你知道，生产力带来回报。就像即使从局部看，那一刻你没有通过把它们彻底 exploit 出来达到最大产出，这也会反哺到某些东西上。我希望如此。我希望我不知道的那件事是……我希望人们在职业生涯里，这很难，因为对吧，整个世界似乎都想要 optimize。
94:08
there are places, there are times when you're actually doing exploit, when you're trying to run as
有些地方、有些时候，你确实是在做 exploit，是在试着尽可能
94:12
fast as you can. And then there's also training time. And so maybe people just have to train.
快地跑。然后还有 training time。所以也许人们就是得 train。
94:17
And it's entirely plausible that if you spend time actually hammering away and doing the hard
而且完全有可能，如果你花时间真的埋头苦干、做那些艰难的
94:22
work, even if it goes slower there, that pays dividends into your larger, you know, productivity
工作，即使那边进展更慢，这也会给你更大的、你知道的，生产力
94:29
long term. Like even if locally, that one moment, you are not being maximally productive by not
带来长期回报。就像即使局部地看，那一个时刻，你没有因为不
94:34
exploiting it out of them, that feeds into something. I hope so. I hope the thing I don't know is
把它们身上的东西 exploit 出来而达到产出最大化，那也会反哺到某个东西上。我希望如此。我希望，我不知道的那件事是
94:40
I hope people in their careers, it's hard because right, the whole world seems to want to optimize
我希望人们在他们的职业生涯里，这很难，因为对吧，整个世界似乎都想要 optimize
94:47
everything. And it's your fault. No, I don't. No, I don't. It's my fault, but it's just sort of the,
一切。而且这都是你的错。不，我没有。不，我没有。是我的错，但就只是那种，
94:54
the, you know, and sometimes you have to set aside time, like at Google, especially in my group,
那种，你知道，而且有时候你得留出时间，比如在 Google，尤其是我所在的组，
94:59
we have this concept of 20% time, which I still I very, very strongly in my own group try to protect.
我们有 20% time 这个概念，我在自己的组里仍然非常、非常努力地想要保护它。
95:05
It's like you can do whatever you, if you want to learn stuff, if you want to try stuff,
就像你可以做任何你想做的事，如果你想学点什么，如果你想试点什么，
95:09
you don't even have to tell me, you don't even have to tell me if I probably shouldn't tell me,
你甚至都不用告诉我，你甚至都不用告诉我，如果我大概不该告诉我，
95:13
you know, just do stuff for exactly for learning. And also because that's where the sort of creative
你知道，就纯粹为了学习去做点事。而且也因为那正是那种创意
95:17
juices are, I don't want to so occupy people's time where they have nothing like where they can't
灵感所在，我不想把人们的时间占得那么满，让他们什么都没有，像让他们不能
95:22
feel like they can play or learn or try new crazy things. So I know 20% time is unusual. And
觉得自己可以玩、可以学，或者去尝试新的疯狂东西。所以我知道 20% time 很不寻常。而且
95:30
there just seems to be this strong impetus in the world to just like like is that optimize and
就好像这个世界上有一股很强的驱动力，就是，像是要 optimize 然后
95:34
squeeze everything out. But you do lose something when you hyperoptize or overfit.
把一切都榨干。但当你 hyperoptize 或者 overfit 的时候，你确实会失去一些东西。
95:41
So you're overfit to productivity. Yes. So that's right. So I know that might be my advice,
所以你就是 overfit 到生产力上了。是的。所以没错。所以我知道那可能是我的建议，
95:46
might be swimming upstream against perhaps cultural norms. The thing that always comes up for me here
可能是在逆流而上，逆着也许算是文化常规的东西。对我来说，这里总是会冒出来的一件事是
95:51
is that I've, there's a, the problem is not stationary. There's a new skill set that will be the
就是，我有，有一个，问题不是 stationary 的。会有一套新的技能组合，那将会是
95:59
the right skill set for the future. And the question in my mind is always just tangling,
未来正确的技能组合。而我脑子里一直纠结的问题是，
96:06
okay, is this a skill that is an enduring skill that yes, or like maybe it wasn't enduring
好吧，这是一项能持久的技能吗？是的，或者，像是，也许它并不持久
96:14
yesterday, but today it will be enduring because like I've seen that, you know, like for just
昨天，但今天它会变得持久，因为就像我见过那样，你知道，就像只是
96:20
as an obvious one, you become sort of like a manager when you're, yes, when you're using a
最明显的一点是，当你……对，当你使用一个
96:25
absolutely, I tell people this transfer well, some of them don't, but it's a lot of them do.
绝对是，我告诉人们这个 transfer 得很好，有些人不行，但很多人确实可以。
96:30
And so that as a manager, you lose track of, of the details of what's going on and you trust
所以作为管理者，你会搞不清正在发生的事情的细节，然后你信任
96:37
your people or agents or whatever to have that managed so that they can report up to you and
你的手下或 agents 或随便什么来把那些管好，这样他们能向你汇报，并且
96:44
answer, you know, sort of the high level questions and get the judgment about the little things
回答，你知道，那种高层面的问题，并且对小事情做出判断
96:50
correctly. And, and so that is that what we've come to is that we're just like middle managers now.
正确地。而且，而且所以我们就到了这一步：我们现在就像中层管理者一样了。
96:57
Well, I mean, I don't know. Again, I have a little management work, but you can't, I don't,
嗯，我是说，我也不知道。再说，我做过一点管理工作，但你不能，我不，
97:05
you don't want to be an empty suit. In other words, because the things might get it wrong,
你不想当个空架子。换句话说，因为这些东西可能会搞错，
97:09
especially LM's that are sort of really weird, they don't make the same kind of mistakes that
尤其是那些有点特别奇怪的 LM，它们犯的错误跟
97:13
humans make. And so you can't, you know, fully trust them, you have to be rigorous and like,
人类犯的不一样。所以你不能，你知道，完全信任它们，你必须很严谨，然后
97:19
you know, poke at it and make sure, although you should be poking it software that you write yourself
你知道，去戳一戳它、确认一下，虽然你也应该去戳你自己写的软件
97:24
to, I mean, you shouldn't trust yourself. That's one thing I've learned. What did five and say,
我是说，你不该信任你自己。这是我学到的一件事。Feynman 怎么说来着，
97:28
you know, you absolutely can't fool yourself and you're the easiest person to fool.
你知道，你绝对不能骗自己，而你是最容易被骗的人。
97:33
Yeah. So, um, so that might be an, I don't know if I can quantify what's enduring, but somehow
是啊。所以，嗯，所以那可能是一个，我不知道我能不能量化什么才是持久的，但不知怎么地
97:41
fundamentalness. I mean, really learning a domain that is about the world, for example.
根本性。我是说，比如真正学习一个关于世界的领域。
97:47
So this is why I like, well, biology or physical sciences, I think those will, those are
所以这就是为什么我喜欢，嗯，生物学或者物理科学，我觉得那些会，那些是
97:52
enduring sort of fundamental things. Math is very enduring, but even things like rigor and checking
那些很持久、很根本的东西。数学非常持久，但即便是严谨和核查
97:58
and that sort of thing, which goes back to maybe management that, that you want to really make
以及那一类事情，也许可以追溯到管理，就是，你真的很想
98:03
sure that the LM's are producing the right things or they haven't cheated in some way. But again,
确保这些 LM 产出正确的东西，或者它们没有以某种方式作弊。但话说回来，
98:08
you should be doing that to yourself too. Yeah. So, I think there's some enduring and also just
你也应该对自己这么做。对。所以，我觉得有些东西是持久的，而且还有
98:14
the enduring value of sort of creativity and thinking out of the box. And so I think those are
创造力和跳出框框思考的那种持久价值。所以我觉得那些都是
98:19
enduring. I don't know. There's something very fundamental about all of those. So you mentioned
持久的。我不知道。所有这些都有某种非常根本的东西。所以你提到了
98:24
Feynman. Um, if you don't mind me changing gears. Yeah, yeah, yeah. I took a class from
Feynman。呃，如果你不介意我换个话题。对，对，对。我上过一门课，是
98:29
Feynman. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Yeah. Not just any class. Yes. The, the, the,
Feynman 教的。对。对。对。对。对。对。对。对。可不是随便什么课。是的。那，那，那，
98:32
his sort of physics of computation class. I did it with, in fact, Hopfield and, and, uh,
他那种 physics of computation 课。其实我是和 Hopfield 一起做的，还有，还有，呃，
98:39
Carver Mead. Uh, that was fun. At the time, I don't think, and it may be Feynman knew. I felt like
Carver Mead。呃，那挺有意思的。当时，我不觉得，而且可能 Feynman 知道。我感觉
98:44
none of us knew what even the problem was. I mean, I guess Feynman was trying to say, oh, let's do
我们谁都不知道问题到底是什么。我是说，我猜 Feynman 想说的是，哦，咱们来做
98:49
quantum simulation, which I guess is, uh, it turned out to be the right answer. But yes, at the time,
quantum simulation，我猜这就是，呃，结果证明那才是正确答案。但没错，当时，
98:54
it was, oh, well, I guess, uh, it was cool. First of all, it was, uh, DARPA funded it. So you
就是，哦，好吧，我猜，呃，那挺酷的。首先，它是，呃，DARPA 资助的。所以你
99:00
were supposed to go once a week to get a primary dinner. Uh, but I just went every week. Uh,
本来应该每周去一次，好拿一份 primary dinner。呃，但我就是每周都去。呃，
99:05
anyway, for the students. Yeah. So just for a little bit more context, this class was basically
反正，是给学生准备的。对。所以再补充一点背景，这门课基本上
99:11
the class right after Feynman. And I forget who else proposed the concept of a quantum computer
就是 Feynman 之后紧接着的那门课。我忘了还有谁也提出了 quantum computer 这个概念
99:17
without really knowing what it was, but knowing that there was some sort of, well, there was,
当时并不真知道那是什么，但知道有某种，嗯，有，
99:21
there was plenty of room at the bottom as I, which I think was in the very early. This, I should
底下还有很大空间，正如我……我觉得那是在很早的时候。这个，我应该
99:26
get an 82. Okay. Oh, that was when he taught it. And he did say there was plenty of room at the
拿个 82 分。好吧。哦，那是他教这门课的时候。他确实说过底下还有很大空间。
99:32
bottom. But the way the class was structured, it was, um, uh, it was like a guest lecture on
但课的结构是，嗯，呃，就像周二有一场客座讲座，
99:38
Tuesday. And then Feynman would stand up on Thursday and explain why that was all wrong.
然后周四 Feynman 会站起来，解释为什么那全错了。
99:46
Which is pretty fun. And then we encountered something, which other people have also encountered,
这还挺好玩的。然后我们遇到了一样东西，别人也遇到过，
99:50
something called the Feynman effect. Maybe he was so charismatic or something. He would explain
叫 Feynman effect。也许他太有魅力了还是怎么的。他会解释
99:54
things and you would say, yes, yes, I understand. And then you walk out to think, no, no, I don't,
一些东西，你会说，对，对，我懂了。然后你走出去一想，不，不，我没懂，
99:58
I didn't. So it was kind of fun. But a lot of the guests, people, I mean, maybe they sort of
我没有。所以还挺好玩的。但很多嘉宾，那些人，我是说，也许他们算是
100:05
showed the chaos. I mean, a lot of people, I think it was, uh, Danny Hillis came, uh, there
展现出了那种混乱。我是说，很多人，我觉得是，呃，Danny Hillis 来了，呃，有
100:11
was all these interesting guest lectures. So I would say in the union of all of them, I think
所有这些有趣的客座讲座。所以我会说，把它们全都合在一起看，我觉得
100:14
showed the sort of mass confusion of what was going on because there was a lot of, like, oh,
展现出了当时那种集体混乱，因为有很多，像是在说，噢，
100:18
should we make computers reversible? Because we have to make sure, you know, can they even be
我们该不该让计算机变得 reversible？因为我们必须确保，你知道，它们到底能不能
100:22
reversible? And can this sort of the bottom limit of, of the heat per operation B zero? Or is there
reversible？还有，这个每次操作的 heat 下限，能不能是 B zero？还是说存在
100:27
some sort of thermodynamic limit? That was like a big deal. I don't think that's, I don't think
某种 thermodynamic limit？那当时好像是个大事。我不觉得那是，我不觉得
100:31
that's a big deal now. Yeah. But you're talking about the land hour limit rate or not. What's on them?
那现在是个大事。对。但你说的是 land hour limit rate，还是不是？那它们上面有什么？
100:36
Uh, well, do it at the time. Most of the, we mean with all these, this question about, like, can
呃，嗯，当时就那么做吧。大部分，我是说，所有这些，这个问题，就像是，能不能
100:39
you have revered, like billiard ball computers? And can they be reversible and things? So,
你能有，像是，reversible 的 billiard ball computers 吗？而且它们能是 reversible 之类的吗？所以，
100:44
and also sort of like what kind of computing that's why Danny Hillis came, like, you know,
而且也有点像是，什么样的 computing，这就是为什么 Danny Hillis 来了，就像，你知道，
100:47
I think that was in the era of the connection machine. And what, what original thinking machines,
我觉得那是在 connection machine 的时代。还有，什么，什么最初的 thinking machines，
100:52
Miriam Roddy's, but the original one in the 80s. So he came. What was the state of, like,
Miriam Roddy's，但 80 年代那个最初的。所以他来了。那，像是，当时是什么状态，
100:58
general computation in 1982? Oh, I mean, at this point, you do a two-rounding era. We had zero
1982 年的 general computation？哦，我是说，这时候，你算是处在一个 two-rounding era。我们当时是 zero
101:05
I mean, I remember what I, sorry, I got there. Uh, I was cover meets, um, uh,
我是说，我记得我，抱歉，我到那儿了。呃，我当时 cover meets，嗯，呃，
101:10
Sister Edmond, we had a Vax 11750 that maybe did a MIP, one million, what a million operators.
Sister Edmond，我们有一台 Vax 11750，可能能做一个 MIP，一百万，什么，一百万 operators。
101:16
And the whole research group shared an 80 megabyte drive that was the size of a dishwasher.
整个研究组共用了一个 80 megabyte 的 drive，大小跟洗碗机差不多。
101:22
It was, it was very exciting. Uh, so what about complexity through you? What, what was it? Because I
那真是，那真是非常令人兴奋。呃，那 complexity 从你这边呢？那，那是什么？因为我
101:27
know that there's a lot of interest in quantum complexity and how it relates to gravity right now.
我知道现在大家对 quantum complexity 以及它和 gravity 的关系都很感兴趣。
101:33
And I wonder, I don't have a mind in my mind about when complexity, we could try to look it up.
然后我在想，关于 complexity 是什么时候，我脑子里也没有这个印象，我们可以试着查一下。
101:37
I don't think there's, of course, the whole, there's a whole hierarchy of, you're right,
我不觉得有，当然了，整个，有一整套 hierarchy，你说得对，
101:40
quantum complexity classes. I think that was developed after that, because this was in 1982,
quantum complexity classes。我觉得那是后来才发展出来的，因为当时是 1982 年，
101:47
in fact, I, I'm really not even sure there was a gate, but we never talked about the gate model.
其实，我，我甚至真的不确定当时有没有 gate，但我们从来没谈过 gate model。
101:51
There was no gate model of quantum. I mean, quantum gate. Yeah. Yeah. Yeah. Yeah. I forget. I'm
当时还没有 quantum 的 gate model。我是说，quantum gate。对。对。对。对。我忘了。我
101:55
sorry. I think a lot of those, yeah, I don't think those came out until, like, the, what, 90s or
抱歉。我觉得其中很多，对，我觉得那些直到，大概，那个，什么，90 年代才出来，或者
101:59
something. Yeah. I mean, I, I remember reading Gilson and Chong the classic quantum for the information.
什么的。对。我是说，我，我记得读过 Gilson 和 Chong 那本经典的 quantum for the information。
102:05
Yeah. Quantum information textbook, which, uh, which brings up a lot of those points. Now, I think
对。quantum information 教材，它，呃，它提到了很多那些点。现在，我觉得
102:10
that I, that textbook was written in what the late 90s are early 2000s. Yes. That's right.
那个，我，那本教材是在，怎么说，90 年代末还是 2000 年代初写的。对。没错。
102:15
And I think that was the first textbook, which, like, put down kind of the general knowledge of
而且我觉得那是第一本教材，它有点像把……的一般知识写下来的
102:20
the, the, the field, but I could be wrong. I guess I, I think now that when I was young, I don't
这个，这个，这个领域的，但我可能弄错了。我猜，我，我现在觉得，我年轻的时候，我不
102:25
get it, I take it just for granted that this is, this is the block that everyone. Yeah. I know
明白，我只是想当然地觉得这就是，这就是每个人……的那个 block。对。我知道
102:29
this is a profession. Well, you remember in the, in the early 80s, um, I mean, there was a whole
这是一门专业。嗯，你记得在，在 80 年代初，呃，我是说，有一整个
102:34
bunch of excitement around neural networks. And it's really interesting, because again, we really did
大家对 neural networks 兴奋得不得了。而且真的很有意思，因为再说一次，我们当时真的
102:39
not know what we were doing. No one knew what they were doing. There was, like, an interesting,
并不知道自己在做什么。没人知道自己在做什么。当时有点像，一个有意思的，
102:42
like, neural networks, then SVMs and neural networks. Oh, no, no, this is before that.
像，neural networks，然后是 SVMs 和 neural networks。哦，不不，这是在那个之前。
102:45
Oh, okay. This is in the 80s. Everyone said, everyone said, this is, it has the capability of
哦，好吧。这是在 80 年代。所有人都说，所有人都说，这个，它有能力
102:49
revolutionizing computing. But what does, well, I mean, there's a tremendous, um, excitement
彻底改变 computing。但它到底，嗯，我的意思是，有一种巨大的，呃，兴奋
102:55
around hot field networks. Uh, in fact, Neurips came out of, uh, because, uh, there was a workshop
围绕 hot field networks。呃，事实上，Neurips 就是源于，呃，因为，呃，当时有一个 workshop
103:03
at Snowbird that was not only private, but everyone tried to crash. And so they, they spotted up
在 Snowbird，那个活动不光是私人的，但每个人都想混进去。于是他们，他们就搞出了
103:08
Neurips. So that was snowbird is a, is a skiing trip with a, with a computation conference attached.
Neurips。所以那就是，Snowbird 是一个，是一个滑雪旅行，还附带一个，一个 computation conference。
103:12
That, that's right. That's why I learned to ski. Yeah. Yeah. I didn't notice skiing. And I kept
对，对，没错。
103:16
going. Anyway, that workshop came out of the Santa Barbara workshop in 1995, which came out of
这就是我为什么去学滑雪。
103:22
some local things at Caltech, which was called hopfests. So, so yeah, people thought, oh,
对。对。
103:27
wow, something involved. In fact, ironically, there was like, yeah, yeah, something about
我没注意到滑雪。然后我就一直继续。
103:31
associated memory. And if you actually dig down into what transformers are, they are associated
总之，那个 workshop 源自 1995 年的 Santa Barbara workshop，而那个又源自 Caltech 的一些本地活动，叫 hopfests。
103:36
memories. Yeah. So in fact, it was even a paper called, you know, uh, hot field networks are
所以，所以，对，人们会想，哦，哇，这里面有点东西。
103:40
all you need. Uh, so, yes, yeah, forget about that. That title was a reference to that paper, uh,
事实上，讽刺的是，当时有，对，对，关于 associated memory 的东西。
103:46
to the era here. So it's actually, it's like, oh, well, the whole thing is sort of come full circle.
而如果你真正深挖 transformers 是什么，它们就是 associated memories。
103:53
And in fact, I think we did, uh, we have, um, collectively as a, as a field, uh, revolutionized
而且事实上，我觉得我们做到了，呃，我们，嗯，作为一个领域，作为一个领域，呃，彻底革新了
103:58
computer science, but it was the, the sort of the hopes and dreams completely outstrip, uh,
computer science，但它其实是那种，那种希望和梦想完全超出了，呃，
104:03
the capabilities because again, we effectively had zero compute. Yeah. I think there's,
能力，因为还是那句话，我们实际上完全没有 compute。对。我觉得有，
104:08
the history of machine learning is different paradigms as like compute versus memory,
machine learning 的历史就是不同的范式，就像 compute 对 memory，
104:15
scaling versus data become available and different levels. That's right. And the fact that I think
scaling 对 data 变得可用，以及不同的层级。没错。而且我觉得
104:22
people don't realize that the reason why neural networks one is there that one compute limited
人们没意识到，neural networks 之所以赢了，就在于那个 compute limited 的
104:27
thing, although again, with, with now that we have transformers, things are getting memory
东西，不过话说回来，现在有了 transformers，事情又变得 memory
104:31
limited again, but, but, and the fact that they ride on top of a blast. And so the fact that
limited 了，但是，但是，而且它们其实是跑在 BLAS 之上的。所以这个事实就是
104:36
blast was being optimized by things like GPUs, I mean, I don't actually know if, in fact,
blast 正在被 GPUs 这类东西优化，我是说，其实我并不知道是不是真的，实际上，
104:41
I'm pretty sure the brains don't work by matrix multiply, but, but it was just that, that, that,
我很确定大脑不是靠 matrix multiply 运作的，但是，但是，就只是说，那，那，那，
104:46
that the algorithms co evolved with the, with the hardware. And, and that's, that's why we're here.
就是 algorithms 跟，跟 hardware 共同演化。而且，而且这就是，这就是我们为什么会走到今天。
104:52
Who knows? I mean, if we go down some other path where people really cared about some other
谁知道呢？我是说，如果我们走上另一条路，人们真正在意的是另一种
104:57
compute, who knows what would architecturally end it up with? I don't know. I mean,
compute，谁知道最终在架构上会变成什么样？我不知道。我是说，
105:01
didn't know a lot of the start with like, I think people were hacking play stations, a trait
一开始的很多事我也不知道，就像，我觉得人们当时在 hack play stations，a trait
105:04
to a train neural network or something or to do, I guess soup, maybe it was even before that.
去 train neural network 之类的，或者去做，我猜 soup，也许甚至在那之前。
105:08
It's for like super, there's a funny story by a friend of mine from grad school,
这有点像 super，我研究生院的一个朋友有个很好笑的故事，
105:16
Brian Katzenzar, a video where he, Brian, sorry, if I get this wrong by saying, he came to
Brian Katzenzar，有个视频，里面他，Brian，抱歉，如果我说错了，他来到 Nvidia，然后，然后一直很难打开局面。
105:22
Nvidia and, and was having a lot of trouble getting traction. And, and he basically, they were,
然后他基本上，他们，他们就是，他们加倍、三倍地下注在游戏上。
105:30
they were doubled, interpled down on, on gaming. And he basically, one day had an meeting with
然后他有一天和 Jensen 开了个会，说服了他，说，你知道，就像，这，我们有这么多人在用 CUDA 做 BLAS，基本上，尤其是做 deep learning，然后说服了 Jensen。
105:36
Jensen and convinced him, let's, you know, like, this is, we have all these people using CUDA for
他说，那大概就是一段 15 分钟的对话，然后他们第二天就让整个公司转型了。
105:43
Blas, basically, and, and for deep learning in particular and convinced Jensen. And he said, it was
我是说，我有朋友，Dave Kirk，他是第一任首席科学家，还有 Bill Dalai，他们俩都是我读研时的朋友。
105:50
like a 15 minute conversation and they pivoted the whole company to the next day. I mean, I have
所以，对，但我不知道具体是怎么回事。
105:56
friends, Dave Kirk, who's the first chief scientist and, and Bill Dalai were both friends of mine
现在，记住，人们，
106:01
from grad school. So, yes, but I don't know the details of way. Now, remember that people,
从研究生院开始。所以，是的，但我不知道方式的细节。现在，记住，人们，
106:08
like in, in Hinton's group even in the around 2010, they were using GPUs to do the deep learning,
就像，在，在 Hinton 的组里，甚至大概在 2010 年左右，他们就已经在用 GPUs 做 deep learning 了，
106:15
but, but even as a, even in, I would say 2007, 2008, they weren't that much faster than CPUs.
但，但就算，就算在，我会说，2007、2008 年，它们也没比 CPUs 快多少。
106:21
Again, they, they, they only started really exceeding. In fact, I think this not a coincidence
再说，他们，他们，他们是到后来才真正开始超过。其实，我觉得这不是巧合
106:26
in the era of, you know, the original image net and some of the speech, some of the speech recognition
在，你知道，最初的 ImageNet 那个时代，还有一些 speech，一些 speech recognition 的
106:30
stuff. So, that's, I think this not a coincidence that they really researched when, when, when
东西。所以，我觉得，这不是巧合，他们真的研究了，是在什么时候，什么时候，什么时候
106:35
GPUs passed CPUs. So, there's that, that's not a coincidence either. Yeah. I really want to know,
GPUs 超过了 CPUs。所以，就是这样，那也不是巧合。对。我真的很想知道，
106:40
how does one get the opportunity to name an asteroid? Oh, well, again, is Caltech. There was a,
一个人怎么才能有机会给一颗小行星命名？哦，好吧，又是，是 Caltech。有一个，
106:48
there was so, there were some wonderful people, Jean and Carolyn Schumaker, and they were teaching
有，所以，有一些很棒的人，Jean 和 Carolyn Schumaker，他们在教书
106:53
a class in, in planetary science. I like planetary science. And so, yes, as part of that class,
一门，呃，planetary science 的课。我喜欢 planetary science。所以，是的，作为那门课的一部分，
107:00
they took us, they took us through the sort of asteroid discovery thing. Now, this is in the 80s.
他们带我们，他们带我们走了那种 asteroid discovery 的流程。现在，这是 80 年代。
107:06
So, now, there's all sorts of amazing, amazing systems. Well, there's for a while, this is
所以，现在有各种超厉害、超厉害的系统。嗯，有一阵子，这是个
107:11
something called linear, which automated the thing. But this was before anything was automated.
叫 linear 的东西，它把这件事自动化了。但这是在任何东西被自动化之前。
107:16
So, yeah, it was just part of a class, and what you would do is you go, the steps, even though we
所以，是的，这只是课的一部分，而你要做的就是，你要走那些步骤，尽管我们
107:20
did it out of order in the class, but the steps we used to do, this was 40 years ago, is you would
在课上没按顺序做，但我们以前做的那些步骤，这是 40 年前了，就是你会
107:25
go to Palomar. There would be a fast telescope, which literally now, museum pieces in their
去 Palomar。那儿会有一台 fast telescope，现在它真的成了他们的
107:30
visitor center, but at the time it was a real thing. You would put a piece of film in it. You would
游客中心里的博物馆展品，但当时它是真实存在的东西。你会往里面放一张胶片。你会
107:34
take a picture, and then you'd wait for a few more minutes to take a same picture of the same
拍一张照片，然后你会等几分钟，再拍一张同样的
107:38
point of sky. You put in a stereoscope, like back at Caltech. You would see if anything
天空中的点。你把它放进 stereoscope 里，就像在 Caltech 那时候一样。你会看看有没有什么东西
107:42
flee would look around. You would see if anything floated, because it's this, it would move by a tiny
逃跑，会四处看。你会看到有没有什么东西漂浮，因为它就是这个，它会移动一点点
107:48
bit, so it would pop up. Yeah, you see it. Because it's different in different eyes, then you would
，所以它会冒出来。对，你看到它了。因为它在不同的眼睛里不一样，然后你就会
107:52
be able to see it as something that was projected in a different place. That's right. So, you would pop
能够把它看成是投射在不同位置的东西。没错。所以，你会
107:57
out at you literally. And then you would go to a measuring microscope, and you would take measurements
真的在你眼前跳出来。然后你会去用 measuring microscope，你会测量
108:03
of known star references, and where this floater is. And so, if you got accurate enough
已知的 star references，以及这个 floater 在哪里。所以，如果你测
108:08
measurement, then you would send it off to a person, I don't think it's him anymore.
得足够准确，然后你会把它发给一个人，我觉得已经不是他了。
108:12
There's the same Brian Marson at the Minor Planet Center in Arizona. And then he had a big
在 Arizona 的 Minor Planet Center，也有同一个 Brian Marson。
108:18
software system to do, to sort of piece together these called apparitions. And if you found,
然后他有一个很大的 software system 要做，基本上就是把这些所谓的 apparitions 拼在一起。
108:23
and if you happened to have done an observation, which was the final operation, which allowed his
如果你发现了，而且你碰巧做了一次 observation，而那次 observation 是最后一步，让他的 software 能把它连成一个大的 orbit，那你就能拿到 discovery rights。
108:27
software to connect it into one big orbit, then you would get discovery rights. You could name the
你就可以给 asteroid 命名。
108:32
asteroid. But now it's like amazing. There's this observatory now in Chile called the Vera Ruben
但现在这就很惊人。
108:39
Observatory. And there's this amazing telescope called the Simone Survey Telescope.
现在 Chile 有一个 observatory，叫 Vera Ruben Observatory。
108:44
It essentially automates this process. It essentially can just take many frames of the sky.
还有一个很惊人的 telescope，叫 Simone Survey Telescope。
108:49
It's an utterly stunning instrument. And so, they discovered 11,000 asteroids in six weeks.
它基本上把这个 process 自动化了。
108:54
Oh, wow. Yeah. So, it's now, well, it depends, I guess, stun.
哦，哇。是啊。所以，现在嘛，呃，得看情况，我猜，stun。
109:00
That's not off. Yeah. So, but there are probably millions of asteroids.
那倒没错。是啊。所以，但可能有几百万颗 asteroids。
109:04
There's still an opportunity. It's true. But I don't even know if they bother me.
还是有机会的。确实。但我甚至都不知道它们会不会让我烦。
109:09
So, I don't know, maybe you don't need, I don't know. But people are doing occult, sorry,
所以，我不知道，也许你不需要，我不知道。但人们在搞 occult，抱歉，
109:12
I can now talk about this forever. There's this stuff called occultation. Like, one of my asteroids,
我现在可以一直聊这个了。有种东西叫 occultation。比如，我的 asteroids 里有一颗，
109:17
I was talking, I was sitting email to a amateur. If an asteroid happens to pass in front of a star,
我当时在聊，我正坐着给一个业余爱好者写邮件。如果一颗 asteroid 恰好从一颗恒星前面经过，
109:24
it dips you just like when they discover exoplanets. But if you're super lucky, it'll dip
它就会变暗，就像他们发现 exoplanets 时那样。但如果你超级走运，它会变暗
109:30
and then it'll dip again because there's a moon. And so, one of my asteroids, this amateur found
然后又会再变暗一次，因为有一颗卫星。所以，我的 asteroids 里有一颗，这个业余爱好者发现了
109:35
a little moon around it. So, that was, I think, on the asteroid. Yeah. Interesting.
它周围有颗小卫星。所以，我觉得那是在小行星上。对，挺有意思的。
109:42
So, that's pretty cool. So, there's still room for fear. Your asteroid has a moon too.
所以，这挺酷的。所以，还是有让人害怕的余地。你的小行星也有卫星。
109:46
Yes. In fact, apparently. It's actually actually just a little bit of a story there
是的。事实上，显然是这样。其实，其实这里面还有个小故事。
109:49
because it's not really my asteroid that I found two of them. I named one of them after my dad.
因为其实那不是我的小行星——我发现了两颗。我把其中一颗以我爸爸的名字命名了。
109:54
I waffled. So, Carolyn named it after a professor at University of Washington. And I said,
我犹豫了。所以，Carolyn 以 University of Washington 的一位教授的名字给它命名了。然后我说，
109:59
oh, but I wanted to name it. And so, I wind it. And she was nice. And so, she gave me one of hers.
哦，但我想给它命名。所以我就抱怨了一下。她人很好。所以，她就把她的一颗让给了我。
110:04
And so, then the name that one after my mom. And that's the one that has the moon.
然后，我就把那颗以我妈妈的名字命名了。而那颗就是有卫星的。
110:07
Oh, okay. So, the asteroid named after my mom has a moon. How big is it?
哦，好吧。所以，以我妈妈命名的那颗小行星有卫星。它有多大？
110:13
It's like the main body they think is about four kilometers across. I'm trying to get the units
就像他们认为的主体，大概横跨四公里。我在努力把单位
110:19
right. And the moon is about one kilometer. So, it's actually, it's not, it's a pretty big
弄对。而且那颗卫星大概一公里。所以，其实，它不是，它是个相当大的
110:24
binary. They're not quite twins. But, yeah. Did other people in that class?
binary。它们不算完全是双胞胎。但是，对。那个班里其他人呢？
110:29
Just every three years. I think there was one other person. It's a little bit of a
每三年才一次。我觉得还有另外一个人。这有点
110:33
crapshoot. Yeah. And the fact that I found two was quite unusual. Was there a reason why you,
碰运气。对。而且我找到了两个，这相当不寻常。是不是有什么原因，你，
110:37
was it just pure luck? Or was it just pure luck? They're all they do it trying to do it.
只是纯粹靠运气吗？还是只是纯粹靠运气？他们全都在做，都是在试着做。
110:41
I know. I know. I know. I tried to be careful. Yeah. Yeah. How does one become, get an
我知道。我知道。我知道。我尽量小心了。对。对。一个人怎么成为，拿到
110:46
Oscar? Yeah. I'm kind of your word. Well, again, maybe it's at the right place. I, my,
Oscar？对。我有点，你的话。嗯，再说，也许是在对的地方。我，我的，
110:52
he says, via his name, L bar. And I was a student intern actually in 1986 at a place called
他说，他的名字是 L bar。其实我1986年时是个学生实习生，在一个叫
110:58
Shumberjé, which is a, well, this can be a place, but they had an AI lab. And so they were,
Shumberjé 的地方，那是一个，嗯，这可以算个地方，但他们有一个 AI 实验室。于是他们，
111:02
we were all doing sort of computer graphics research. And we were all thinking, you know,
我们都在做某种 computer graphics 研究。我们都在想，你知道，
111:07
at the time, this was revolutionary. I realized it's now considered incredibly poor. Like,
当时，这挺革命性的。我意识到现在这被认为糟糕得难以置信。就像，
111:11
you could actually use like physics simulators to make computer graphics movies. And at the time,
你其实可以用像 physics simulators 这样的东西来制作 computer graphics 电影。而当时，
111:16
I was like, wow, that's really cool. So, I said, you know, you could use the theory of elasticity
我心想，哇，那真的很酷。所以我说，你知道，你可以用 theory of elasticity
111:20
to make floppy things. And, and so I said, a lot of sort of like, here's the theory of elasticity.
来做软趴趴的东西。然后，然后我说，很多类似这样的，这里是 theory of elasticity。
111:24
And I wrote a last simulator. And I made, you know, fabric and, you know, stretchy things and,
然后我写了一个 last simulator。然后我做了，你知道，布料和，你知道，有弹性的东西，和，
111:28
and stuff. So they said, oh, wow, that's cool. And so you sort of descendants of that became
之类的。然后他们说，哦，哇，这挺酷的。然后，算是从那个衍生出来的东西，后来就变成了
111:34
a lot of the physics simulators that people used in, in, you know, Pixar and their various movies.
很多人们在，在，你知道，Pixar 和他们的各种电影里用的 physics simulators。
111:39
So yes, I tell, I tell interns sort of have jokingly, well, if you do a really good jobs in
所以是的，我跟，我跟实习生有点半开玩笑地说，嗯，如果你在
111:43
an intern, you can get it. So how, like, how long was that between the time you did that work?
实习期间做得特别好，你就能拿到它。所以，那，从你做那个工作到……有多久？
111:51
And then 20 years. It was 20 years. Which is actually not a typical, right? Because they want to,
然后20年。就是20年。这其实并不典型，对吧？因为他们想，
111:56
when they give you a category where they want to make sure, like, oh, yeah, it's sort of well-used
当他们给你一个类别时，他们想确保，就像，哦，对，它算是被广泛使用
111:59
and everyone uses it and stuff. So yeah, but by, of course, obviously, like in that 20 years,
而且大家都在用之类的。所以是的，但是，当然，很明显，就像在那20年里，
112:03
like, well, everyone doesn't solve yes, but, you know, this was specific work done for a specific
嗯，好吧，不是所有人都解决了，是的，但是，你知道，这是为某个特定的
112:07
movie. No, it was like a paper in cigarette. It was, it was a paper. And then Pixar just became,
电影。不，它就像香烟里的纸。就是，就是一张纸。然后 Pixar 就变成了，
112:13
like, I think, somewhat important core to, like, a lot of the, yes. And in fact, some of my friends,
就像，我觉得，在某种程度上成了，很多……对，很重要的核心。而且事实上，我的一些朋友，
112:17
in fact, a lot of my friends did a lot of these simulators. So yeah. And curious about, since you
其实，我的很多朋友做了很多这类 simulators。所以，是的。而且我很好奇，既然你
112:22
have been quantum computing adjacent and worked on quantum computing directly at Google,
一直和 quantum computing 沾边，还在 Google 直接做过 quantum computing，
112:27
applied science, it's, well, I think you're not currently on that. But I'm curious to see,
applied science，这个，嗯，我觉得你现在不在做那个了。但我很好奇，
112:31
what are you still dabbling in? Oh, you're still dabbling. Yeah. Yeah. Yeah. Yeah. So what do you see
你现在还在涉猎什么？哦，你还在涉猎。对。对。对。对。所以你觉得
112:36
the trajectory of quantum computing over the, over the years going? Because this is one of those
quantum computing 这些年来，这些年来会怎么发展？因为这是那种
112:42
things, which I guess kind of like fusion was sort of had a sense of first like being very exciting.
东西，我猜有点像 fusion，一开始有种非常让人兴奋的感觉。
112:47
And then seeming like it wasn't going anywhere for a long time. And then maybe now we're seeing
然后有很长一段时间，它好像都没什么起色。然后也许现在我们又看到了
112:51
hints again of it being exciting. And I'm, or maybe that's like my sort of quantum adjacent.
一些迹象，让人觉得它又开始让人兴奋了。而我，或者说那可能就是我这种跟 quantum 沾点边的角度吧。
112:58
I think a lot of people have gone through that. I tend to average, I tend to average things out
我觉得很多人都经历过这个。我倾向于平均，我倾向于把事情平均一下
113:03
over the day. I guess I'm old enough now or I sort of average things out over the decades. And
在一天里。我猜我现在年纪够大了，或者我会把事情放在几十年尺度上平均来看。而且
113:06
it's just on progress. But I mean, like, see, going from, you know, starting from
它就是一直在进展。但我的意思是，就像，你看，从，你知道，从
113:12
Feynman just trying to figure out even what this means as a concept all the way up to now where,
Feynman 还在试图搞清楚这作为一个概念到底意味着什么，一直到现在的
113:17
you know, I guess there's this recent willow result of quantum error correction, which actually
你知道，我猜有最近这个 willow 的 quantum error correction 结果，它实际上
113:21
seems genuinely capable. The right, the right scaling laws and stuff. I mean, does this,
看起来真的有能力。正确的，正确的 scaling laws 之类的东西。我的意思是，这，
113:27
is this like a, do you still think that, or would you say that quantum is 20 years away,
这是不是，呃，你现在还是这么认为吗，还是你会说 quantum 还要 20 年，
113:33
or even till like some simple, but practical algorithm, which actually succeeds? Or I mean,
或者甚至要等到某个简单但实用的 algorithm 真正成功的时候？或者我是说，
113:40
is it, do you see this accelerating? Or do you think this is still going to be something
你觉得它是在加速吗？还是你觉得这仍然会是某种
113:44
that's log? Because I mean, there's a lot of areas where we saw, oh, this is very, you know,
log 式的东西？因为我是说，有很多领域我们看到，哦，这个非常，你知道，
113:49
this advanced very quickly. And it was very unexpected. And I'm wondering if this is a thing that you
它发展得非常快。而且非常出乎意料。我在想，这是不是一件你
113:54
think will be quick or will not be or is this, I don't know. I think sort of in between, it's not
觉得会很快、还是不会很快的事，或者这是，我不知道。我觉得算是介于两者之间，这不是
114:01
a purely software thing because we need to build systems that are large enough and stable enough
一个纯 software 的事，因为我们需要构建足够大、足够稳定的 systems
114:05
to be able to, you know, be quantum computer instead of a quantum apparatus. We're in there
才能，你知道，成为 quantum computer，而不是一个 quantum apparatus。我们就在其中
114:10
what they call the nisk era. That was a, is he in for intermediate scale quantum, which was,
他们所谓的 NISQ era。那是一个——NISQ 也就是 intermediate scale quantum，那个……
114:16
I think, coined by John Prescott. It's not a very good term. Sorry, but I guess that's the
我想，是 John Prescott 造出来的。这不是个很好的术语。抱歉，但我觉得这就是我们
114:20
term we have. And there are, you know, the quantum team made this wonderful paper, which I think
有的说法了。而且，你知道，quantum 团队做了一篇很棒的论文，我想
114:27
I'm co-author on one of many for something called the quantum echoes algorithm, which could be
我是它的 co-author 之一，这篇论文讲的是叫 quantum echoes algorithm 的东西，它可能
114:34
applicable now, which fundamentally it's it's in the style of of Feynman's proposal, which is
现在就能用，它本质上其实属于 Feynman 那种提议的风格，也就是
114:40
essentially it's, well, the technical term is you could try to fit Hamiltonian to observe data
基本上，嗯，技术上的说法是，你可以试着把 Hamiltonian 拟合到观测数据上
114:46
like it in NMR. What that means is, yeah, you have a physical model that's parameterized and you
就像在 NMR 里那样。那意思就是，对，你有一个 parameterized 的物理模型，然后你
114:52
use the, use the quantum computer to kind of adjust the parameters and try to figure out it
用，用 quantum computer 来调这些 parameters，试着把它搞清楚
114:58
inside of a loop to what the right, so you could, for example, decode in NMR parameters. So that
在一个循环里面，朝着正确的方向，所以你可以，比如说，在 NMR 参数里解码。所以那
115:03
is a practical thing that's used now. The question is, it will be big enough to make breakthroughs.
是一个现在就在用的实用东西。问题是，它会不会足够大，能做出突破。
115:09
That's still TBD. The quantum team at Google has been very executing again, amazingly well
那还是 TBD。Google 的 quantum 团队又一次执行得非常好，惊人地好
115:18
against a Broadmap that that Hartwood Nevin laid out a few years ago and they're just continuing
对照 Hartwood Nevin 几年前制定的 Broadmap，他们只是在继续
115:25
to march down this thing where they're scaling up. And when they hit their last milestone,
沿着这条路前进，他们在 scaling up。当他们达到上一个里程碑时，
115:30
they should people have a quantum computer that does amazing things and they're continuing to
他们应该让人们拥有一个能做惊人事情的 quantum computer，而且他们继续
115:35
march it along. So yeah, it's, I think on the scale of a few to several years, I haven't
把它推进下去。所以是的，我觉得在几年到十几年的尺度上，我没有
115:41
sort of kept up on exactly what they're saying. So you should ask Hartwood exactly when that's
有点跟上他们到底在说什么。所以你应该问 Hartwood 那到底是什么时候
115:45
going to happen. But yeah, they're sort of marching along. So the main, the interesting question is,
会发生。但是啊，它们算是就这么往前推进着。所以最核心、最有趣的问题是，
115:51
will the superconducting computing be the winner or will one of the other sort of alternate
会是 superconducting computing 胜出，还是其他某种替代
115:56
technologies? And that's still TBD. I still think superconducting is a very promising thing
技术？这还都是 TBD。我还是觉得 superconducting 是非常有前景的东西
116:02
because it is, it is very scalable. So yeah, no, I don't think it's 30 years away. I don't think
因为它，它非常 scalable。所以是啊，不，我不觉得它还要 30 年。我不觉得
116:07
it's tomorrow. I don't think there's going to be, I don't think unless, well, even if there's a
它明天就能实现。我不觉得会有，我不觉得，除非，嗯，就算有一个
116:10
sudden hardware breakthrough, these are very finicky things. So it's like someone might have
突然的 hardware breakthrough，这些东西也非常难搞。所以就像有人可能有一个
116:14
a brilliant idea, but it will still be a while before because fundamentally, at least in the Nisk era
绝妙的想法，但还是要等一段时间，因为从根本上说，至少在 Nisk era
116:19
or for a while, these are fundamentally analog computers. And so that, there tend to be very,
或者说在一段时间内，这些东西从根本上就是 analog computers。所以这就，往往会有非常，
116:22
very, very finicky. So I wouldn't expect all of a sudden, you know, something, some phase change
非常非常难搞。所以我不会指望突然之间，你知道，某个东西、某种 phase change
116:29
happens. No one has figured out how to scale up cubits in a way that interact and just to kind
会发生。没人搞清楚怎么 scale up qubits，让它们能相互作用，并且达到某种
116:35
of arbitrary size or still in the order of 100, 200 cubits, I think, or you mean in terms of,
任意大小，或者仍然在 100、200 个 qubits 这个量级，我觉得，或者你是说在……方面，
116:41
okay, so in terms of, well, yeah, yeah, exactly. So for superconducting cubits,
好吧，所以就……来说，嗯，对，对，没错。所以对于 superconducting qubits，
116:48
yeah, people, I mean, although people are working on it, the ratio of the number of physical
是啊，人们，我是说，虽然人们正在研究这个，但 physical qubits
116:52
cubits, the logical cubits are still relatively large for the error rates that you need. And so maybe
和 logical qubits 的数量比例，对于你需要的 error rates 来说仍然相对很大。所以也许
116:57
they'll be breakthrough there. I don't know. Or there are other things that might have,
他们会在那方面取得突破。我不知道。或者还有别的东西可能会有，
117:02
you don't need that ratio to be so high because a lot of it has to do with the 2D connectivity
你不需要那个 ratio 那么高，因为很多都跟 2D connectivity 有关
117:08
of the chips that you lay out your cubits out of 2D chips. And for things like neutral atoms,
在这些 chips 里，你是用 2D chips 来铺你的 qubits。而对于像 neutral atoms 这样的东西，
117:14
they, in theory, can connect anything to anything else. But of course, in practice, we don't really know,
理论上，它们可以把任何东西连到任何其他东西上。但当然，在实践中，我们并不真的知道，
117:21
we don't actually know what the limitations of neutral at least. I don't know what the limitations
我们其实不知道 neutral atoms 至少有哪些限制。我不知道有哪些限制
117:25
of what neutral at it, maybe super experts know. Superconducting cubits in particular have us
neutral atoms 到底是什么，也许超级专家知道。Superconducting qubits 尤其有
117:30
sort of tension where you want to have nice clean resonators, which, you know, you do get a
某种张力，就是你想要有很好、很干净的 resonators，你知道，你确实能得到一个
117:36
clean resonator by decoupling from the environment. And then you get good interactions by
干净的 resonator，靠的是 decoupling from the environment。然后你通过
117:40
coupling resonators, which involves covering to the environment. So it's always true. But at
coupling resonators 来获得好的 interactions，而这又涉及到 coupling to the environment。所以这总是成立的。但至少
117:44
least there, but there, yes, but there's like the whole environment. Yeah, yeah. And then there's
在那里，但是那里，是的，但那里有整个 environment。对，对。然后还有
117:49
little tiny principles that you want to go through to your neighbors. This is not like a fundamental
那些你想跟邻居们讲清楚的小小小原则。
117:54
uncertainty relationship or something. That's right. This is a technological imitation that
这不像是什么 fundamental uncertainty relationship 之类的东西。
117:58
difficulty or something. Yeah. The hardware team in Google Quantum is very, very skilled. They're
没错。
118:04
very, very skilled. So yeah, they're really good at making these designs and making these things
这算是 technological imitation 的那种难度之类的吧。
118:10
actually work. So I find them impressive. Yeah. It's exciting to see that advance. Yeah. Yeah. Yeah.
嗯。
118:16
Yeah. I guess I'll just have to hold a breath and wait. Just wait. Yeah. I guess I'm just very patient.
Google Quantum 的 hardware 团队真的非常非常厉害。
118:21
So I start working 20 years ahead of that. That's what Dave Bacon who runs the software team in
他们真的非常非常厉害。
118:29
Google Quantum always he teases me like, John, like, oh, no, I can't work in quantum computing.
所以嗯，他们真的很擅长做这些设计，让这些东西真正跑起来。
118:33
This was like 10 years ago, because it's going to be, you're always 20 years ahead of time.
这大概是 10 年前的事了，因为到时候会是——你总是提前 20 年。
118:38
So I'm going to have to wait for 20 years, like, okay, well, you know, 10 years.
所以我得等 20 年，就像，好吧，嗯，你知道，10 年。
118:42
Halfway there. Halfway there. I don't, I don't take that as an absolute. There's a hearty
已经过半了。已经过半了。我不，我不把这当成绝对。有一个 hearty
118:46
rule. Yeah. Yeah. Well, I also started working on fusion 10 years ago. We'll see.
规则。对。对。嗯，我 10 年前也开始搞 fusion 了。到时候看吧。
118:53
Maybe you're actually causal. Like you start working at something at the reality just, you know,
也许你其实是 causal。就像你开始做某件事，然后现实就，你知道，
118:57
catches up, catches up. I guess so. Maybe who knows? I don't know. I started working. I love
赶上来了，赶上来了。我猜是吧。也许谁知道呢？我不知道。我开始搞了。我热爱
119:01
convolutional nets in the early 90s. And, uh, Yon, uh, Likun claimed. I coined the term
90 年代初的 convolutional nets。而且，呃，Yon，呃，Likun 声称。我创造了这个词
119:08
convolutional net, uh, as far as I can tell, that, that may be true. I, because everyone called it
convolutional net，呃，据我所知，那，那可能是真的。我，因为大家都叫它
119:12
Lynette because it was a very specific thing and they all worked for you on. I said, well, I don't
Lynette，因为它是一个非常具体的东西，而且他们都在为你做这个。
119:15
work for you on. I don't want to call it Lynette. So I called it a convolutional net. So I don't
我说，嗯，我不是为你做这个。
119:19
know. There's a more generic term. It's a good term. Yeah. Before you go, is there anything you want
我不想把它叫 Lynette。所以我把它叫 convolutional net。
119:24
the audience to take away any messages you want? Oh, I think Eris example of it, but I,
所以我不知道。还有一个更通用的术语。这是个好术语。对。
119:30
I'm really amazingly excited about the potential of AI for science. I think it's, I think it's
在你走之前，你有没有什么想让听众带走的，或者任何你想传达的信息？
119:35
going to be an amazing power tool for scientists. And I think scientists won't be replaced. I think
哦，我觉得 Eris 就是个例子，但我，我真的特别兴奋于 AI 在科学上的潜力。
119:41
it's, in fact, I'm hoping that they'll spend all the time on, again, the creative stuff,
我觉得这，我觉得它会成为科学家的一个了不起的强大工具。
119:46
on the rigorous stuff, on the philosophy stuff. And so I think it's going to, I think it's going
而且我觉得科学家不会被取代。
119:50
to be way cool. I hope so too. Yeah. My intuition is that a lot of, a lot of, I've heard a lot of
那会很酷。我也希望如此。是啊。我的直觉是，很多，很多，我听过很多人
119:57
people say that that I hope they're right. And I hope it's not just sort of coping with
说，我希望他们是对的。而且我希望这不只是在应付
120:03
reality that's uncomfortable. The other question we almost forgot to ask, if you could remove a
让人不舒服的现实。还有一个我们差点忘了问的问题，如果你能移除一个
120:10
bottleneck in your industry, which, whatever you, however you wanted to find that by fiat,
你行业里的 bottleneck，随便你，不管你想怎么靠一纸命令把它找出来，
120:18
then like magic, by magic, what would that be? If I could get a magic wish, I would say,
然后像变魔术一样，靠魔法，那会是什么？如果我能得到一个魔法愿望，我会说，
120:24
someone please make the everything lab that you could like send JSON blog to and it will do any
拜托谁做个 everything lab，你可以往里发 JSON blog，它就会做任何
120:30
experiment at all. Automated. But it have to be anything. So essentially, so I guess we have to solve
实验。自动化。但它得什么都能做。所以本质上，所以我想我们得解决
120:35
the sort of AI complete robotics problem, I guess. But if we did, then that would be stunning. Because
那种 AI complete 的 robotics 问题，我猜。但如果我们做到了，那会非常惊人。因为
120:40
right now, things like era, it's all computational. So someone has to gather the data. So yeah,
现在，像 era 这样的东西，全都是 computational。所以得有人去收集 data。所以，是的，
120:46
if we could just break that, oh, oh, that would be so amazing. That would be so utterly amazing.
如果我们能打破这一点，噢，噢，那会太棒了。那会简直太棒了。
120:51
Yeah. Great. So I really appreciate you taking the time to see us. And I think you,
是的。太好了。我真的很感谢你抽时间来看我们。而且我觉得你，
120:59
you kind of flew in and adjusted your schedule a little bit. Yeah. I was sort of flying over
你算是飞过来的，还稍微调整了一下日程。是的。我当时正飞过
121:04
San Francisco to get home. And so I guess that I landed in San Francisco. So you made a big effort
旧金山回家。所以我想我是在旧金山降落的。所以你费了很大劲
121:09
to be here. We really appreciate that. It was really fun to talk to you. Yeah. Well, since it's
才来到这里。我们真的很感激。跟你聊天真的很有趣。是的。嗯，既然
121:14
been a blast here. Okay. Cool. Thank you for having me. Yeah. You're welcome. Thank you.
在这儿玩得这么开心。好的。酷。谢谢你邀请我。是的。不客气。谢谢。