Latent Space
Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
00:00
00:03
Okay, we're here at Periodic.
好，我们现在在 Periodic。
00:06
We're here today with Liam and Doge from Periodic.
今天和我们在一起的是来自 Periodic 的 Liam 和 Doge。
00:08
Welcome and thanks for having us at yours.
欢迎，也谢谢你们让我们来你们这儿。
00:11
Yeah, great to be here in.
是啊，很高兴能来这里。
00:12
Yeah, thanks for coming in.
是啊，谢谢你们过来。
00:14
I want to start off with one of these quotes
我想先聊聊其中一句引言，
00:16
that I love from your website.
就是你们网站上我很喜欢的那句。
00:18
It says, intelligence is necessary but not sufficient.
它说，智能是必要条件，但还不够充分。
00:21
New knowledge is created when ideas are found
新知识是在想法被发现
00:23
to be consistent with reality.
与现实一致时产生的。
00:25
It seems so straightforward, but why is it non-obvious?
这看起来很简单，但为什么它反而并不显而易见？
00:29
I think that's sort of the thesis and premise
我觉得这算是那个核心论点和前提
00:31
behind Periodic that Doge and I thought
在 Periodic 背后，Doge 和我当时想的
00:33
when we were building this, which is,
就是我们在做这个的时候，也就是：
00:36
you can't just think your way to a solution.
你不能光靠想就想出解决方案。
00:39
The universe is so complicated that in order
宇宙如此复杂，以至于为了
00:42
to actually push the frontier of knowledge
为了真正推动知识的边界
00:44
and to make progress, you need to create these conjectures
而且想取得进展，你就得提出这些猜想
00:48
and then actually see whether or not it holds.
然后真的去看看它到底成不成立。
00:51
No amount of rereading that textbook or paper
不管你把那本教科书或那篇论文重读多少遍
00:53
and just thinking and like disappearing into room
还是只是在那儿想，像把自己关进房间里消失了一样
00:55
is going to allow you to think through all possible
都没法让你把所有可能的
00:59
experimental outcomes, expected and unexpected
实验结果，不管是预期之内的还是意料之外的，全都想清楚
01:03
and you really need this like iterative process
你真的需要这种像是 iterative process 一样的东西
01:05
and that's our core belief for putting this together
这就是我们把这些整合到一起的核心信念
01:08
and that's why from the very beginning
这也是为什么从一开始
01:10
where we need to have the AI systems,
我们就需要有 AI systems，
01:12
the simulations of the physical world
对物理世界的 simulations
01:14
but then also to build up the physical high throughput experiment
但也要搭建起物理层面的 high throughput experiment
01:18
and we think a different type of intelligence emerges from that.
而且我们认为，从中会涌现出一种不同类型的 intelligence。
01:21
I think it's also notable like with OpenAI and Google,
我觉得还有一点也很值得注意，就像 OpenAI 和 Google 一样，
01:25
you don't have that resources in those big labs.
那些大实验室里并没有那样的资源。
01:27
You have to come out and do your own thing
你得站出来，做你自己的事
01:30
because you are kind of inventing your own playbook
因为你其实是在发明自己的一套打法
01:32
as you go along.
边做边摸索。
01:34
Yeah, I mean, a team like this has never existed, I think.
对，我是说，我觉得像这样的团队以前从来没存在过。
01:36
We try to bring together salt state chemists,
我们试着把 salt state 化学家、
01:39
salt state physicists, experimentalists,
salt state 物理学家、实验人员、
01:41
theorists, hardware engineers,
理论人员、硬件工程师、
01:43
LLM experts, computer scientists
LLM 专家、计算机科学家聚到一起
01:45
and some of these technologies are very recent
而且这些技术里有一些是最近才出现的
01:47
like the high throughput experimentation,
比如 high throughput experimentation，
01:50
the robotic arms have only been tried in the last three years
robotic arms 也就是过去三年
01:53
or so, the force field expertise,
左右才有人尝试，还有 force field expertise，
01:55
one of these force fields are also very recent.
这些 force fields 里有一个也是最近才有的。
01:58
But yeah, we felt like having a lab is very important,
不过呢，我们觉得有一个 lab 非常重要，
02:00
having this focus and bringing all these people together
有这种 focus，并且把所有这些人都聚到一起
02:02
to work together is very important.
一起合作，是非常重要的。
02:05
And I guess like the closest examples from history
而且我觉得，历史上最接近的例子
02:08
are place like Bell Labs,
大概就是像 Bell Labs 这样的地方，
02:09
where incredible theorists and experimentalists
那里有非常厉害的理论家和实验家
02:12
and chemists work together
还有化学家一起合作，
02:13
and achieved incredible things.
做出了很了不起的事情。
02:15
So we're trying to do the same here, yeah.
所以我们也想在这里做同样的事，对吧。
02:17
I'd like to talk about force fields in a bit
我待会儿想聊聊 force fields
02:19
but before we get to the, yeah, like what are they?
但在我们聊到那个之前，嗯，就是，它们到底是什么？
02:22
Before we get to that, I think one of the,
在我们聊到那个之前，我觉得其中一个，
02:23
just for context, one of the biggest differences
先交代一下背景，最大的差异之一
02:25
between some of our audience are AI engineers
就是我们听众里，有些是 AI engineers，
02:30
and some of the audience is scientists.
而有些听众是科学家。
02:32
So let's go and talk to the engineers real quick
所以我们就先快速去跟 engineers 聊聊，
02:35
and say like what is the difference?
问问看，区别到底是什么？
02:37
If you've been working at Open AI is a big lab,
如果你一直在 Open AI 这种大 lab 工作，
02:39
what's the difference between what you do there
你在那儿做的事情有什么不一样？
02:41
and what is here?
那这里是什么？
02:43
I think you've already hinted at this,
我觉得你已经暗示过这一点了，
02:44
but maybe explicitly what's the,
但也许明确地说，那个，
02:46
how do you have to reframe your thought process?
你要怎么重新构建你的思考过程？
02:49
Well, I think like one interesting thing is
嗯，我觉得一个有意思的点是
02:52
now our reinforcement learning environments
现在我们的 reinforcement learning environments
02:55
literally derive from our physical labs.
真的就是源自我们的实体实验室。
02:59
Our data comes from our physical labs
我们的数据来自我们的实体实验室
03:01
and this is sort of our ultimate truth.
而这可以说是我们的终极真理。
03:04
It's not enough just to do optimization
只是做 optimization 是不够的，
03:07
and some answers that were known in some papers
还有一些在某些论文里早就知道的答案，
03:10
or textbooks because we're going beyond that.
或者教科书里的答案，因为我们要超越这些。
03:12
And I think that's one of the biggest differences.
而且我觉得这是最大的区别之一。
03:15
But along with that, there's issues of, say,
但与此同时，还有比如说
03:18
decision making under uncertainty.
decision making under uncertainty 的问题。
03:19
So when you're doing optimization against math,
所以当你在针对数学做 optimization 的时候，
03:22
there's a high precision to it.
它本身有很高的 precision。
03:23
You're not really dealing with variance
你其实并不是在处理 variance
03:26
or uncertainties or you know, aberrant measurements,
或者 uncertainties，或者你懂的，aberrant measurements，
03:28
whereas that's very key to our process.
而这对我们的流程来说非常关键。
03:31
Like, for example, when we are doing materials discovery loop,
比如说，当我们在做 materials discovery loop 的时候，
03:36
things don't come off the out of the furnace labeled, right?
东西从炉子里出来时并不会带着标签，对吧？
03:40
Even the labeling process can be stochastic and noisy.
就连 labeling process 本身都可能是 stochastic 和 noisy 的。
03:43
And sometimes you'll have aberrations
而且有时候你会遇到 aberrations。
03:45
between different machines.
在不同机器之间。
03:47
Maybe we'll have insufficient telemetry.
也许我们的 telemetry 会不足。
03:49
So we thought we ran it at this temperature,
所以，我们以为是在这个 temperature 下跑的，
03:53
but really the temperature was, you know, some delta away.
但实际上，那个 temperature，你知道，差了某个 delta。
03:56
And an intelligence that can take in this noisy data
而一种能够接收这些 noisy data 的智能，
04:00
and make intelligent decisions like a scientist would
并且能像科学家那样做出明智决策，
04:04
is sort of a different set of reasoning strategies.
这算是另一套不同的 reasoning 策略。
04:07
So I think there's a huge amount of commonalities.
所以我觉得这里面有非常多共通之处。
04:09
So the standard mid training, reinforcement learning,
所以标准的 mid training、reinforcement learning，
04:13
the construction of tool using agents,
tool using agents 的构建，
04:17
variance reduction, you know, making sure
variance reduction，你知道的，就是确保
04:19
and for it doesn't have mismatches
它不会出现不匹配，
04:21
between like training and inference.
比如 training 和 inference 之间。
04:22
But we have to go beyond that and really think about,
但我们必须超越这些，真正去思考，
04:26
how do you do accurate work when there's
当存在
04:29
a high amount of uncertainty?
高度不确定性时，你怎么才能把工作做准确？
04:31
How do you make really incredibly efficient use
你怎么才能真的极其高效地利用一组有限的 data？
04:34
of a limited set of data?
所以我们正在大幅地 scaling things up，
04:36
So we're scaling things up significantly,
但它仍然和一个 digital environment 非常不一样，
04:38
but still it's very different than a digital environment
在那里你可以任意增加更多 environments
04:41
where you can arbitrarily add more environments
或者更多 roll outs。
04:44
or more roll outs.
我们并没有同样的 capacity。
04:45
We don't have that same capacity.
所以 sample efficiency 是另一个关键点。
04:46
So sample efficiency is another key thing.
04:48
This seems a little bit abstract to me,
这对我来说有点抽象，
04:50
and it is a theme which has been on the science pods
而且这个话题在科学播客上已经出现一阵子了，
04:53
for a bit, but I think it might be helpful
不过我觉得这可能会有帮助，
04:55
when you talk about experimental uncertainty
当你谈到 experimental uncertainty 的时候，
04:58
to explain like what would a specific process
去解释一下，比如你当时在实验里做的一个具体流程会是什么样？
05:01
that you were doing experimentally look like?
还有，比如说，如果是人类，
05:03
And like what would a human,
人类会怎么处理这个问题？
05:05
how would a human approach this problem
05:06
before we jump into how do you go into the AI side?
在我们开始聊你是怎么进入 AI 领域之前，
05:09
Do you have like a specific example
你有没有那种具体的例子，
05:11
of a type of material you make
比如你做的一种材料，
05:12
and like what sort of uncertainty in the measurements
还有比如测量中会有什么样的 uncertainty，
05:15
you would experience in the process?
是你在过程中会遇到的？
05:17
I think there are many dimensions to this.
我觉得这个有很多维度。
05:18
So one of them is I think every scientific experiment
所以其中一个维度是，我觉得每个科学实验
05:21
has to do some dimensional reduction,
都得做一些 dimensional reduction，
05:23
and this is very different than coding or math.
而这跟 coding 或 math 非常不一样。
05:25
So when you're doing math or coding,
所以当你在做 math 或 coding 的时候，
05:28
all the context could be available to the human or the LLM.
所有的 context 都可以提供给人类或者 LLM。
05:31
It's a bunch of axioms, some corollaries,
它就是一堆 axioms、一些 corollaries，
05:33
people that drive from them.
还有人们从它们推导出来的东西。
05:35
It's a bunch of like functions and API calls.
它就是一堆像 functions 和 API calls 这样的东西。
05:38
So everything you need to reason through is available to you,
所以你需要推理的所有东西都摆在你面前，
05:41
and then you just have to be very smart
然后你就只需要非常聪明就行了。
05:43
and like figured out.
然后，就好像搞明白了。
05:44
In physics, as you know, we start with more atoms
在物理里，你也知道，我们一开始有的 atoms 更多
05:47
than we could ever store on a computer.
比我们能在电脑上存下的还要多。
05:49
So clearly we'll have to go
所以很明显，我们得从
05:51
from the original number of dimensions
原始的 dimensions 数量
05:54
and bits that represent the system
以及代表这个 system 的 bits
05:56
to the number of bits we can fit in the computer.
到我们能塞进电脑里的 bits 数量。
05:58
And this was I guess the original premise
而这，我猜，就是最初的前提。
06:01
of thermodynamics.
thermodynamics 的。
06:02
They were originally really confused
他们一开始真的特别困惑
06:03
about how steam engines worked,
搞不懂蒸汽机到底是怎么运作的，
06:05
but then turns out you can find like 5, 4, 6 thermodynamic variables
但后来发现，你好像能找到 5 个、4 个、6 个 thermodynamic variables
06:09
that explain what's going on pretty well,
就能把情况解释得相当清楚，
06:11
which is incredible.
这太不可思议了。
06:12
That's the beauty of physics.
这就是物理的美。
06:13
Like you have a room full of atoms,
就像你有一个房间，里面全是 atoms，
06:15
and there are 10 to the 23 or 10 to the 27th.
06:19
There's something atoms in this room,
06:21
and yet you need temperature and pressure
06:24
and simple variables, and that tells you everything.
06:27
They'll tell you mostly what you need to know
06:29
for most cases, right?
06:30
So let's say you make a material
06:32
and you could do something like look
06:34
at an x-ray structure.
在一个 x-ray structure 上。
06:35
And this thing is a very lossy projection, right?
而且这个东西是一个非常 lossy projection，对吧？
06:39
It's like not actually a structure,
它其实并不真的算一个结构，
06:42
or it doesn't tell you the true structure.
或者说它并没有告诉你真正的结构。
06:43
It tells you some principal components of that.
它告诉你的是其中一些 principal components。
06:47
So how would a human take these simple principal components,
所以人类会怎么拿这些简单的 principal components，
06:51
and then kind of reason about them,
然后某种程度上对它们进行推理，
06:54
and then how do you extend that to an AI?
然后你又怎么把这个扩展到 AI 上？
06:58
So on the theoretical physics side, right?
所以在 theoretical physics 这边，对吧？
07:00
We've decided that energy, the average value of energy,
我们已经确定 energy、energy 的平均值，
07:03
and the fluctuations are energy is very important.
以及 energy 的 fluctuations 都非常重要。
07:06
So all our thermodynamics have been right from this,
所以我们所有的 thermodynamics 都是从这里来的，
07:08
and then we realize reaction barriers are also very important.
然后我们意识到 reaction barriers 也非常重要。
07:11
The kinetics is very important.
kinetics 非常重要。
07:14
So a human kind of tries to reason through this
所以人类会有点像是在试着推理这个
07:16
extremely complex system by coming up
极其复杂的系统，通过想出
07:18
with some reduced-dimensional descriptors.
用一些 reduced-dimensional descriptors。
07:22
People also really like atomistic picture,
人们也真的很喜欢 atomistic picture，
07:24
like do you think about what the local neighborhoods look like,
比如你会不会去想 local neighborhoods 长什么样，
07:27
whether they from Okta Hidra, Tetra Hidra,
它们是不是来自 Okta Hidra、Tetra Hidra，
07:30
and that especially chemists really benefit from that?
而且尤其是化学家真的能从中受益？
07:33
Okay, so that's on the theoretical side.
好，所以那是理论这边。
07:35
And on the experimental side,
而在实验这边，
07:36
you would try to collect as much data as possible.
你会尽量收集尽可能多的数据。
07:38
So you'd have a lab notebook,
所以你会有一本 lab notebook，
07:40
and you'd write everything you see, et cetera.
然后把你看到的一切都写下来，等等。
07:42
But with the understanding, of course, that you'll miss a lot.
但当然，前提是你明白，你会漏掉很多。
07:45
So as Liam mentioned, there are all these issues
所以就像 Liam 提到的，会出现各种各样的问题，
07:47
that come up, for example, your furnace will degrade over time,
比如你的 furnace 会随着时间退化，
07:51
because as you're using the furnace,
因为当你使用 furnace 的时候，
07:53
some of the things you're baking will evaporate
你烘烤的一些东西会蒸发，
07:56
and cover the heating element.
然后覆盖住 heating element。
07:58
So every time you use a furnace, it's worse than worse.
所以每次你用 furnace，都糟上加糟。
08:01
Another issue is that you have a furnace,
另一个问题是，你有一个 furnace，
08:03
but the temperature isn't perfect uniform,
但 temperature 并不是完全均匀的，
08:05
so where you put in the furnace will affect the result.
所以你把东西放在 furnace 里的位置会影响结果。
08:07
We don't have this issue yet,
我们目前还没有这个问题，
08:09
so we will maybe someday,
所以也许有一天我们也会碰到，
08:10
the optical instruments suffer a lot from vibrations.
optical instruments 受 vibrations 的影响很大。
08:13
So like every time somebody walks,
所以就像每次有人走动，
08:15
that I remember when I was doing my PhD,
我记得我读 PhD 的时候，
08:17
one lab had this big issue,
有一个实验室遇到了一个大问题，
08:18
because sometimes the results would be very different
因为有时候结果会和其他时候差很多，
08:21
than other times, and they brought it down to,
他们最后把原因归结到，
08:23
at a certain time in the night,
在夜里的某个时间，
08:24
upstairs somebody would be walking.
楼上有人在走路。
08:26
And that vibration affected the laser setup.
那种 vibration 影响了 laser setup。
08:30
So yeah, science is very hard,
所以啊，搞科研真的很难，
08:31
but this one makes it special, right?
但这一条让它很特别，对吧？
08:33
The other thing Liam and I have been talking about is,
我和 Liam 一直在聊的另一件事是，
08:35
a lot of the current improvements focus on math
现在很多改进都集中在 math
08:37
and coding and theoretical computer science
以及 coding 和 theoretical computer science 上
08:39
because it's easier for all of us,
因为这对我们所有人来说都更容易，
08:41
but in real life, most things that require intelligence
但在现实生活中，大多数需要智能的事情
08:43
are actually more like science.
其实更像科学。
08:45
Like there's all the uncertainty, there's all the noise,
就像充满了不确定性，也充满了噪声，
08:48
there's all the missing context,
有那么多缺失的 context，
08:50
but you have to be the intelligent
但你得是那个有智能的
08:51
being able to figure out what to do next.
存在，能够弄清楚下一步该做什么。
08:53
I think like a really interesting point to build on that too
我觉得，同样可以顺着这一点展开的一个很有意思的点是，
08:55
is the optimal reasoning strategies to do math,
做数学的最优 reasoning 策略，
08:58
to do theoretical computer science,
做 theoretical computer science 的，
09:00
to do these types of things,
做这类事情的，
09:01
may not be the optimal reasoning strategies to do science.
可能并不是做科学的最优 reasoning 策略。
09:05
How do you define a RL reward function
你怎么定义 RL reward function
09:08
when your output is a noisy crystal structure
当你的输出是一个有噪声的 crystal structure
09:11
or some number of, I don't know,
或者一定数量的，我也不知道，
09:15
like collective variables or something?
比如 collective variables 之类的？
09:17
Maybe I'll give one part of the loop
也许我会先讲 loop 里的一个部分
09:20
and I'll just kind of dig into that.
然后我就稍微深入讲讲这个。
09:22
So as part of the materials discovery loop,
所以作为 materials discovery loop 的一部分，
09:25
we have to first figure out what to make.
我们首先得弄清楚要做什么。
09:27
So that's the stability of the atoms together
所以这就是原子放在一起时的稳定性
09:30
as well as do we expect it to have the properties we want.
以及我们是否期望它具有我们想要的性质。
09:34
So we don't just want a novel configuration of the atoms,
所以我们不只是想要一种新颖的原子排列方式，
09:37
we want them when put together to have the property of interest.
我们想要它们在组合在一起时，具备我们感兴趣的那种性质。
09:40
Next is how do you synthesize it?
接下来就是，你怎么把它合成出来？
09:42
How do you actually make the thing?
你到底怎么把这个东西做出来？
09:44
That's also non-trivial.
这同样也不简单。
09:45
So even if you know something will hold together
所以即使你知道某个东西能稳定地结合在一起
09:47
the process of actually, you know,
其实这个过程，你知道，
09:49
what are the processing conditions
processing conditions 是什么
09:51
to actually bring that forth is highly non-trivial.
要真正把它做出来，是非常 non-trivial 的。
09:53
But again, like the thing, once you've done those two steps
但再说一遍，就像那件事，一旦你完成了那两步
09:56
and you've made something, it doesn't come off labeled.
并且做出了某个东西，它出来时并不会带着标签。
09:58
So you actually have to characterize it
所以你得实际去 characterize 它
09:59
and figure out what you made.
弄清楚你做出来的是什么。
10:01
So in that case, rather than thinking about the oral environment
所以在这种情况下，与其去考虑 oral environment
10:05
of we're just going to kick off and experiments
呃，我们就直接开始，做点实验吧
10:09
way to a couple of days and try to do like an update
等个几天，然后试着做个 update 之类的
10:11
on like, did you find a room temperature
比如说，你有没有找到室温
10:13
superconductor or not?
superconductor，还是没找到？
10:14
That's completely invisible.
那完全看不见。
10:15
Yeah, I was gonna ask that.
对，我正想问这个。
10:16
Yeah, like, you can't wave your cheeks.
是啊，就，你没法挥动你的脸颊。
10:18
You're sitting for a week or something.
你坐了一个星期还是怎么着。
10:19
Yeah, it's completely infeasible
是啊，这完全不可行
10:20
because you don't have enough agents
因为你没有足够的 agents
10:23
to reduce the variance efficiently.
来高效地降低 variance。
10:25
The rollout time is going between the different instruments.
rollout 时间都花在不同仪器之间来回切换上了。
10:29
You're waiting, there's just some physics limits
你就在那儿等，而且实验本身还有一些物理限制
10:31
of the experiments for like the actual synthesis
比如实际的 synthesis
10:35
or furnace time.
或者 furnace time。
10:36
So it's just too slow, too noisy.
所以就太慢了，也太 noisy 了。
10:39
So the way we think about the AI program
所以，我们看待这个 AI program 的方式
10:42
is basically constructing agents
基本上就是构建 agents
10:44
around this set of data that we've produced.
围绕我们产出的这组数据。
10:47
And so in the case of characterization,
所以在 characterization 的情况下，
10:49
you're looking for our environments
你要找的是我们的 environments
10:51
that are rewarding the identification
会奖励识别出
10:53
of the phases actually present.
实际存在的 phases。
10:55
Are you able to give in the raw experimental data?
你能把 raw experimental data 给进来吗？
11:00
Fit that effectively.
要把那个有效地 fit 好。
11:01
So the way we do this is we'll shoot X-rays at the material.
所以我们的做法是，我们会朝材料打 X-rays。
11:06
X-rays have a frequency or a wavelength roughly
X-rays 的 frequency 或 wavelength，大概
11:09
measured or commensurate to the spacing between atoms.
和原子之间的间距相当，或者说跟它匹配。
11:12
So you get nice diffraction patterns
这样你就能得到很漂亮的 diffraction patterns
11:14
that kind of give a fingerprint of the crystal structure.
它们有点像 crystal structure 的指纹。
11:16
And doing this identification
而做这种 identification，
11:19
of what actually is present is highly non-trivial.
也就是判断实际到底存在什么，是非常不简单的。
11:22
And so some of our early AI work has been
所以，我们早期的一些 AI 工作一直是
11:25
on basically producing systems to do this.
基本上就是在打造系统来做这件事。
11:27
And this allows us to move much more quickly, too,
而且这也让我们能推进得快得多，
11:31
because if you're running so many experiments,
因为如果你跑这么多实验，
11:34
you really quickly become bottlenecked on your ability
你很快就会在自己的能力上遇到瓶颈，
11:36
to understand what was produced.
去理解产出了什么。
11:38
But again, now, the reinforcement learning environment
但话说回来，现在这个 reinforcement learning environment
11:41
is fairly straightforward, which is you're
其实挺直接的，也就是你
11:43
going to reward the identification of the phases present.
会奖励对存在的 phases 的识别。
11:46
You're going to penalize spurious phases or things
你会惩罚 spurious phases，或者那些
11:49
that don't have any like real chemical plausibility.
完全没有什么真正的 chemical plausibility 的东西。
11:53
So that's like a little, it's the same thing,
所以这有点像，稍微有点，其实是一回事，
11:55
where you can kind of do this for the different pieces.
你可以对不同部分做类似的事情。
11:58
And when you stitch this together,
当你把这些拼接在一起时，
11:59
that's the end-to-end discovery loop.
那就是 end-to-end discovery loop。
12:02
Maybe like another piece would be,
也许另一个部分会是，
12:05
as we build up a large basket of experimental data,
随着我们积累起一大堆实验数据，
12:09
you can think of another thing of basically time stamping
你可以把它想成另一件事，基本上就是对
12:13
the state of the world.
世界的状态做 time stamping。
12:14
So you could say, at this date, this
所以你可以说，在这个日期，这
12:16
was our experimental evidence up until that point.
就是我们截至那一刻的实验证据。
12:19
And you can construct reinforcement learning environments
而且你可以构建 reinforcement learning 环境，
12:21
where you say, OK, given that body of experimental evidence
在里面你说，OK，给定那批实验证据
12:25
and given the choices, what was the next choice
以及那些选择，下一个选择是什么
12:30
that the scientist made, or what was an outcome of that experiment.
科学家做出来的那个，或者那个实验产生的结果。
12:34
And this is really interesting, too,
而且这也真的很有意思，
12:36
because this allows us to do types of programs that
因为这让我们能做一些类型的 programs，这些
12:39
would be harder externally, because if the pre-trained model
在外部会更难，因为如果这个 pre-trained model
12:43
has already memorized some of this data,
已经记住了其中一些 data，
12:46
and then you try to do reinforcement learning on that,
然后你试着对它做 reinforcement learning，
12:49
if it already knows the answer and you construct reinforcement
如果它已经知道答案，而你又构造 reinforcement
12:52
learning tasks that depend on that answer, it can fake work.
learning tasks，这些任务依赖那个答案，它就能假装在工作。
12:56
And I mean, because it's like, oh, well,
而且我的意思是，因为这就像是，哦，好吧，
12:59
it already knows the answer.
它已经知道答案了。
13:00
So it doesn't actually have to do the hard physics reasoning.
所以它其实并不需要去做那些困难的 physics reasoning。
13:02
It doesn't have to actually do the calculation simulations.
它其实并不需要真的去做 calculation simulations。
13:05
And it gets the right answer.
然后它就能得到正确答案。
13:07
And then you reinforce that policy
然后你就 reinforce 那个 policy
13:09
and you upweight those reasoning strategies.
然后你 upweight 那些 reasoning strategies。
13:11
Those reasoning strategies will not generalize to novel systems.
那些 reasoning strategies 不会 generalize 到 novel systems 上。
13:14
So it's not useful to us.
所以这对我们没什么用。
13:16
But as we build up baskets of experimental data
但随着我们积累起一批批实验数据，
13:19
and like stitching things together,
把各种东西拼接在一起，
13:20
having the lineage to the whole process,
并且保留整个过程的脉络，
13:22
that gives us the ability to construct
这让我们有能力去构建
13:24
new types of environments that are just not plausible elsewhere.
那些在其他地方根本不可能实现的
13:27
I want to ask my question, but I'm worried
新型环境。
13:29
that I'm taking it off the rails.
我想问我的问题，但我担心
13:30
I'm just going to go for it, because again,
那我就直接说了，因为还是那句话，
13:33
I come from the non-scientists, but engineering sort of.
我算是非科学家出身，但偏 engineering 一点。
13:35
I'm familiar with the ML side, but not the physics side.
ML 那边我熟，但 physics 那边我不熟。
13:38
Okay, there's a few versions of this.
好吧，这个有几个版本。
13:41
But I think the basic question is, don't we
但我觉得最基本的问题是，我们难道不是
13:45
have most of the laws of physics worked out?
已经把大部分 physics 定律都搞明白了吗？
13:48
And how come we don't have perfect simulators already?
那为什么我们还没有完美的 simulators 呢？
13:51
Like, this is very basic.
就，这很基础啊。
13:54
And he was like, that's cute.
然后他就说，那挺可爱的。
13:56
But I have all these physics books,
但我有这么多 physics 书，
13:59
so what do you mean?
所以你是什么意思？
14:00
What do you mean we're not done?
你说我们还没完，是什么意思？
14:03
Yeah, so there are a couple of things going on here, right?
对，所以这里有好几件事在同时发生，对吧？
14:07
So I think, first of all, we're definitely not done
所以我觉得，首先，我们绝对还没搞定
14:09
with any of the laws of physics.
任何 physics 定律。
14:11
Even the quantum scale, okay,
甚至 quantum scale，好吧，
14:14
but maybe very, very large scale,
但也许是非常非常大的 scale，
14:16
but like, at human scale, at like, you know, material scale,
但就像，在 human scale 上，在，你知道，material scale 上，
14:19
we're like, what's left?
我们就会想，还剩下什么？
14:21
We're not, so it's a really good question.
我们还没，所以这真的是个好问题。
14:24
And it's very interesting why we're not done yet, right?
而且为什么我们还没做完，这非常有意思，对吧？
14:28
So there's a very common story people talk about,
所以人们经常讲一个很常见的故事，
14:32
like when Dirac was figuring out quantum mechanics,
就像 Dirac 在搞清楚 quantum mechanics 的时候，
14:35
like late 1920s, he wrote a textbook about quantum mechanics,
大概在 1920s 末期，他写了一本关于 quantum mechanics 的教科书，
14:39
and he kind of framed it as a sold-down.
14:41
And then there's a phrase people like to quote,
14:43
I'm not sure actually how accurate it is,
14:44
but apparently Dirac said, the rest is chemistry.
14:48
The idea being like he could solve the hydrogen atom.
14:50
Just composed everything.
14:51
Yeah, he could maybe solve like a chain of 1D hydrogen atoms,
14:54
but then he cannot currently solve, say,
14:57
nitrogen and tracking with oxygen,
nitrogen 和 oxygen 的 tracking，
14:58
but that's okay, that's just chemistry.
但没关系，那只是 chemistry。
15:00
Physics really loves single, like atom,
Physics 真的很喜欢单个的，比如 atom，
15:04
really simple systems.
非常简单的 systems。
15:05
Simpsons, all you can solve, yeah.
Simpsons，你都能解，对。
15:07
Very cool cow, yeah, that's very cool cow.
很酷的牛，对，那是头很酷的牛。
15:08
But turns out, I think what we learned in the last 100 years
但事实证明，我觉得过去 100 年我们学到的
15:11
is, first of all, that's not true.
是，首先，那不是真的。
15:13
It's not like a simple extension of what it was
这不像它原本用途的简单延伸
15:15
to do with this hydrogen.
来处理这个 hydrogen。
15:17
And it's not just, the rest is not just chemistry.
而且不只是这样，剩下的也不只是 chemistry。
15:19
There's actually a lot of physics there.
其实那里有很多 physics。
15:21
I mean, I guess one of the things we still haven't figured out
我是说，我猜我们还没搞明白的一件事
15:22
is high temperatures per conductivity.
就是高温下的 conductivity。
15:24
But there are a lot of other things.
但还有很多其他事情。
15:25
And high for you is like 175 or 200.
而对你来说，高大概就是 175 或 200。
15:28
When you say high temperatures per conductivity,
当你说 high-temperature superconductivity 的时候，
15:30
what they mean there is unconventional superconductivity.
那里指的其实是 unconventional superconductivity。
15:32
So there is the conventional superconductivity
所以，有一种 conventional superconductivity，
15:34
that's just mainly driven by electron-4-9 coupling.
它主要就是由 electron-phonon coupling 驱动的。
15:37
And in that case, you can kind of see the isotope effect,
而在那种情况下，你差不多能看到 isotope effect，
15:39
like if you take the same system,
比如说你拿同一个 system，
15:41
but just different weight for one of the elements,
只是让其中一种元素的 weight 不一样，
15:43
you can see the superconducting temperature drops
你会看到 superconducting temperature 下降
15:46
at the rate you expect.
以你预期的速率。
15:47
So that one is conventional.
所以那个是常规的。
15:49
But then higher temperatures for conductors
但接着，conductors 在更高温度下
15:50
that cooperates like the ones that are about 77,
会像那些大约在 77 的一样协同配合，
15:53
Calvin, like 93, Calvin, turns out don't obey that physics.
Calvin，比如 93 Calvin，结果它们并不遵循那套物理规律。
15:56
But we don't know what physics they obey.
但我们不知道它们遵循的是什么物理规律。
15:58
They're just incredible superconductors.
它们就是不可思议的 superconductors。
16:00
But it's not just that.
但不只是这样。
16:01
So that definitely is one of these very popular topics.
所以这绝对是那些非常热门的话题之一。
16:04
We don't know the theory yet for yet.
我们目前还不知道这个理论。
16:06
But there's so many other things we don't understand.
但还有很多其他我们搞不懂的东西。
16:07
For example, even strong correlation
比如，即便是 strong correlation
16:10
in simple quantum mechanical systems,
在简单的 quantum mechanical systems 里，
16:12
we cannot simulate yet.
我们现在也还无法 simulate。
16:14
Density functional theory is a very powerful tool.
Density functional theory 是一个非常强大的工具。
16:17
It's incredibly accurate on some things.
它在某些方面准得惊人。
16:19
But as soon as there's some strong electron correlation,
但只要出现一些 strong electron correlation，
16:22
it actually fails to capture some of the effects.
它实际上就无法捕捉到某些效应。
16:24
I mean, there's so much to figure out.
我是说，有太多东西需要搞清楚了。
16:26
That's why I'm really excited for AI
这就是为什么我真的很期待 AI
16:27
to be applied to this field.
能被应用到这一领域。
16:28
Because theoretically, computationally and experimentally,
因为无论在理论上、计算上还是实验上，
16:31
there's so much to figure out.
都有太多东西要搞清楚。
16:32
And every improvement we make here
而我们在这里取得的每一点进步
16:35
should improve human life.
应该能改善人类生活。
16:36
Because the better we can understand materials,
因为我们越能理解材料，
16:38
all state physics, the better devices we can make for them.
all state physics，就越能为他们造出更好的设备。
16:42
There's this famous quote by Phil Anderson, which
Phil Anderson 有句名言，就是
16:44
is more is different, which is basically
more is different，基本上是说
16:47
you may understand all the basic laws of something.
你可能理解某个东西的所有基本定律。
16:49
But when you add many things, they behave qualitatively,
但当你把很多东西加在一起时，它们会在性质上表现得
16:53
distinct differently from how the simple physics should
与 simple physics 本该有的方式截然不同。
16:56
tell you.
告诉你。
16:57
So understanding how lots of things behave together,
所以，理解一大堆东西一起表现时是怎么回事，
17:00
it seems like it should be simple.
看起来应该挺简单的。
17:02
Beginning laws are simple, but the collective behavior
最开始的规律很简单，但 collective behavior
17:05
is so complicated that it's really hard to model.
却复杂到真的很难 model。
17:07
It's like emergent and universal, which is crazy.
这就像 emergent 和 universal，简直离谱。
17:11
And we're also things with deep learning models, right?
而且我们用 deep learning models 也能看到这些东西，对吧？
17:13
We see power laws everywhere in deep learning.
我们在 deep learning 里到处都能看到 power laws。
17:15
We don't understand.
我们并不理解。
17:15
But it's very reminiscent of physics,
但这非常让人联想到物理学，
17:17
where there's emergence and universality.
在物理学里存在 emergence 和 universality。
17:20
There's been many theory papers in the physics world
物理学界已经有很多理论论文
17:23
about emergent and neural networks.
是关于 emergent 和 neural networks 的。
17:24
Absolutely, yes.
绝对是，没错。
17:26
And then my other follow up question
然后我另一个追问
17:28
was also just on the process of let's call it the process
也正好是关于那个过程，我们姑且称之为过程
17:33
and the end result.
以及最终结果。
17:36
You want to have these properties
你想要具备这些属性
17:37
that you're targeting for materials,
也就是你为材料所瞄准的那些属性，
17:39
and then you have to figure out the process to get there.
然后你必须弄清楚达到这个目标的过程。
17:41
Is it worth separating these things
把这些东西分开值得吗？
17:43
so that you can have, like you just
这样你就能，就像你直接
17:44
train a model that perfectly reverse engineers
训练一个模型，它能完美 reverse engineer
17:47
any process whatsoever, given whatever theoretical and state
任何流程，无论给定什么样的 theoretical 和 state
17:50
you want?
你想要？
17:51
Is that meaningful?
那有意义吗？
17:53
Maybe going back to the question we read about how
也许回到我们读到过的那个问题，关于
17:56
since there are more than Avogadro's number of atoms,
既然原子的数量比 Avogadro's number 还多，
17:58
we'll never know the exact context of what happened.
我们永远无法知道所发生事情的确切 context。
18:01
I think that might be a reason why we'll never
我觉得那可能就是一个原因，为什么我们永远
18:04
be able to reverse engine everything perfectly.
无法完美地 reverse engineer 一切。
18:06
But we're just trying to reverse engineer sufficiently
但我们只是想 reverse engineer 到足够的程度。
18:08
to be able to improve the performance
能够提升 performance
18:11
of the materials, basically.
也就是材料的 performance，基本上。
18:13
Well, you're asking, is there a forward model
嗯，你是在问，有没有一个 forward model
18:15
where you put in an input and you can predict the output?
你把一个 input 放进去，就能预测 output？
18:17
And you want to create a surrogate model, which
然后你想创建一个 surrogate model，它
18:22
can address it.
可以解决这个问题。
18:22
Like the input is, like the final state
比如说，input 就是，比如 final state
18:24
of that, the configuration of atoms,
那个的，atoms 的 configuration，
18:26
and then the prediction is.
然后 prediction 就是。
18:27
Because without the work, almost.
因为如果没有这些工作，几乎……
18:28
Like you can have one team do the process side,
比如你可以让一个团队负责 process side，
18:31
the other team do the property side,
另一个团队负责 property side，
18:32
and then just race them.
然后就直接让他们比赛。
18:36
Yeah, I think there's a lot of progress we can make
对，我觉得我们可以取得很多进展
18:39
in terms of these, like splitting up
在这些方面，比如把
18:41
these agentic workflows to these different areas.
这些 agentic workflows 拆分到这些不同的领域。
18:44
I think there's a lot of empirical work
我觉得有很多 empirical work
18:46
we can make on synthesis prediction as well.
我们也可以在 synthesis prediction 上做。
18:50
Where it's like taking in that configuration of atoms,
这就像是把原子的构型拿进来，
18:53
what tools can you use, given this experimental evidence,
有了这些实验证据，你能用什么工具，
18:56
given prior literature, prior papers,
有了之前的文献、之前的论文，
18:59
how do you actually get to these things?
你到底要怎么才能真正得到这些东西？
19:00
And you ultimately are testing.
而你最终其实是在做测试。
19:02
Are you able to actually replicate it or not?
你到底能不能真的复现它？
19:04
The reason I say this maybe also is just
我这么说，可能也只是
19:07
thinking about it in terms of a company.
从一家公司的角度来想这件事。
19:09
The way that TSMC is kind of like a fab for semis,
TSMC 有点像做 semis 的 fab，
19:13
but they don't design as semis, you
但他们不设计 semis，你
19:15
could be the TSMC of materials where people come to you
可以成为材料领域的 TSMC，人们会来找你
19:18
with like, well, this year's what I want.
带着类似“嗯，今年我想要的就是这个”的需求。
19:19
And then you can figure out the process.
然后你就能搞清楚 process。
19:21
Yeah, it's like a matter of compiler.
对，这就像 compiler 的问题。
19:24
So it's like giving these requirements.
所以这就像是把这些 requirements 给出来。
19:27
Yeah, yeah, yeah, but it's like giving these requirements
对对对，但这就像是把这些 requirements 给出来
19:29
is this actually like a valid set of configurations?
这实际上算是一组有效的 configurations 吗？
19:32
Can this exist?
这能存在吗？
19:33
Is it like the objective function?
这是不是就像 objective function 一样？
19:35
I don't know, it's too big.
我不知道，它太大了。
19:37
Yeah.
对。
19:38
Why don't we kind of talk about like one of our missions
要不我们稍微聊聊，比如说我们其中的一个任务？
19:40
as synthesis superintelligence.
作为 synthesis superintelligence。
19:42
So I think it's in line with this.
所以我觉得这和这个是一致的。
19:44
I want to double click on something
我想深入聊一下某个点
19:45
that you said a little while back, which
就是你不久前说过的，那个
19:49
was you talked about phases.
就是你谈到了 phases。
19:51
And I think this is going to a doge's comment
而且我觉得这会说到 doge 的评论
19:54
about what is, like emergence.
关于什么是，比如说 emergence。
19:58
Like the concept of a phase transition,
就像 phase transition 这个概念，
19:59
I think might be foreign to a lot of the audience.
我觉得这对很多听众来说可能有点陌生。
20:01
So can you explain what is a phase transition?
那你能解释一下什么是 phase transition 吗？
20:03
And then why is this like a helpful signal
然后，为什么这会是一个有用的 signal，
20:06
for something like an RL environment?
对于像 RL environment 这样的东西来说？
20:08
Phase transitions are fascinating.
Phase transitions 非常迷人。
20:10
I mean, I think the maybe the one that people can most easily
我是说，我觉得人们最容易
20:13
relate to in their life is ice melting probably,
在生活中联想到的那个例子，可能就是冰融化，
20:17
or water boiling.
或者水沸腾。
20:18
Like, you know, you increase the temperature of ice,
就像，你知道的，你把冰的温度升高，
20:21
and it still looks like ice, and it behaves like ice.
它看起来还是冰，表现得也像冰。
20:24
But there's a certain temperature
但到了某个温度，它的温度就不再上升了，然后就变成液体。
20:26
at which it just stops raising its temperature,
所以它突然从 solid phase 变成 liquid phase。
20:29
and then turns into liquid.
当然，另一个跟我们非常相关的，就是 ising model。
20:30
So it suddenly goes from the solid phase to the liquid phase.
20:34
Another one that's very relevant to us, of course,
20:36
is the ising model.
20:38
And I think computer science and mathematicians
而且我觉得 computer science 和数学家们
20:39
also study this maybe under different names.
可能也在研究这个，只是叫法不同。
20:41
But basically, you can have spins up and down,
但基本上，spin 可以朝上，也可以朝下，
20:44
and they have different interaction terms
而且它们有不同的 interaction terms，
20:47
if they're both upwards, ones up, ones down.
如果它们两个都朝上，或者一个朝上、一个朝下。
20:49
And at high enough temperature,
而且在 temperature 足够高的时候，
20:51
they usually just randomly up or down entropy winds.
它们通常只是随机地朝上或朝下，entropy 占上风。
20:53
You lower the temperature.
你把 temperature 降下来。
20:55
It's still looking the same.
它看起来还是老样子。
20:56
But then at some point, suddenly,
但后来在某个时刻，突然，
20:58
they perfectly set the line.
它们完美地定下了那条线。
21:00
Yeah, the phase transition.
对，就是 phase transition。
21:01
Phase transition is really nice,
phase transition 真的很妙，
21:02
because it makes physics life easier
因为它让做物理这件事轻松多了
21:04
for studying certain things.
尤其是在研究某些东西的时候。
21:05
They tend to have certain spatial correlations
它们往往会有某些 spatial correlations
21:09
that really help us as well.
那也真的帮了我们很多。
21:11
OK, so in our lab, phase transitions
OK，所以在我们的实验室里，phase transitions
21:13
come about usually because we mix the precursors
通常会发生，是因为我们会混合 precursors
21:16
that are different crystals, basically.
而它们基本上是不同的 crystals。
21:19
But then we raise the temperature
但之后我们会升高温度
21:21
or do something else to encourage them to react.
或者做点别的来促使它们 react。
21:23
And then atoms start reacting and they
然后 atoms 开始 react，它们
21:25
form into this new crystal.
就形成这个新的 crystal。
21:28
And this shows up in the XID pattern
而这会体现在 XID pattern 里
21:30
because usually if the geometry of the atoms
因为通常来说，如果原子的几何结构
21:32
change drastically, the XID pattern
发生剧烈变化，XID pattern
21:34
changes a lot.
也会变化很大。
21:35
But what really happens in a typical practical materials
但在一个典型的、实际的 materials discovery
21:38
discovery campaign is, when you first
campaign 里，真正发生的是，当你第一次
21:41
try something, it doesn't work.
尝试某个东西时，它并不奏效。
21:43
And the outcome isn't just a clear yes or no,
而且结果并不只是一个明确的“是”或“否”，
21:45
but it's usually a very mixed phase.
但通常它是一个非常 mixed 的 phase。
21:47
It will usually have some of the precursors.
它通常会含有一些 precursors。
21:49
It might have some amorphous phase.
它可能会有一些 amorphous phase。
21:51
And then it will have a bunch of phases
然后它还会有一堆 phases
21:52
that you maybe didn't try to make.
是你可能根本没打算做出来的。
21:54
So that becomes a challenge.
所以这就成了一个挑战。
21:55
And that's why we found it very difficult to bring together
这也是为什么我们发现很难把
21:58
simulations and AI to do the characterization.
simulations 和 AI 结合起来做 characterization。
22:02
So the simulations can always say, oh, this phase doesn't
所以 simulations 总是可以说，哦，这个 phase 看起来不
22:04
look like what we predicted, but this is
像我们预测的那样，但这是
22:06
an exciting variation of it.
它的一个令人兴奋的变体。
22:08
Let me now do force field calculation
现在让我来做 force field calculation，
22:10
and to see if it's still stable.
然后看看它是不是还稳定。
22:11
Or like the AI can say, based on the previous experiment,
或者，AI 可以说，基于之前的实验，
22:14
and this one, this is probably not
以及这一次，这大概不是
22:16
the phase we want to make.
我们想做的那个 phase。
22:17
So let's change the conditions.
那咱们换个条件吧。
22:18
Yeah, for listeners XRD is X-ray diffraction,
对，给听众说明一下，XRD 就是 X-ray diffraction，
22:21
which you already described.
你之前已经描述过了。
22:22
Yeah, in some cases, there's like a phase
对，有些情况下，会有一种 phase，
22:24
that just hasn't been recorded before.
只是之前从来没被记录过。
22:27
It doesn't exist in any papers or database.
它在任何论文或 database 里都不存在。
22:29
And so the AI has to actually go and sort of put
所以 AI 实际上得去，差不多是放
22:31
easy tools to say, well, what configuration atoms
一些简单的工具，用来说，嗯，atoms 的 configuration 是什么
22:35
would actually explain these phases?
实际上能解释这些 phase 吗？
22:37
But this becomes incredibly important
但这一点变得极其重要
22:38
for directing the scientific process.
来引导科研过程。
22:40
Because let's say we have a phase target in mind,
因为假设我们心里有一个 phase target，
22:44
we really need to sort of hill climb that
我们真的需要多多少少去 hill climb 它
22:46
to get better measurements on that.
才能对那个得到更好的测量结果。
22:48
So if it's 1%, and we're like, actually,
所以如果它是 1%，而我们觉得，其实，
22:50
we want to increase this phase purity.
我们想提高这个 phase purity。
22:52
We want to have a very good measurement
我们想有一个非常好的 measurement
22:53
on this experimental data to say, you know,
在这个 experimental data 上，来说，你知道，
22:55
what is president, what's not.
什么算 president，什么不算。
22:57
So one of the advantages is you have just a distinct,
所以其中一个优势就是，你有一个很独立的，
23:00
like, variable that you can just observe.
就像，一个你可以直接观察的 variable。
23:02
Like, going back to the same and about,
就像，回到同一件事，关于，
23:04
how do you deal with uncertainty?
你如何处理 uncertainty？
23:05
One of the, it seems like part of your answer
其中一个，看起来像是你答案的一部分
23:07
is you make your observable signature
就是让你的 observable signature
23:10
somewhat unambiguous.
在某种程度上没有歧义。
23:12
There's no interpretation there.
这里面没有解读的空间。
23:13
Is that right?
是这样吗？
23:14
Is that like the logic behind this?
这就是背后的逻辑吗？
23:15
Or is this like literally you just need
还是说，这其实就是你只需要
23:17
to define a new phase?
定义一个全新的 phase？
23:18
And so this is the thing you care about.
所以这就是你在意的东西。
23:20
I mean, I think there's still ambiguity
我是说，我觉得在很多这类情况下还是有歧义。
23:21
in a lot of these cases.
我是说，就，你知道吗，有个特别傻的地方就是 replicates。
23:22
I mean, like, you know, a very dumb thing is like replicates.
我们会跑 replicates。
23:26
We run replicates.
那当然，对对。
23:27
Of course, yeah, yeah.
但不是，我是说，我觉得它们，你知道，
23:28
But no, I mean, I think they're, you know,
这是个很有挑战性的任务，因为它不是
23:29
it's a challenging task because it's not
完全 deterministic，而且还有像，
23:31
just fully deterministic and there are like,
23:34
well, there are like ambiguities
嗯，会有一些歧义
23:36
and other things.
还有其他一些东西。
23:38
Two different phases can actually be consistent
两个不同的 phases 其实可以是一致的
23:41
with the same pattern.
跟同一个 pattern 相符。
23:42
And so I think that's why it's really important
所以我觉得，这就是为什么真的非常重要
23:44
for the system to use some chemical intuition
让系统用一些 chemical intuition
23:47
to say like, okay, well, this would be highly unlikely
来说，嗯，好吧，这可能性会非常低
23:51
given the synthesis conditions,
考虑到 synthesis conditions，
23:53
given the prior knowledge.
考虑到 prior knowledge。
23:56
And so that can be used to help disambiguate.
所以那可以用来帮助 disambiguate。
23:58
So you're somewhat injecting priors there
所以你在那里某种程度上是在注入 priors
24:01
based on what you expect?
基于你所预期的？
24:03
Yeah, I mean, I think thermodynamics
对，我是说，我觉得 thermodynamics
24:05
is the biggest priors, right?
是最大的 priors，对吧？
24:06
Right.
对。
24:07
And then physics is a big prior.
然后 physics 是一个很大的 prior。
24:09
Another one that we really benefit
另一个我们在这里真正受益的东西
24:10
from here is multimodality or materials characterization.
就是 multimodality 或者 materials characterization。
24:13
So like we can do XRD and yes,
所以就像我们可以做 XRD，而且没错，
24:15
some phases might look similar on the XRD pattern,
有些 phases 在 XRD pattern 上可能看起来很像，
24:18
but then we can also measure the electrical properties,
但接着我们也可以测 electrical properties，
24:21
we can measure their morphology,
我们可以测它们的 morphology，
24:22
like some kind of electron microscopy.
比如某种 electron microscopy。
24:24
And then some of the phases that look similar on XRD
然后有些在 XRD 上看起来相似的 phases
24:26
will look different than some of these dimensions.
会和其中一些 dimensions 看起来不一样。
24:28
And that kind of multimodality really helps.
而那种 multimodality 真的很有帮助。
24:30
And again, AI is really helpful here
而且再说，AI 在这里真的很有帮助
24:32
because like humans have limited context
因为人类本身 context 就有限
24:34
and computation power as well.
computation power 也一样有限。
24:35
So if you give humans like 10 different modalities
所以如果你给人类比如 10 种不同的 modalities
24:37
at the same time and say,
同时，然后说：
24:39
analyze these all consistently, it's a bit difficult.
把这些都一致地分析一下，就有点难。
24:42
But it's actually not super intelligent for an AI to do it.
但其实让 AI 来做这件事，并不显得有多 super intelligent。
24:45
It's great, yeah.
这很棒，对。
24:46
Yeah, it's like the super intelligence
对，这就像是 super intelligence
24:47
is it can just do so much calculations
就是它能做特别多的计算
24:49
that can look at everything.
能看遍所有东西。
24:51
And this kind of gets to that phrase
这就有点说到那个说法了
24:53
about like this is better decision making under uncertainty
大概就像这是更好的 decision making under uncertainty
24:56
where now it's stitching together the signal
现在它是在把 signal 拼接起来
24:59
across many different instruments
跨很多不同的 instruments
25:01
longitudinally across different experiments
longitudinally 跨不同的 experiments
25:03
across the replicates and kind of getting
跨这些 replicates，然后有点慢慢得到
25:05
to this like underlying better model
这种，有点像，更底层、更好的 model
25:08
rather than just over-inducing on one measurement
而不是只在某一个 measurement 上 over-inducing
25:11
from one instrument.
来自某一个 instrument。
25:12
And let me just like, so you know,
然后让我就，怎么说呢，你知道，
25:14
you started I think a while ago
你开始做这个，我觉得已经有一阵子了
25:16
also saying that there's just more data
而且也是在说，data 就是多到
25:17
than you can fit in any reasonable computer
任何一台说得过去的电脑都装不下
25:19
or a process or anything.
或者一个 process，或者随便什么。
25:21
So at some point you have to throw out data
所以到了某个时候，你就得把 data 扔掉
25:23
even though you have all these replicates,
即使你有所有这些 replicates，
25:24
even though you have like,
即使你有，比如说，
25:25
presumably like 10 different ways
大概有 10 种不同的方法
25:27
to measure the same thing.
来测量同一个东西。
25:28
It's because what if one of your temperature thinks is wrong?
因为万一你 temperature 的某个想法是错的呢？
25:31
So like when do you throw that out?
所以，比如你什么时候会把它扔掉？
25:33
How do you decide what to do?
你怎么决定该怎么做？
25:34
As a data-hungry ML guy, I just want everything, right?
作为一个对数据如饥似渴的 ML 人，我就是什么都想要，对吧？
25:36
And then I just, you know,
然后我就，你知道，
25:37
you learn everything.
你就会把所有东西都学一遍。
25:39
I don't know what I don't know
我不知道自己不知道什么
25:40
and just don't know the machine at it.
而且就是搞不懂这件事里的机器。
25:42
Yeah, so we don't throw in data.
对，所以我们不是把 data 丢进去。
25:44
I think maybe what I meant was potentially,
我觉得，可能我想说的是，也许，
25:47
like if God was observing the experiment,
比如，如果上帝在观察这个实验，
25:50
there is more data in the experiment
这个实验里的 data 比
25:51
than humans can fit in a computer.
人类能塞进一台电脑的还多。
25:53
Unfortunately us models cannot observe all the atoms.
不幸的是，我们这些 models 没法观察到所有的原子。
25:57
So even though there's more data than we could store,
所以，即使有比我们能存下的还多的 data，
26:01
we actually can't access that data.
我们其实也根本访问不到那些 data。
26:02
For example, I cannot trace all the atoms positions.
比如说，我没法追踪所有原子的位置。
26:06
So unfortunately, we're still in a regime
所以很不幸，我们还处在这样一个阶段
26:07
where any data we collect is very precious
在这个阶段，我们收集到的任何数据都非常宝贵
26:09
and we don't delete any of it.
而且其中的任何数据我们都不删。
26:11
Yeah, the only way of thinking of it
对，唯一能想它的方式
26:12
to connect to the AI side is like,
把它和 AI 那边联系起来，就是，
26:15
think about probing, like linear probing
想想 probing，比如 linear probing
26:17
of a foundation model, right?
对一个 foundation model 做，对吧？
26:18
This model has all of this data.
26:20
So God has this entire picture
26:23
of what the material is doing.
26:24
And then the experiment is this little tiny linear probe
26:27
which has like eight bits or something about foot.
26:30
And so you're taking this humongous thing
26:33
and course graining it down to a little bit of information
26:35
and then it's now either hit traditionally human's job
26:39
or now your agent's job to reconstruct
或者说现在你的 agent 的任务就是去重建
26:41
from those eight bits of information.
从那八个 bit 的信息里。
26:43
What is actually happening inside.
里面实际在发生什么。
26:45
Like it's an analog where you can't have full visibility
就像它是一个 analog，你没法完全看清
26:48
over every parameter way,
每一个 parameter way，
26:50
over every activation for that input.
每一个针对那个 input 的 activation。
26:52
Yeah, you're getting some sort of subset on it
对，你拿到的是它的某种 subset
26:54
or some like course grained features across the whole thing.
或者一些类似 coarse grained features，覆盖整个东西。
26:58
Yeah.
嗯。
26:59
But it's probably find that you probably do find
但很可能你会发现，你很可能确实会发现
27:02
Maxwell's Demon argument, but interesting
Maxwell's Demon 这个论证，但挺有意思的
27:04
because that kind of-
因为那种——
27:05
You can use to recite it because I don't.
你可以用它来复述，因为我不行。
27:08
He's mentioned it before, I don't know.
他之前提过，我不确定。
27:09
I'm actually forgetting it.
我其实已经忘了。
27:10
So you know, in the 1800s,
所以你知道，在 1800 年代，
27:12
people realized that entropy
人们意识到，entropy
27:14
always system has to increase or, okay,
system 总是必须增加，或者说，好吧，
27:16
for a closed system or stay the same,
对于一个 closed system 来说，要么保持不变，
27:19
but it can't go down as the second law of thermodynamics.
但按照 second law of thermodynamics，它不能下降。
27:23
And there was a thought experiment
然后有一个 thought experiment
27:25
that was asking what would happen
在问会发生什么
27:27
if there was a very tiny, all-knowing demon
如果有一个非常微小、全知的 demon
27:30
that could kind of look at atoms
它有点像是能看看 atoms
27:32
and then whenever an atom had low energy,
然后，每当一个 atom 的 energy 很低时，
27:34
kind of open a door and lead it into this other room
有点像打开一扇门，把它引到另一个房间里
27:37
and keep doing this to reduce entropy of the system
然后一直这么做，来降低 system 的 entropy
27:40
which would break second law of thermodynamics.
可这就会打破 second law of thermodynamics。
27:42
And for a long time,
而且在很长一段时间里，
27:43
I think that wasn't a very satisfying answer
我觉得那并不是一个很令人满意的答案
27:44
for why that was the case.
来解释为什么会是那样。
27:46
But then there was a more recent paper
但后来有一篇更近期的 paper
27:48
in the 1900s, more recently to us.
在 1900 年代，离我们更近一些的时候。
27:51
Landa came up with the argument that
Landa 提出了一个论点，
27:53
you'd have to spend energy to delete information.
你必须消耗 energy 才能 delete information。
27:56
And Maxwell's Demon would basically have access
而 Maxwell's Demon 基本上能够 access
27:58
to so much information by doing this thing he's doing
到这么多 information，就靠他正在做的这件事
28:00
that he'd have to delete some.
以至于他必须 delete 一些。
28:02
And for a minute, deleting information
而且有那么一瞬间，delete information
28:04
he'd have to spend energy
他就必须消耗 energy
28:05
which would then make sure entropy increases.
这样就能确保 entropy 增加。
28:07
So even though we're not at the level
所以，尽管我们还没到
28:09
where we can be Maxwell's Demon,
能成为 Maxwell's Demon 的那种水平，
28:10
I think your question is very visionary
我觉得你的问题非常有远见，
28:13
towards the future where we store so much data,
是在展望未来：我们存了那么多数据，
28:15
we have to delete some.
不得不删掉一些。
28:17
Yeah, no, I wasn't obviously going to that,
嗯，不是，我显然没打算说到那儿，
28:19
it's a net extent.
这是一个净范围。
28:21
It does remind me of the air conditioning,
这确实让我想到空调，
28:23
that's kind of how air conditioning works.
空调差不多就是这么运作的。
28:24
But also, I think it's just a question of,
但另外，我觉得问题就在于，
28:28
why don't you just buy every sensor in the world
你为什么不干脆把世界上所有的 sensor 都买下来，
28:29
and just ridiculously over-instrument,
然后荒谬地 over-instrument，
28:32
over-replicate everything.
把所有东西都过度复制。
28:33
How we do that, right?
我们就是这么做的，对吧？
28:34
Yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah.
对对对对对对对对对对。
28:36
Right.
对。
28:37
Yeah, I think it's a matter in front.
嗯，我觉得这是摆在眼前的问题。
28:39
Yeah, exactly.
嗯，没错。
28:40
Are you getting visibility to individual atoms
你能看到单个原子吗
28:42
or is this infeasible?
还是这根本不可行？
28:43
But no, we add a huge amount of telemetry
但不是，我们会加大量 telemetry
28:46
and that basically reduces the number of hidden variables,
而这基本上会减少 hidden variables 的数量，
28:49
if you will.
如果你愿意这么说的话。
28:49
Yeah, just out of curiosity,
嗯，就是好奇问一下，
28:51
it is like, do you worry that you're just in California?
就是，你会不会担心你们就只是在 California？
28:54
Like, don't you want to do this in Nepal
比如，你不想在 Nepal
28:57
and Australia?
和 Australia 做这个吗？
28:59
Yeah, we could.
嗯，我们可以。
28:59
I mean, we have plans to open more labs.
我是说，我们有计划开更多实验室。
29:02
Because obviously gravity and where you face
因为很明显，重力、你面向哪里
29:05
and where you are matters.
还有你在哪里，这些都很重要。
29:07
Also, expertise can change, right?
而且，专业能力也是会变的，对吧？
29:09
Like, here, we have certain kinds of expertise.
就像在这里，我们有某些类型的专业能力。
29:12
For example, Stanford Physics Department,
比如说，Stanford Physics Department，
29:13
Engineering Department has certain kinds of expertise
Engineering Department 有某些类型的专业能力，
29:16
that really helps us because we can hire
这对我们真的很有帮助，因为我们可以招聘，
29:18
from there our resources can collaborate.
从那里，我们的资源也能协作。
29:20
Like, as you said, the geography can matter.
就像你说的，地理位置会有影响。
29:22
But then different cities can have some differences
但不同的城市也会有一些差异
29:25
so expertise and physics, chemistry.
所以，专业知识，还有物理、化学。
29:27
Yeah, oh, I mean, we have Zoom.
嗯，哦，我是说，我们有 Zoom。
29:28
Yeah, for that.
对，就为那个。
29:29
Thanks.
谢谢。
29:30
Oh, yeah.
哦，对。
29:31
It's like your experiments are going to be affected
就像你的实验会受到影响，
29:33
by where you are.
受你所在位置的影响。
29:34
Yeah.
对。
29:35
Yeah, I mean, the future of periodic is not a single lab.
对，我是说，periodic 的未来不是一个单一的 lab。
29:38
I mean, we need to get up, like, you're ready to.
我是说，我们得起来，就像，你已经准备好。
29:40
And, like, go to space and, like, yeah,
然后，就像，去太空，然后，就像，对，
29:43
I think it's necessary to have this sort of network of labs.
我觉得得有这种 lab 网络。
29:46
And, like, yeah, I think as there's just kind
然后，就像，对，我觉得，就像，只是在
29:47
of pointing out, like, there's an optimal place
指出，就像，有一个最优的位置
29:49
for each of these based on expertise
对每一个来说，基于专业能力
29:52
or, you know, materials constraint.
或者，你知道，材料限制。
29:53
There's many different constraints permitting.
有很多不同的 constraints 允许这种情况。
29:56
OK, so one of them is, so one thing I've been curious about
OK，所以其中一个就是，嗯，我一直很好奇的一件事
29:59
is how do you deal with computational tools
就是你如何处理 computational tools
30:02
or, like, how do you integrate computational tools
或者，比如说，你怎么把 computational tools
30:04
into a decision-making workflow, especially
整合进一个 decision-making workflow，尤其是
30:07
when, let's say, computation and experiment
当，比如说，computation 和 experiment
30:10
don't necessarily, you know, meet.
不一定，你知道，能对得上。
30:13
It's also, like, theory in terms of, like, high-level,
它也是，就像，theory 层面上的，就是，high-level 的那种，
30:16
like, understanding.
比如说，理解。
30:17
And, yeah, how does that sort of work together?
然后，嗯，这些大概是怎么一起配合的？
30:22
Yeah, so, I mean, one thing that's really helpful here
对，所以，我是说，这里有一个特别有帮助的地方
30:24
is there's certain things that are easier to do in simulation
就是有些事情在 simulation 里做起来更容易
30:28
and more accurate in simulation.
而且在 simulation 里更准确。
30:29
There's certain things that are easier to do in experiment
有些事情在 experiment 里做起来更容易
30:31
and more accurate in experiment.
而且在 experiment 里更准确。
30:32
I mean, in general, of course, experiments are more accurate,
我是说，一般来说，当然，experiments 更准确，
30:35
but certain measurements are very hard to make with experiment,
但有些测量用实验非常难做，
30:38
so you make a bad version of it, so you get inaccurate data.
所以你只能做个很差的版本，于是得到的数据就不准。
30:42
Like, let me just give you some examples.
比如说，我直接给你举几个例子。
30:43
So one of them is we use density-functional theory
所以其中一个就是，我们用 density-functional theory
30:47
to estimate formation-antil-pure materials.
来估算 formation-antil-pure materials。
30:50
Right.
对。
30:51
Click pause.
点一下暂停。
30:52
Can you explain density-functional theory
你能解释一下 density-functional theory 吗？
30:54
information entropy?
information entropy？
30:55
Yeah, that's right, enthalpy.
对，没错，enthalpy。
30:57
Yeah, so density-functional theory
对，所以 density-functional theory
30:58
is probably the most commonly used simulation method
大概是最常用的 simulation method
31:01
for materials.
用于材料。
31:03
It kind of comes from Kohn, Honimberg,
它算是来自 Kohn、Honimberg
31:06
theorem, and Kohn got a Nobel Prize for this.
theorem，而 Kohn 因此拿了 Nobel Prize。
31:09
When he realized that, sure, quantum mechanics
当他意识到这一点时，当然，quantum mechanics
31:11
is very expensive, because you basically
这非常昂贵，因为基本上
31:13
have this exponential Hilbert space for every wave function,
每个 wave function 都要有这个 exponential Hilbert space，
31:16
so you have to solve this very difficult problem,
所以你必须解决这个非常困难的问题，
31:18
especially as the system size grows.
尤其是当 system size 变大时。
31:21
Kohn and Honimberg realized that turns
Kohn 和 Honimberg 意识到，事实证明
31:24
that you don't need to do with that exponential space.
你不需要处理那个 exponential space。
31:27
All you have to do is think about the charge density,
你只需要考虑 charge density，
31:29
at least for the ground state of quantum mechanical systems.
至少对于 quantum mechanical systems 的 ground state 来说是这样。
31:32
So it's actually the theorem is pretty simple.
所以其实这个 theorem 挺简单的。
31:35
So even, I think, an undergraduate in physics can understand.
所以甚至我觉得，一个学 physics 的本科生也能理解。
31:39
But the idea is all the properties of a quantum mechanical
但核心思想是，一个 quantum mechanical
31:42
system ground state is a functional of the charge density,
system 的 ground state 的所有性质，都是 charge density 的 functional，
31:46
which is amazing, because charge density
这太神奇了，因为 charge density
31:48
is just this three-dimensional object,
只是一个 three-dimensional 的东西，
31:51
whereas the Hilbert space and wave functions
而 Hilbert space 和 wave functions
31:53
are exponential.
则是 exponential 的。
31:55
So that was a really interesting observation
所以那真的是一个很有意思的观察
31:57
and then Kohn published another paper
然后 Kohn 又发表了一篇论文
31:59
I think with his post-doc, Sham, Kohn Sham wave functions,
我想，和他的博士后 Sham 一起，Kohn Sham wave functions，
32:03
kind of had a practical way to solving
算是有了一个实用的办法来求解
32:06
some approximation of quantum mechanical properties or materials.
对 quantum mechanical properties 或材料的一些近似。
32:09
And today, we use it very much, both in our lab and externally.
而今天，我们用得非常多，既在我们实验室里，也在外部。
32:13
And then formation enthalpy is one of the measurements
然后，formation enthalpy 是其中一种测量
32:16
we can make to understand the energy of the material.
我们可以做，来理解材料的能量。
32:19
Like, basically, we want to ask, what
就是说，基本上，我们想问的是，这个材料具有的能量是多少？
32:22
is the energy that this material has?
因为如果那个能量比其他材料高，这些原子就能进来，
32:25
Because if that energy is higher than other materials,
那么这个材料就不太可能被制造出来。
32:28
these atoms can come into, then it's
所以 stable 的意思就是，它处在由其他材料及其能量构成的 convex hull 上。
32:31
unlikely that this material will be able to be made.
这样说够清楚了吗？
32:33
So stable just means it's from the convex hall
32:35
of other materials and their energies.
32:40
Was that clear enough?
32:41
Or should we maybe a simple high-level point?
还是说，我们也许该先讲一个简单的高层要点？
32:44
So density-functional theory is a way
所以 density-functional theory 是一种方法
32:47
of taking something which is exponentially hard
用来处理某个会指数级变难的东西
32:50
in number of electrons or number in your system size.
在 electrons 数量上，或者你 system size 里的数量上。
32:53
And then reducing it approximately
然后通过近似把它化简
32:56
was some introduced error to, like,
代价是会引入一些误差，比如说
32:59
maybe usually an incubed approximation.
也许通常是一个 incubed approximation。
33:02
So now you can compute quantities, which are, you know,
所以现在你可以计算一些 quantities，这些 quantities，你知道，
33:05
you can now compute things which are just
你现在可以计算一些原本
33:06
uncomputable, but at some cost.
uncomputable 的东西，但需要付出一些代价。
33:09
If you could do it perfectly, we wouldn't need a lab.
如果你能把它做到完美，我们就不需要实验室了。
33:12
But, you know, we can't.
但是，你知道，我们做不到。
33:14
This might be a good call back to our, actually,
这也许正好可以呼应一下我们，其实，
33:17
our second episode ever, Heather Coolick had said,
我们有史以来的第二期节目，Heather Coolick 曾说过，
33:21
she famously said, there is no alpha-fold for materials,
她有一句名言：没有针对 materials 的 alpha-fold，
33:24
not just computationally, but the ground truth just
不只是 computationally，而是 ground truth 本身就
33:26
doesn't exist.
并不存在。
33:27
Like, we don't really know what material crystal structures
就像，我们其实并不知道 material crystal structures
33:30
look like, and DFT is generally our best route towards.
长什么样，而 DFT 通常是我们最好的途径。
33:34
Yeah, I think so.
嗯，我觉得是这样。
33:35
I mean, maybe two things I would add is,
我是说，也许我会补充两点，
33:37
DFT by definition doesn't have to be an approximation.
DFT 从定义上讲并不一定非得是 approximation。
33:40
Like, the cone-honum bacterium shows that it can be exact,
就像，cone-honum bacterium 表明它可以是精确的，
33:43
but I think as you're pointing out the exchange correlation
但我觉得，正如你指出的，exchange correlation
33:45
functional, we don't yet have access to.
functional，我们目前还接触不到。
33:48
And then for the charge density-based models,
然后，对于 charge density-based models，
33:50
we don't know what the kinetic energy functional is.
我们不知道 kinetic energy functional 是什么。
33:52
And then, even if we had perfect DFT,
然后，即使我们有完美的 DFT，
33:54
I still think we'd need a lab, because even if we had perfect DFT,
我还是觉得我们需要实验室，因为即使我们有完美的 DFT，
33:59
we cannot fit 10 to the 23 atoms in the computer.
我们也没法把 10 的 23 次方个原子塞进计算机里。
34:04
You may have said that, but we know.
你可能是那么说过，但我们知道。
34:06
You know, OK, OK.
你知道的，OK，OK。
34:07
Yeah, but with exponentially growing compute,
对，但随着 compute 指数级增长，
34:10
you can argue that in in-cubed scaling size,
你可以说，在 in-cubed scaling 的规模下，
34:13
eventually you could, in principle, calculate things.
最终你原则上是可以把东西算出来的。
34:15
You just wait till otherwise long enough.
你只要等得足够久就行。
34:17
Massive parameters, yeah.
海量 parameters，对。
34:19
But as long as in-cubed is a good scaling, you can.
但只要 in-cubed 是个好的 scaling，你就可以。
34:24
As the non-scientist, I will make an observation,
作为非科学家，我来做个观察，
34:26
which I don't know if you want to throw around
不知道你想不想拿出来聊聊。
34:28
or anything like that.
或者类似那样的东西。
34:29
So I came from finance.
所以我来自金融行业。
34:31
We had the Gaussian copula, which was a way of pricing,
我们有 Gaussian copula，它是一种定价方式，
34:34
like, credit default swaps by doing correlations
比如，通过做 correlation 来给 credit default swaps 定价，
34:37
and reducing everything into a single kernel.
然后把所有东西都压缩进一个单一的 kernel。
34:40
And it feels also spiritually similar to the VAE
而且感觉在精神上也和 VAE 很像
34:44
in terms of condensing all these things
在把所有这些东西浓缩起来这方面
34:46
into a single sort of parameter.
变成某种单一的 parameter。
34:48
I wonder if they're the same trick everywhere.
我想知道是不是到哪儿都是同一个套路。
34:51
Your sounds, you know, a bit more high-dimensional than mine,
你的声音呢，你知道，比我的要稍微 high-dimensional 一点，
34:55
which, you know, like, VAE is just like, you know,
这个呢，你知道，就像，VAE 就只是，你知道，
34:57
one single thing.
一个单一的东西。
34:59
But like, sounds the same.
但，就，听起来是一样的。
35:01
Yeah, the charge density file is still quite a big file.
对，charge density 文件还是相当大的一个文件。
35:04
Yeah, right.
是啊，没错。
35:05
You hand wave a bit.
你有点含糊带过了。
35:06
But like, it's good enough.
但怎么说呢，这就够用了。
35:07
Like, this is the highest order bit.
就，这是 highest order bit。
35:09
Like, yeah.
就，对。
35:10
I personally don't know a better method
我个人不知道有更好的方法
35:12
to predict the stability of a new material.
来预测一种新材料的稳定性。
35:15
It's definitely not perfect, but it's definitely
它肯定不完美，但它肯定
35:17
better than other methods I can think of.
比我能想到的其他方法都好。
35:21
And yeah, so going back to your question,
然后，对，所以回到你的问题，
35:22
the reason we use them is, for certain things,
我们用它们的原因是，对某些事情来说，
35:25
it's easier to simulate than experiment.
做 simulation 比做实验更容易。
35:27
And they kind of complement each other well.
而且它们某种程度上能很好地互补。
35:30
And then we can, like, do this in a loop.
然后我们就可以，像是，在一个 loop 里做这件事。
35:31
Like, we feel like simulations will never
就像，我们觉得 simulation 本身永远不会
35:33
be enough by themselves.
足够。
35:34
But in the loop of simulations, AI and experiments,
但在 simulation、AI 和实验的 loop 里，
35:38
I think we can make progress much faster than before.
我觉得我们能比以前快得多地取得进展。
35:41
One thing the simulations that you do
你们做的那些 simulations 有一个特点，
35:42
is scale up much more quickly.
就是能 scale up 得快得多。
35:44
I mean, there's now, like, compute
我的意思是，现在，比如说，compute
35:46
is much easier to bring online than bringing on new labs.
比起拉来新的实验室，要容易上线得多。
35:51
How do you avoid, or let's say you scale up a bunch of DFT,
你怎么避免，或者说，假设你 scale up 一大堆 DFT，
35:54
how do you avoid in having your models over index
你怎么避免让你的 models over index
35:56
on that and still be grounded that the real world
在那上面，同时仍然 grounded 地认识到，现实世界
36:01
is ultimately the truth you care about.
才是你最终真正在乎的真相。
36:03
If you, you know, how do you scale up the lab as well?
如果你，你知道，你怎么把 lab 也 scale up 呢？
36:06
Yeah, you scale up the lab.
对，你把 lab scale up。
36:08
I mean, I could imagine scaling up DFT to millions,
我是说，我可以想象把 DFT scale up 到几百万，
36:11
hundreds of millions given modern compute.
在现代 compute 下，到几亿。
36:13
Whereas your lab, I would imagine,
而你的 lab，我猜，
36:16
you're hundreds, thousands a day.
你一天是几百、几千。
36:18
I don't know, like, something still, like, fairly,
我不知道，就像，某种东西还是，就像，相当，
36:21
even in a really high-throughput case,
即使是在非常 high-throughput 的情况下，
36:23
like, you know, you're still fairly limited
就，你知道，你还是挺受限的
36:26
by comparison, right?
相比之下，对吧？
36:27
Yeah, I mean, one thing is the humans,
对，我是说，一个方面是人类，
36:31
but also the LAMs are kind of pretty
但另外，LAMs 其实也挺
36:33
aware of the limitations of DFT.
清楚 DFT 的局限。
36:35
For example, even if we can do it,
比如说，就算我们能做，
36:37
you said 100 million trials with DFT,
你说用 DFT 做一亿次 trials，
36:39
we could never actually try their microstructure.
我们也根本没法真的去试它们的 microstructure。
36:42
And microstructure is this idea that, like,
然后 microstructure 就是这么一个概念，就比如说，
36:44
DFT usually simulates the perfect crystal,
DFT 通常会模拟 perfect crystal，
36:46
but crystals usually actually aren't perfect,
但 crystal 通常其实并不完美，
36:49
and their microstructure, which is the structure
而它们的 microstructure，也就是那种
36:51
at, like, a medium order, can really affect the properties.
在，就比如说，一个 medium order 上的结构，真的会影响 properties。
36:54
Other issues, of course, like, some properties
当然，还有其他问题，比如说，有些 properties，
36:57
like superconducting temperature
比如 superconducting temperature，
36:59
cannot be easily simulated by DFT.
没法很容易地用 DFT 模拟。
37:01
Again, like, even if you do 100 million trials with DFT,
再说，就像，就算你用 DFT 做一亿次试验，
37:04
you won't be enough.
也还是不够。
37:04
So we have to rely on heuristics, experiments, et cetera.
所以我们得靠 heuristics、实验等等。
37:08
And so there's a huge filter from all those calculations
所以从所有那些计算，会有一个巨大的筛选，
37:11
to actually what gets executed in the lab.
再到实验室里实际执行的东西。
37:13
Is that human-mediated or a LAM, or a mix?
这是 human-mediated，还是 LAM，还是混合？
37:17
Can be a mix.
可以是混合的。
37:18
Like, for example, for a long-time materials project,
比如说，对于一个长期的材料项目，
37:20
which was the best open source DFT database,
哪个是最好的 open source DFT database，
37:23
used experimental calibration for DFT.
用 experimental calibration 来校准 DFT。
37:26
So they would actually use experimental data to calibrate DFT.
所以他们实际上会用 experimental data 来校准 DFT。
37:29
We also do this internally here.
我们内部这里也这么做。
37:31
Like, whenever there's a certain chemical system
就像，每当有某个特定的 chemical system
37:33
we're interested in, you get experiments,
我们感兴趣时，你会拿到 experiments，
37:35
you get simulations, and you can calibrate them.
你会拿到 simulations，然后你可以对它们做 calibration。
37:37
This is another thing.
这是另一件事。
37:38
I've talked to a lot of biologist friends,
我跟很多生物学家朋友聊过，
37:41
and they are always very confused when you say materials
每次你说到材料的时候，他们都特别困惑，
37:45
you can't just do an XRD structure
你不能只做一个 XRD structure
37:47
and actually know what the structure is.
就真的知道 structure 是什么。
37:48
Because if you're in the biology world,
因为如果你是在生物学领域里，
37:50
you can look at a crystal structure of a protein
你可以看一个 protein 的 crystal structure，
37:52
and typically, more or less, with sub-a,
而且通常，或多或少，能以 sub-a 的精度
37:55
order things from accuracy reconstructed.
把东西的 order 重构出来。
37:58
What is the difference between the materials, XRD,
材料里的 XRD 有什么区别，
38:02
you know, X-ray diffraction, sorry.
你知道，X-ray diffraction，抱歉。
38:07
Experimental XRD for materials versus,
材料上的实验性 XRD，对比，
38:11
let's say, the biology world.
比如说，生物领域。
38:12
Like, what information do you lose
就像，你会丢失什么信息
38:13
and why is this sort of like a lossy projection?
以及为什么这有点像一种 lossy projection？
38:16
It's hard not to sound like crazy
很难不让自己听起来像疯了
38:18
when answering this question,
在回答这个问题的时候，
38:19
but do you not feel like organic chemistry and biology,
但你不觉得有机化学和生物学，
38:24
almost stem from a lower VC dimension,
几乎都源自一个更低的 VC dimension，
38:26
like lower comical complexity model?
像是 comical complexity 更低的模型？
38:29
Like, you can almost represent as a one-de-sequence
就像，你几乎可以把它们表示成一个 one-de-sequence
38:32
and it can almost be compiled.
而且它几乎可以被 compiled。
38:33
So like, there's a sense in which it's a simpler
所以就像，从某种意义上说，它更简单
38:37
in the comical complexity sense.
在 comical complexity 的意义上。
38:38
One D?
One D？
38:39
I mean, that just DNA is like just...
我是说，DNA 就有点像只是……
38:42
In your case, yeah, yeah.
在你这种情况下，对，对。
38:44
Or RNA.
或者 RNA。
38:45
I think most people argue it's more 2D, but...
我觉得大多数人会说它更偏 2D，但是……
38:47
Okay, I'm not a biologist, yeah.
好吧，我不是生物学家，对。
38:50
I'm sure they're right.
我确定他们是对的。
38:51
But with inorganic chemistry,
但说到 inorganic chemistry，
38:54
virtually it's crazy.
简直可以说是疯了。
38:55
It's clearly not coming from such a low comical complexity model
这显然不是来自一个复杂度低得那么可笑的模型
38:58
because that 3D inorganic crystals can host metals, insulators,
因为那种 3D inorganic crystals 可以容纳 metals、insulators、
39:04
superconductors, diamond, it's like really crazy different.
superconductors、diamond，差别真的超级大。
39:07
Like, people tried, but we could never find a one-de-representation
就像，人们试过，但我们一直找不到一种 1D representation
39:10
that is very hard to find a smile string
也就是说，很难找到一个 SMILES string
39:12
for 3D inorganic crystals.
用于 3D inorganic crystals。
39:14
And the way the atoms interact with each other
而且原子彼此相互作用的方式
39:16
is quite strange, like covalent bands,
还挺奇怪的，比如 covalent bands，
39:19
ionic bands, metallic bands.
ionic bands、metallic bands。
39:22
But yeah, this is very interesting.
不过说真的，这非常有意思。
39:24
I wonder sometimes, like, you know,
我有时候会想，就，你知道，
39:26
have you ever looked into issues with solid electrolytes?
你有没有研究过 solid electrolytes 的问题？
39:29
Like, why people cannot replace liquid electrolytes
比如说，为什么人们不能把 liquid electrolytes 替换掉
39:31
in batteries with solids?
在电池里，用 solids？
39:33
But they're always dandrised that form.
但总是会有 dendrites 形成。
39:35
Like, basically the lithium will form these structures
就，基本上 lithium 会形成这些结构。
39:38
that go into solid electrolyte and crack it.
会进入 solid electrolyte 并把它弄裂。
39:40
And you look at that and you think
然后你看着这个就会想，
39:42
this is so primitive compared to biology.
跟生物学相比，这也太原始了。
39:44
Like, in biology, these systems can do such complex
比如说，在生物学里，这些系统能取得如此复杂的
39:47
nano-technological progress,
nano-technological 进展，
39:51
whereas we can't even like make two interfaces work with each other
而我们甚至都没法让两个 interfaces 彼此配合，
39:54
because lithium will break it.
因为 lithium 会把它弄坏。
39:56
Yeah, but then I guess inorganic stuff
对，但那样的话，我猜就是 inorganic 那类东西
39:58
can be very resistant to high temperature.
可以非常耐高温。
40:00
You can make space shuttles,
你可以造航天飞机，
40:01
you can make silicon, compute Mars law.
你可以造 silicon，compute Mars law。
40:04
So yeah, it's a trade-off.
所以啊，这就是一种 trade-off。
40:05
Yeah, we talked about this with several other guests.
对，我们和其他几位嘉宾聊过这个。
40:08
I think one of the big differences in my mind
我觉得，在我看来，其中一个很大的区别是
40:11
is that biology, we have a toolkit that, you know,
就是，在生物学里，我们有一个 toolkit，它，你知道，
40:14
that you can basically borrow, like,
基本上你可以借用，就像，
40:17
from millions of years of evolution.
来自数百万年的进化。
40:19
Yeah.
对。
40:20
That has already given us all the tools
那已经给了我们所有的工具
40:22
and we can just reproduce those.
而我们只要把它们复现出来就行。
40:23
But also evolution also constrains the VC dimension
但进化也限制了 VC dimension
40:27
we fairly low for, like proteins and so on.
对于像 proteins 之类的，我们相当低。
40:30
Though, like, emergent beyond behavior
不过，就像，那种 emergent beyond behavior
40:32
of large-scale systems like cells
属于像 cells 这种大规模系统的
40:34
could be much more complicated.
可能会复杂得多。
40:35
Yeah, and consciousness.
对，还有 consciousness。
40:37
Yeah, yeah, yeah.
对，对，对。
40:38
Yeah, but maybe even just like, more specifically,
对，但也许甚至可以更具体地说，
40:40
for, like, if you just look at an X-ray,
比如说，如果你只是看一张 X-ray，
40:43
you had a perfect crystal and you shoot an X-R-D at it
你有一个 perfect crystal，然后朝它打一个 X-R-D
40:48
and you get the, you know, you get the spectrum.
然后你得到那个，你知道，你得到 spectrum。
40:50
That still doesn't actually tell you uniquely
那实际上还是不能唯一地告诉你
40:53
what the crystal structure is, right?
crystal structure 是什么，对吧？
40:55
So, and that's something that...
所以，而且那是某种……
40:56
What?
什么？
40:57
It's just average behavior.
那只是 average behavior。
40:59
But, like, for example, this proportionation
但是，就像，比如说，这个 proportionation
41:01
is a very complicated thing, right?
是非常复杂的一件事，对吧？
41:02
Where it's possible that you're the perfect crystal
在这种情况下，有可能你就是那个 perfect crystal
41:04
and on average, it looks like it's perfect X-R-D.
而且平均来说，它看起来像是完美的 X-R-D。
41:06
But in reality, there's some other pattern going on
但现实中，还有另一种 pattern 在发生
41:10
where, like, atoms slightly shift to the right or left,
就是，atoms 会稍微往右或往左偏一点，
41:12
but on average, they're in the middle.
但平均下来，它们是在中间的。
41:14
Yeah, it's a difficult life, yeah.
是啊，这日子挺难的，是啊。
41:16
Or is with proteins, you basically know
或者对于 proteins 来说，你基本上知道
41:18
there's some sort of underlying protein
有某种底层的 protein
41:20
with false some rules.
带着一些 false 规则。
41:21
So, you can combine a forward model with it.
所以，你可以把 forward model 跟它结合起来。
41:23
Right.
对。
41:24
And then you can kind of move towards, like, you know,
然后你差不多就可以转向，就，你知道，
41:26
single crystal, X-R-D, and that can help.
single crystal、X-R-D，那会帮上忙。
41:28
When we started the science pod,
我们刚开始做 science pod 的时候，
41:30
there were all these, like, elements of science, AI and
有各种，就，科学的元素，AI 和
41:33
science, AI and math, AI and physics and all these things.
科学，AI 和数学，AI 和物理，还有所有这些。
41:35
And there was just, like, kind of pissing contest,
而且那就有点像在比谁更牛，
41:37
like, which is harder.
就，哪个更难。
41:38
And, like, I think, like, you can make the scientific
而且，就是，我觉得，就是，你可以提出一个科学上的
41:40
argument that my material science has started.
论点，说我的 material science 已经开始了。
41:42
Well, we can pick it, necessarily, for hardness.
嗯，我们一定可以为了硬度来挑它。
41:45
I mean, there's actually, like, a lot of areas
我是说，其实，就是，有很多领域
41:47
where it is easier.
在那里会更容易。
41:49
I mean, we have these, like, great simulators
我是说，我们有这些，就是，很棒的 simulators
41:50
for large classes of materials.
针对大类材料。
41:53
Whereas in biology, it's incredibly difficult
而在 biology 里，那就难到离谱。
41:55
to model a cell or an organ or a full organism.
来给一个 cell、一个 organ，或者一整个 organism 建模。
42:00
So, I mean, that's a huge advantage.
所以，我是说，那是个巨大的优势。
42:01
Yeah.
对。
42:02
Isn't that interesting that at the smallest level,
是不是很有意思，在最微小的层面上，
42:05
it is one-dimensional or two-dimensional?
它是 one-dimensional 还是 two-dimensional？
42:07
And then, like, it just has many orders more scale
然后，就像，它只是多了很多个数量级的 scale
42:10
and structure.
和 structure。
42:11
And, like, and that actually introduces the complexity.
而且，就像，这实际上引入了 complexity。
42:14
It's weird.
这挺奇怪的。
42:15
Like, where's, like, maybe the complexity of materials
就，怎么说呢，也许 materials 的 complexity
42:18
is more the microstructure?
更多是在 microstructure 上？
42:19
Yeah, I mean, I still get, like, different complexity
对，我是说，我还是会觉得，会有不同的 complexity
42:21
at the different length scales for materials as well.
在 materials 的不同 length scales 上也是这样。
42:24
So yeah, microstructure and beyond.
所以对，microstructure 以及更往后的层面。
42:27
Yeah.
对。
42:28
Just sort of, like, a brief, like, sort of palate cleanser.
就有点像，一个简短的、嗯，换个口味的小插曲。
42:31
In terms of science fiction, let's say
从科幻的角度来说，假设
42:32
you can invent whatever materials you want.
你可以发明任何你想要的材料。
42:35
What is more valuable?
什么更有价值？
42:37
I have a list here.
我这里有个清单。
42:39
There's obviously room temperature
显然有 room temperature
42:40
supreme directors.
supreme directors。
42:42
But also, you know, you mentioned batteries.
不过，你知道，你也提到了电池。
42:43
I've always thought, like, batteries.
我一直觉得，就像，电池。
42:45
Just, like, do better than living in my own.
就是，像是，比我自己的活法要更好。
42:47
You're good, like, you know.
你挺好的，就是，你知道吧。
42:49
That's stood around for, like, 100 years.
那东西已经搁那儿，差不多，100年了。
42:51
And then, the last one is carbon nanotubes.
然后，最后一个就是 carbon nanotubes。
42:54
Mostly for the space elevators.
主要是给 space elevators 用的。
42:56
I don't know if, like, those are three that come to mind.
我不知道，就是，脑子里能想到的是不是就这三个。
42:59
Are there any, like, people talk about your world?
有没有什么，就是，你们那个世界里人们会聊到的？
43:02
That, like, are the dream?
那种，就是，算是梦想的？
43:03
Definitely, it's for conductors, magnets,
43:06
especially if you can lower the dependence on rare earth
43:09
or, like, transition metals that are hard to source,
43:12
like, cobalt reducing cobalt in batteries.
43:15
One of the things I think is very fundamentally important
43:17
is if we can get close to land our limit
43:21
for compute energy efficiency.
43:24
Can you explain land our limit?
43:26
Yeah.
嗯。
43:27
So if you look at how much energy we spend per,
所以如果你看我们每单位消耗多少能量，
43:31
like, floppin' compute, it is followed
比如说 floppin' compute，它遵循
43:33
the Moore's Law type behavior
Moore's Law 那种规律
43:35
based on exponentially more efficient.
基于指数级更高的效率。
43:37
So we've been spending exponentially less energy
所以我们一直在指数级地减少能量消耗
43:40
per floppin' compute.
每单位 floppin' compute。
43:41
I mean, I haven't checked recently.
我是说，我最近没查过。
43:42
This was the case 10 years ago.
十年前情况就是这样。
43:44
Yeah, that's great.
对，那挺好的。
43:45
I remember when I first learned about it,
我记得我第一次了解到它的时候，
43:46
looked at that, and we were, like, 10, 20
看了那个，我们当时大概还差着 10、20
43:48
in order to magnitude away from it or something.
orders of magnitude 那么远之类的。
43:50
And the last time I looked, it's, we're weirdly close.
而我上次看的时候，已经，我们怪异地接近了。
43:54
Like, we're within a factor of, I don't know,
就像，我们在 factor 以内，我也不知道，
43:56
like, three or four orders of magnitude away?
大概三四个 orders of magnitude 那么远？
43:58
Yeah.
嗯。
43:59
It's actually, like, I was rolled out.
其实，怎么说呢，我被 rolled out 了。
44:01
Yeah, yeah, yeah.
嗯，嗯，嗯。
44:02
It's still over many years.
不过这还是很多年的事。
44:03
Yeah, I'm pretty impressed with it.
嗯，我对它印象挺深的。
44:05
But, like, in your lifetime, you used to say something.
但是，怎么说呢，在你这一生里，你以前也说过一些话。
44:07
Yeah, since, just since I discovered this way,
嗯，自从……就从我发现这种方式开始，
44:09
what 10 or 15 years ago, yeah, no.
大概 10 年还是 15 年前吧，嗯，不是。
44:11
And one of the things we have to do is dissipate heat,
而我们必须做的一件事就是散热，
44:14
because, like, is these computations that are happening
因为，就像，这些正在发生的 computation
44:17
is producing heat, which has to happen,
会产生热量，这是必然的，
44:18
but then we have to dissipate it.
但之后我们得把它散掉。
44:20
So, like, if we can get close to the lend-out limit
所以，就像，如果我们能接近 lend-out limit
44:22
for how efficiently we do computation,
就我们做 computation 的效率而言，
44:24
that's amazing for humanity, right?
那对人类来说就太棒了，对吧？
44:26
Because that means now we're doing computation,
因为这意味着现在我们正在做 computation，
44:28
it's probably one of the most fundamental things
这大概是最根本的事情之一
44:29
we do as humanity, but as efficiently as possible
我们作为人类所做的，但要尽可能高效
44:31
from energy perspective, which is a real currency
从 energy 的角度看，这是一种真正的货币
44:33
in the universe.
在宇宙中。
44:34
That gets affected by quantum computing.
而这会受到 quantum computing 的影响。
44:36
I'm just going to throw it out there.
我就直说了吧。
44:38
You're radically, massively, embarrassingly,
你是彻底地、大规模地、embarrassingly
44:40
parallel compute.
parallel compute。
44:42
So, it definitely gets affected by reversible computing.
所以它肯定会受到 reversible computing 的影响。
44:46
But, honestly, I'm not an extra, I don't understand.
但说实话，我不是个专家，我不懂。
44:48
If you can do reversible computing, now you don't actually
如果你能做 reversible computing，那你其实就不需要
44:51
have to spend energy to do computation,
花能量来做 computation，
44:53
because you're not actually deleting any information,
因为你其实没有删除任何 information，
44:54
it's reversible.
它是可逆的。
44:56
But I don't know if that works.
但我不知道那能不能行。
44:58
I don't know.
我不知道。
44:58
Quantum computing, I assume, has to obey these laws somehow,
Quantum computing，我猜，总得在某种程度上遵守这些定律，
45:03
because they are still like, they have to obey
因为它们还是，呃，它们必须遵守
45:05
to them.
它们。
45:06
So, yeah, it's reversible.
所以，对，它是 reversible 的。
45:09
Well, one of the most memorable conversations
嗯，最难忘的对话之一
45:11
with Elad Gil was, like, he actually
是和 Elad Gil 的，就是，他其实
45:12
was a very skeptical person about quantum computing.
是个对 quantum computing 非常怀疑的人。
45:15
He's like, even if he had it today, there's no applications.
他说，就算他今天有这玩意儿，也没有什么 applications。
45:18
I'm like, whoa, it breaks the security RSA, like that's it.
我当时就想，哇，它把 RSA 的安全性给破了，就这么回事。
45:22
I agree with that.
我同意这一点。
45:24
Like, people talk about how if you're a quantum computer today,
就像，人们会说，如果你今天有一台 quantum computer，
45:26
you could simulate things so much better.
你能把东西 simulate 得好太多了。
45:29
And I asked him, OK, let's accept that.
然后我问他，好吧，我们就接受这一点。
45:31
What would you simulate?
那你会 simulate 什么？
45:32
You're still simulating perfect crystal.
你还是在 simulate perfect crystal。
45:34
Like, you still have the issues that the NFT has,
就像，你还是会有 NFT 有的那些问题，
45:36
if the NFT was perfect in this prediction ability.
如果 NFT 在这种预测能力上完美无缺。
45:39
So, I do think people are kind of glossing over some things
所以，我确实觉得人们有点把一些事情一带而过了。
45:42
because quantum computing is so exciting.
因为 quantum computing 真的太令人兴奋了。
45:44
Like, if we can compute in a quantum logic space
比如，如果我们能在 quantum logic space 里计算，
45:47
instead of classical logic, it's just so exciting
而不是 classical logic，那简直太令人兴奋了，
45:49
that people are glossing over what
以至于人们都忽略了，
45:51
it would actually do when it's made.
它真正被造出来时到底会做什么。
45:53
Maybe that's fine, because maybe once it's made,
也许这没关系，因为也许一旦它被造出来，
45:55
it will do amazing things we can't even imagine.
它会做出我们甚至无法想象的惊人的事情。
45:57
Good.
很好。
45:58
OK.
OK。
45:59
I was going to move to the automating the lab
我本来想转到 automating the lab
46:02
side where you've closed the loop on your experimentation.
这一侧，也就是你在 experimentation 上 closed the loop 的那一侧。
46:05
You know, talk about robotic arms.
你知道，聊聊 robotic arms。
46:08
Talk about, basically, it's just everything you've done here.
聊聊，基本上，就是你在这里做的所有事情。
46:10
My favorite quote from you was that every piece of equipment
我最喜欢你说的那句话是，每一台设备
46:14
in your lab is going to have 140 IQ.
在你的实验室里，会有 140 IQ。
46:16
OK.
OK。
46:17
In the early days, as we were scaling things up,
早期，当我们正在做 scaling、不断扩大规模的时候，
46:19
we realized that we had huge bottlenecks
我们意识到，我们遇到了巨大的 bottlenecks，
46:22
imposed just by operating machinery.
仅仅是因为操作机器而造成的。
46:24
So, for once in a machinery, we would have technicians
所以，有一次在机器里，我们会有技术员，
46:27
and scientists looking for, particularly,
以及科学家，尤其是寻找，
46:30
morphology and trying to see what actually
morphology，并试图看看实际到底是什么
46:34
were we making, see if this is consistent with our intentions.
我们当时在做的，看看这是否符合我们的意图。
46:39
And we really quickly, as we scale up the lab,
而随着我们把实验室规模扩大，我们很快就
46:41
came into these bottlenecks where it just wasn't keeping up.
撞上了这些瓶颈，它根本就跟不上。
46:44
So we started doing some programmatic approaches
于是我们开始做一些程序化的方法
46:47
to capturing the data on SCM.
来采集 SCM 上的数据。
46:50
And the data that was being surfaced
而这些浮现出来的数据
46:53
from this very simple program, where it sort of takes
来自这个非常简单的程序，它有点像会
46:56
the field of view, captures, zooms in, captures,
取 field of view，截取，放大，再截取，
46:59
was just not all that useful.
其实并没那么有用。
47:01
It was a little too dumb for what we really wanted to cut.
它有点太笨了，做不了我们真正想 cut 的东西。
47:06
And so, at that point, it was then a really pertinent
所以，到那个时候，真正关键的是
47:08
to build AI systems directly onto the machines
直接把 AI systems 搭建到这些机器上
47:11
to start controlling these things.
来开始控制这些东西。
47:13
And now they have the full context
而现在它们有了完整的 context
47:15
as to what we were trying to achieve.
知道我们当时到底想实现什么。
47:16
So, what was the intent of the experiment?
所以，这个实验的意图是什么？
47:18
What were we trying to synthesize?
我们当时想 synthesize 什么？
47:20
What were some other experimental evidence?
还有哪些其他的实验证据？
47:22
And it's actually looking in the machine
而且它其实是在机器里面看
47:24
and capturing that data.
并把那些数据 capture 下来。
47:25
And this is really valuable now because, you know,
而这现在真的很有价值，因为，你知道，
47:28
we've been talking about these hidden variables
我们一直在聊这些 hidden variables
47:30
not being able to capture everything,
没法 capture 所有东西，
47:31
but if you can do more intelligent data capture
但如果你能做更智能的 data capture
47:34
at the time of that experiment,
在那个实验的时候，
47:36
your data for future AI systems
你为未来 AI 系统准备的数据
47:38
and future computational predictions is that much better.
以及对未来计算的预测就会好得多。
47:41
You'll have just a richer set of data.
你会拥有更丰富的数据集。
47:44
And so that's sort of what we mean by just, you know,
所以这就是我们所说的，你知道，
47:46
everything on the lab has to be incredibly intelligent
实验室里的一切都必须极其智能，
47:51
to just make the data as useful as possible.
才能让数据尽可能有用。
47:54
So that was sort of the motivation inspiration behind that.
所以这就是那背后的动机和灵感。
47:57
What's the state of the art there in terms of like,
那方面现在最前沿做到什么程度了，就比如，
47:59
putting intelligence on every device, right?
把智能放到每个设备上，对吧？
48:02
Well, I think one interesting aspect is
嗯，我觉得一个有意思的点是
48:07
what's the latency of controlling that instrument?
控制那个仪器的 latency 是多少？
48:10
There's this quality in C, but then there's also
C 里有这种特性，但还有
48:11
device compute, which is.
device compute，也就是。
48:13
That's right.
没错。
48:14
And then there's sort of a time scale
然后还有某种时间尺度
48:16
associated with different like physical processes.
跟不同的，呃，物理过程相关联。
48:18
And if calling out to an API or something
而且如果调用 API 或者什么的话
48:22
is simply too slow, so the amount of reasoning
就是太慢了，所以 reasoning 的量
48:25
or tokens or tool calls is just not matched
或者 tokens 或者 tool calls 就是匹配不上
48:29
to the latency of that actual process,
那个实际过程的 latency，
48:31
then that's infeasible.
那这就不可行。
48:33
Oh, that is right.
哦，对。
48:34
I mean, it's up to like, you know, self-driving cars, right?
我是说，这取决于，呃，你知道，self-driving cars，对吧？
48:36
So I real quick, I'm actually surprised to hear that
那我先快速说一句，我听到这个其实挺惊讶的
48:39
because the latency I'd imagine for reasoning
因为我本来以为 reasoning 的 latency
48:42
seems small compared to a lot of these things
跟很多这类东西比起来好像算小的
48:44
take hours to run, right?
要跑好几个小时，对吧？
48:46
Or maybe, you know, maybe your experiments
或者也许，你知道，也许你的实验
48:49
are much faster, high-through, or something.
要快得多，high-through，或者什么的。
48:51
I'm just surprised to hear that.
我就是听到这个很惊讶。
48:52
So ultimately the experiments do take many hours days,
所以最终，这些实验确实要花很多小时、很多天，
48:57
but there could be particular steps
但可能有些特定的步骤
48:59
where latency wouldn't matter.
在这些步骤里 latency 可能就无所谓了。
49:00
Yeah, like you might want to have final control.
对，比如你可能想要有最终控制权。
49:04
Another thing is how expensive it is,
另一个问题是它有多贵，
49:07
because the same reason some of these models
因为同样的原因，这些 models 中的一些
49:10
can be really slow in analyzing data
在分析 data 时可能会非常慢，
49:12
also makes them very expensive.
这也让它们变得非常贵。
49:14
And then finally, the human patients,
然后最后，就是人类患者，
49:16
like if a human wants to analysis
比如说，如果一个人想要分析
49:18
of a certain anxiety pattern
某种焦虑模式
49:19
and they have to wait two hours
而且他们得等两个小时
49:20
before they get a good result,
才能得到一个好的结果，
49:22
that's very different, I think,
我觉得，那会非常不一样，
49:23
than if they can get it in like two minutes.
比起他们能在大概两分钟内得到结果。
49:25
Oh, okay.
哦，好的。
49:26
Yeah, and then the reasoning,
嗯，然后就是 reasoning，
49:27
it's, you know, you're not doing like a single call
就是，你知道，你不是只发一个 single call
49:30
to that model, right?
给那个 model，对吧？
49:31
Like you're doing potentially many, many tokens
而是你可能要用很多很多 tokens
49:34
to kind of get to these patterns.
才能慢慢得到这些 patterns。
49:35
Like run simulations in the loop.
就像在 loop 里跑 simulations。
49:37
Exactly.
没错。
49:38
Deep research in the loop.
在 loop 里做 Deep research。
49:39
It's not a single tool call.
这不是一个 single tool call。
49:41
It's not a single inference thread.
它不是单一的 inference thread。
49:42
So then the latency can blow up quite a bit.
所以 latency 可能会飙高不少。
49:44
I see what you mean.
我明白你的意思。
49:45
Yeah.
嗯。
49:46
So how much of your latency is tool calls?
那你 latency 里有多少是 tool calls？
49:48
And I assume tool calls is like largely like DFT
而且我猜 tool calls 很大程度上就像 DFT
49:51
or computational signals and therefore, yeah.
或者 computational signals，所以，嗯。
49:53
So how much of your latency is derived from those
那你 latency 里有多少是来自这些的？
49:56
versus like actual reasoning?
还是说，像是真正的 reasoning？
49:59
It's really processed dependent.
这真的很 process-dependent。
50:01
Okay, yeah.
好，嗯。
50:01
Yeah.
嗯。
50:02
And then the other thing I think about is also,
然后我还会想到的另一件事是，
50:04
I guess building up from small things to bigger things,
我猜是从小东西慢慢搭到更大的东西，
50:07
where I assume that the general temptation
我假设，通常的诱惑
50:10
or the typical development is incremental,
或者说典型的发展，是 incremental 的，
50:12
where like everything is human operated
就是那种所有东西都得靠人工操作的情况
50:15
and then you find ways in which to automate it
然后你会找到把它自动化的方法
50:17
and then you sort of build up from there.
然后你就从那儿一点点搭起来。
50:19
I worry that sometimes that that is the way
我担心有时候，那就是那种方式
50:21
that people evolve things,
人们让事物演化的方式，
50:22
but that's a local minima.
但那只是一个 local minima。
50:24
Optimus?
Optimus？
50:25
Sure.
好啊。
50:26
Exactly.
没错。
50:27
When like actually you should get a humanoid
其实吧，那时候你就该搞个 humanoid
50:29
and just put them in there.
然后直接把它们放进去。
50:31
We, I think we're the opinion that solving humanoids
我们，我觉得我们的观点是，解决 humanoids
50:34
would actually be slower to kind of getting
其实反而会更慢，才能算是达到
50:36
to some of our goals.
我们的某些目标。
50:37
Just checking.
只是确认一下。
50:38
Again, a lot of this is just like you do this every day.
还是那句话，这里面很多事就跟你每天做的一样。
50:40
We like see this, but I don't know the reality
我们喜欢看到这个，但我不知道实际情况
50:43
of the situation.
到底怎么样。
50:44
A lot of people are happy humanoids.
很多人都是快乐的 humanoids。
50:45
Yeah.
嗯。
50:46
I mean, I think like maybe one one process is
我的意思是，我觉得可能有一种、一种流程是
50:49
by having like a mix of humans and automation,
通过把人类和 automation 混在一起，
50:53
you can identify really quickly
你就能很快识别出
50:55
like what are some of the bottlenecks
比如有哪些 bottlenecks
50:56
in the experimental process
在实验过程中
50:58
and you start alleviating those bottlenecks one at a time.
然后你开始一个一个地缓解那些瓶颈。
51:01
And so for example, if there's some really tricky dexterity
所以，比如说，如果有一些特别棘手的 dexterity
51:04
task that humans are excellent at
任务是人类非常擅长的
51:06
but you'd have to spend months automating machine,
但要让机器自动化，你得花上好几个月，
51:09
maybe don't spend a ton of time there.
那也许就别在这上面花太多时间。
51:12
And maybe like a more promising thing is something very routine
而也许更有希望的是那种非常常规的事情
51:15
that takes up a huge amount of scientists and technicians
它占用了大量科学家和技术人员的时间
51:17
time and it's easy to automate.
时间，而且它很容易自动化。
51:19
So I think there's like this pragmatism to it.
所以我觉得这里面有一种实用主义。
51:21
But I mean, ultimately what we want from the lab
但我的意思是，说到底，我们想从实验室得到的是
51:24
is a huge quantity of data, high quality data,
海量的 data、高质量的 data，
51:27
diverse data and that those are our goals
多样化的 data，而这些就是我们的目标，
51:30
and full autonomy is a non-goal.
而 full autonomy 不是我们的目标。
51:35
In the service that we use automation
我们使用 automation，
51:36
in the service of achieving the goals on the data.
是为了实现 data 上的这些目标。
51:39
Yeah.
嗯。
51:40
But also, I mean, I think another aspect to automation
但另外，我的意思是，我觉得 automation 的另一个方面
51:43
is robots will just be make fewer mistakes
就是 robots 会少犯错误
51:47
potentially in the lab.
可能在实验室里。
51:48
And it's been really helpful having AI systems
而且有 AI systems 真的很有帮助
51:52
with a full view over all of our data
能全面查看我们所有的 data
51:54
because sometimes if there's like a permutation in data,
因为有时候如果 data 里有类似 permutation 的情况，
51:58
then it can identify it.
它就能识别出来。
51:59
So for example, one of our steps at one point
所以举个例子，我们其中一个步骤有一次
52:02
had a cyclic error because one of the machines
出现了 cyclic error，因为其中一台机器
52:06
was loaded incorrectly.
加载错了。
52:08
And the patterns were inconsistent.
而且这些 patterns 并不一致。
52:10
So the AI was reading through these things and it says,
所以 AI 在读这些东西，然后它说，
52:12
well, given what was run, this is not expected.
嗯，考虑到跑的是什么，这不在预期之内。
52:16
And it's looking at this basket of data long as humanly.
而且它尽可能像人一样长时间地看着这一堆数据。
52:19
And then I realized if I do this cyclic permutation
然后我意识到，如果我做这个 cyclic permutation
52:22
and reverse it, everything is consistent.
然后反过来推，一切就都对得上了。
52:24
And then we're able to go back to the physical infrastructure
接着我们就能回到 physical infrastructure，
52:26
and understand that a mistake had been made in the loading.
并明白是在 loading 的时候出了个错。
52:30
And I think managing data quality is so foundational
而且我觉得，管理 data quality 是非常基础的
52:34
to doing AI in the physical world.
对于在 physical world 里做 AI 来说。
52:37
And these are the types of things that we're building in.
而这些正是我们在内置构建的那类东西。
52:40
And so you can improve the operating process.
这样你就能改进 operating process。
52:42
But at that point, you're like, OK, this
但到了那个时候，你会想，好吧，这个
52:43
is another great opportunity for automation.
是 automation 的另一个绝佳机会。
52:46
How do you make this just so reliable, so durable
你要怎样才能把它做得这么可靠、这么耐用，
52:49
that we never have those types of mistakes again?
以至于我们再也不会犯那类错误？
52:51
Yeah, reducing the noise flow or experimental data
对，减少 noise flow 或者 experimental data
52:54
I think is very important.
我觉得这非常重要。
52:55
Most experimental data in literature
文献里大多数 experimental data
52:57
have such high noise flow that it's actually
都有那么高的 noise flow，以至于它实际上
52:59
usually worse than the FD accuracy.
通常比 FD accuracy 还差。
53:01
Because when different people do experiments,
因为当不同的人做实验时，
53:04
different labs, different parts of the world, different times,
不同的实验室，世界不同地区，不同的时间，
53:07
it really introduces so much fluctuation to the result.
真的会给结果带来很多波动。
53:10
So we are hoping that by standardizing these workflows,
所以我们希望，通过把这些 workflows 标准化，
53:12
one of the biggest benefits will be that the noise flow
最大的好处之一会是，noise flow
53:14
will be lowered.
会降低。
53:16
You've been pretty public about how you use open source models
你一直挺公开地讲你如何使用 open source models
53:19
and finding them for stuff like this.
以及为这类事情去找它们。
53:21
Is it better?
这样会更好吗？
53:22
Is it, basically, I can imagine a situation
是不是，基本上，我能想象出这样一种情况
53:25
where you have maybe the number of models do one task,
就是你有，可能是一批模型一起做一个 task，
53:28
and then you optimize for that task.
然后你针对那个 task 去 optimize。
53:29
And then there's a generalist frontier model
然后还有一个 generalist frontier model
53:31
that supervises everything.
来 supervise 一切。
53:32
Is that a good mental model to have?
这是个好的 mental model 吗？
53:34
Are there more stages to this that I can't think about?
这里面还有更多我想不到的阶段吗？
53:38
I mean, we basically use a mix of open source models
我的意思是，我们基本上会混着用 open source models
53:41
and closed source models.
和 closed source models。
53:43
We're huge beneficiaries of this from our simulation side,
从 simulation 这边看，我们是这件事的巨大受益者，
53:46
building up the ML stack, the simulation stack.
把 ML stack、simulation stack 搭起来。
53:48
We don't need to push on that access.
我们不需要去争取那种 access。
53:51
But there's a lot of areas where the latency is too high.
但有很多领域，latency 太高了。
53:55
The cost is too high.
成本也太高。
53:57
But in some cases, we can actually, you know,
但在有些情况下，我们其实可以，你知道，
53:58
by having access to this data,
通过能获取这些 data，
54:01
we can actually push beyond the frontier
我们其实可以突破 frontier，
54:03
of what these systems can do, under even some,
突破这些系统能做到的极限，哪怕是在一些，
54:06
like the highest reasoning efforts.
比如最高 reasoning efforts 的情况下。
54:08
And so it's not necessarily just about cost
所以这并不一定只是成本问题，
54:10
or prey to efficiency, but in some cases,
或者受效率所困，而是在某些情况下，
54:13
you can push beyond it when you have access to data
当你能获取到 data 时，你可以突破它，
54:15
that no one else has.
而这些 data 是别人没有的。
54:17
One way to characterize this is in terms of compute efficiencies.
一种描述方式是从 compute efficiencies 的角度来看。
54:21
So by having access to this data,
所以，只要能拿到这些 data，
54:24
you can be that much more compute efficient
你的 compute efficiency 就能高那么多，
54:25
compared to some of the frontier models.
跟一些 frontier models 相比。
54:29
And that's been able to be like
而且这一直可以说是
54:31
a really instrumental thing in our program.
我们项目里一个非常关键的东西。
54:33
Yeah, I'm so data quality or data access
对，我是说，data quality 或者 data access
54:37
as a trade-off of compute.
作为 compute 的一种 trade-off。
54:38
Like, you know, I think there's some amount of exchange rate
就，你知道，我觉得多多少少存在一种汇率
54:41
of dollars for compute for dollars for data
就是美元换 compute、跟美元换 data 的
54:44
that people, I feel like the pendulum might be swinging
然后大家，我感觉这个钟摆可能正在往
54:47
out towards data.
data 那边摆。
54:48
I mean, obviously, you would agree with that.
我是说，显然，你会同意这一点。
54:50
But like, yeah, talking about the noise
但就像，对，聊到 noise
54:52
for experimental data, like we need to have high quality data.
对于 experimental data 来说，我们得有 high quality data。
54:55
If you throw a huge amount of compute
如果你砸进去海量 compute
54:57
against a bunch of noise, you're not
面对一大堆噪声，你没法
54:59
going to have a good thing emerge.
让什么好东西涌现出来。
55:00
And it's harder for you because you actually try to be sparse.
而且这对你来说更难，因为你其实是在试图做到 sparse。
55:02
Like, you actually try to get no results
就像，你实际上是在试着得到 no results，
55:04
and learn from that and that's right.
并从中学习，这才是对的。
55:06
Yeah.
是啊。
55:07
Yeah, you've talked about no results
是啊，你谈过 no results，
55:08
and several other venues.
还有其他几个 venue。
55:10
Like, what does a no result look like for materials?
就是，在材料方面，一个 no result 长什么样？
55:12
And how do you use that effectively?
那你怎么有效地利用它呢？
55:14
Especially when I think you're overall
尤其是，我觉得你整体上
55:16
thinking it was probably quite sparse, usually,
认为它通常可能挺稀疏的，
55:18
in terms of success.
从成功这个角度来说。
55:20
I mean, in a no result could be,
我是说，一个 no result 可能是，
55:22
we intended to produce some structure
我们本来打算做出某种结构
55:24
and then all of the evidence points to us
然后所有证据都指向我们
55:27
not producing that structure.
没有生成那个 structure。
55:28
You ever intend to not produce the structure?
你有没有故意不生成那个 structure？
55:30
Of course.
当然。
55:31
Yeah, absolutely.
对，绝对。
55:32
You do the negative controls.
你做 negative controls。
55:34
Absolutely.
绝对。
55:34
Yeah, okay.
嗯，好的。
55:35
Like, there are impurity phases that
就像，有一些 impurity phases，它们……
55:36
will kill your property.
会毁掉你的 property。
55:38
Yeah.
对。
55:39
Okay.
好。
55:40
You can be toxic, you know.
你知道，你可能会很 toxic。
55:41
Exactly.
没错。
55:42
100%.
100%。
55:43
Okay.
好。
55:44
And a negative result, you know, in some cases,
而且 negative result，你知道，在某些情况下，
55:46
is actually, you know,
其实是，你知道，
55:47
we intended to make this the following thing
我们本来是想把它做成下面这个东西
55:49
and then we're actually able to identify
然后我们实际上能够识别出
55:51
a new structure that hadn't previously been identified.
一个之前没被识别出来的新 structure。
55:55
So, it's negative in some respect.
所以，从某种角度来说，这是负面的。
55:58
Like, we intended to do something else,
就是，我们本来是想做点别的，
56:00
but something else emerged from the data.
但别的东西从 data 里冒出来了。
56:02
Yeah.
对。
56:03
I mean, also in general, like,
我是说，而且一般来说，就像，
56:04
when you're training a machine learning algorithm,
当你在训练一个 machine learning algorithm 的时候，
56:06
especially like, they're very basic double
尤其是，比如说，它们其实很基础，而且
56:07
if it's a classification algorithm.
如果它是一个 classification algorithm，那就更是如此。
56:10
If you don't have negative samples,
如果你没有 negative samples，
56:12
you can't really train if everything is positive.
如果全都是 positive，你其实根本没法训练。
56:13
And this is a particularly bad problem in material science
而这在 material science 里是一个特别糟糕的问题
56:16
because people usually publish crystals
因为人们通常只发表 crystals
56:19
they could synthesize,
他们可以合成，
56:20
but they usually don't publish
但他们通常不会发表
56:22
if they fail to synthesize a crystal.
如果他们没能合成出晶体。
56:23
Sometimes they might.
有时候他们可能会。
56:25
And also, like, we never know really
而且，就像，我们其实永远不知道
56:27
if a crystal is not synthesizable ever, right?
一个晶体是不是永远都无法合成，对吧？
56:30
It could just be a skeleton.
它可能只是个骨架。
56:31
Exactly.
没错。
56:32
It could be a skeleton.
它可能是一个骨架。
56:32
It's actually a skeleton.
它其实就是一个骨架。
56:33
Like, synthesis method issue technology, yeah.
就像 synthesis method 的问题，技术上的，对。
56:36
So, I think it really helps us
所以，我觉得这真的帮到我们
56:38
when we do our own experiments
当我们自己做 experiments 的时候
56:40
and get negative results in the context of what we tried.
然后在我们尝试的语境下得到 negative results。
56:42
So then we can't even train a classification algorithm.
所以那样的话，我们甚至连一个 classification algorithm 都训练不了。
56:46
I think there's also another valid thing too
我觉得还有另一个也成立的点。
56:49
of when you have this sort of string of negative results
就是当你有一连串这样的负面结果的时候
56:53
and then finally through process iteration,
然后最终通过 process iteration，
56:56
you're able to get to that positive result.
你才能得到那个正面结果。
56:58
It's a really interesting set of,
这真的是一组很有意思的，
56:59
I'll call it like process engineering type data.
我会把它称为类似 process engineering 类型的数据。
57:03
So it's like through iteration,
所以这就像是，通过 iteration，
57:05
how did you actually get to that correct result?
你实际上是怎么得到那个正确结果的？
57:07
Because so much of material science
因为 material science 中有太多
57:11
has this ambiguity
有这种模糊性
57:12
and actually how something was made.
以及某个东西实际上是怎么做出来的。
57:15
And so, yeah.
所以，对。
57:17
And so there's actually, even if there's a known material,
所以其实，即使有已知的材料，
57:20
it can be highly non-trivial to replicate that.
要复现它也可能非常不简单。
57:23
Some things are in a high school textbook
有些东西在高中课本里就有
57:26
but other things are really at the frontier
但另一些东西真的处在最前沿
57:28
and we're building up that know how as well.
而且我们也在积累那方面的 know-how。
57:31
And then building up the system that,
然后搭建起这样一个系统，
57:33
given the string of negative results,
在一连串负面结果之下，
57:34
how do you actually get to that positive case?
你到底怎么才能真正走到那个 positive case？
57:36
Yeah, going to the classifier result,
对，接下来看 classifier 的结果，
57:37
I would almost assume that if you are doing everything
我几乎会假设，如果你所有事情都
57:40
in-house, most of your results will actually be negative
在公司内部自己做，你的大多数结果其实都会是负面的，
57:43
rather than positive, which is kind of ironic
而不是正面的，这有点讽刺，
57:45
because the literature only gives you positive results.
因为文献里只给你正面的结果。
57:47
So if you were sort of cold starting this,
所以如果你有点是从 cold start 开始做这件事，
57:49
it seems like the problem is actually the reverse
那看起来问题其实正好相反，
57:51
that you have an abundance of negative results
你有一大堆 negative results，
57:55
and not enough positive.
而 positive 不够。
57:56
And as you said, if you were doing per-experiment labels,
而且就像你说的，如果你是按 per-experiment labels 来做，
57:58
it would be mostly negative.
那大部分都会是 negative。
57:59
But if you were doing per campaign
但如果你是按 per campaign 来做，
58:03
and assuming campaigns and when we succeed,
而且假设 campaigns，以及当我们成功的时候，
58:06
then it could be a bit more balanced.
那可能就会更平衡一些。
58:08
I see.
我明白了。
58:09
The reasoning traces could almost be over
reasoning traces 几乎可以覆盖
58:11
entire campaigns. Absolutely.
整个 campaigns。绝对是。
58:13
OK.
OK。
58:13
This is the key thing.
这是关键的一点。
58:14
This type of data basically doesn't exist anywhere else
这类数据基本上在其他任何地方都不存在
58:17
and we spend so much of our time getting
而我们花了那么多时间获取
58:20
the full lineage of the scientific process into the model.
把科学过程的完整脉络都放进 model 里。
58:24
And so tracking all this data, the conversations,
所以追踪所有这些 data、这些对话，
58:27
the intuitions, what was executed in the lab
那些直觉，实验室里执行了什么，
58:30
or the computations run, what was the code written,
或者跑了哪些 computations，写了什么 code，
58:32
stitching all this together is so valuable.
把这一切拼在一起，非常有价值。
58:35
And I think the overall goal is
而且我觉得，整体目标是
58:38
rather than training on the final output of science,
与其在科学的最终 output 上做 training，
58:41
you're training on the process of doing science.
不如在做科学的过程上做 training。
58:44
This feels very much like what you would try to do with RSI,
这感觉非常像你在 RSI 上会尝试做的事情，
58:47
where you are having a model train to train better models.
也就是让一个 model 去训练，从而训练出更好的 models。
58:52
You're now having a model train to make better experiments.
你现在是让一个 model 去训练，从而做出更好的实验。
58:56
But the difference is it's not going back
但区别在于，它不会回到
58:58
into the core model for, so it's
core model 里去，所以它
59:00
unless the physics models are developing
除非 physics models 正在发展、
59:02
are improving chips which then improve it.
在改进 chips，而 chips 又会反过来改进它。
59:05
And this is the wider, yeah, higher level RSI.
而这是更广泛的，对，更高层次的 RSI。
59:09
That's the data flywheel.
这就是 data flywheel。
59:10
But does that mean that is it plausible
但这是不是意味着，有没有可能
59:13
that in, you know, Fable 6 or something or GVT8
就是说，你懂的，在 Fable 6 之类的，或者 GVT8 里
59:19
could just have a, which has been tuned
就能直接有一个，已经被 tuned
59:22
on these sort of higher order reasoning traces,
在这些更高阶的 reasoning traces 上，
59:26
could actually just do this without any logic.
实际上就能不靠任何逻辑做到这件事。
59:30
Because it's sort of the same thinking process, right?
因为这差不多是同一种思考过程，对吧？
59:33
Yeah, I mean, I think there's decision making
对，我是说，我觉得这里面有 decision making
59:36
under uncertainty, obviously in machine learning,
在不确定性下，显然在 machine learning 里，
59:38
there's noise in running like the AI loops.
运行时会有 noise，就像 AI loops 那样。
59:42
But we do think there's like a different set of challenges
但我们确实觉得，这像是一组不同的挑战
59:44
when you're actually interfacing with the physical world.
当你真的在和物理世界交互的时候。
59:46
But also I think there's another piece too,
但另外我也觉得还有另一块，
59:48
which is getting the compression of everything
也就是把一切的 compression
59:51
into weights is still really valuable.
放进 weights 里，这件事仍然非常有价值。
59:54
If inference time reasoning was sufficient,
如果 inference time reasoning 就已经足够了，
59:57
all of the frontier labs would have stopped training
所有 frontier labs 本来都会停止 training
59:59
at like GPD4 and were like, okay, from now on out,
大概在 GPD4 的时候，然后就会说，好吧，从现在开始，
60:02
we're going to get really good at inference time improvements.
我们会把 inference time improvements 做得特别好。
60:06
And so we think that by, you know,
所以我们觉得，通过，你知道，
60:09
because of the differences between, you know,
因为，你知道，这两者之间的差异，
60:11
physical sciences and machine learning,
物理科学和 machine learning，
60:13
getting that compressed into our own weights
把这些压缩进我们自己的 weights 里
60:16
will lead to like different types of systems
会导致像是不同类型的系统
60:18
and different types of capabilities.
以及不同类型的能力。
60:20
But also, we feel like even if like Fable 7
但另外，我们也觉得，即使像 Fable 7
60:23
gets really good at, they're given better than what it is today,
变得非常擅长，它们得到的也会比今天更好，
60:27
it will still have to run experiments to get results.
它仍然得跑 experiments 才能得到结果。
60:30
And the reason for the right machine learning
而原因就在于，对吧，machine learning
60:32
is really good at what's been trained on.
非常擅长它被训练过的内容。
60:34
But scientific discovery is almost by definition
但 scientific discovery 几乎从定义上就是
60:36
what you haven't been trained on.
你没有被训练过的东西。
60:38
And that's why we're building these labs
这就是为什么我们在建这些实验室
60:40
so that whether open models or closed models
这样不管是 open models 还是 closed models
60:44
can use these labs to tinker with the universe
都能用这些实验室去折腾宇宙
60:46
because we don't feel like you can make a big discovery
因为我们不觉得你能搞出什么重大发现
60:48
without trying things.
要是不去试的话。
60:50
Yeah, there's not going to be like,
对，不会有什么，比如说，
60:52
no one's going to like zero shot,
没人会喜欢 zero shot，
60:53
the reattemptive production.
那种 reattemptive production。
60:54
I think that would be pretty cool.
我觉得那会挺酷的。
60:56
Yeah, yeah, I mean, as someone who was more closely
对，对，我是说，作为一个以前跟
60:59
with little wet labs before,
小型 wet labs 更密切接触过的人，
61:00
it's certainly, I am more skeptical about zero
当然，我对 zero shotting
61:03
shotting like scientific results than some.
科学结果这种事，会比一些人更怀疑。
61:05
But it is something that I think some people might have.
但我觉得这可能是一些人会有的东西。
61:08
So yeah, I mean, I think it's really important to distinguish
所以，对，我是说，我觉得区分这一点真的很重要，
61:12
the results we see in math and theoretical physics
我们在数学和理论物理里看到的结果
61:15
from the physical world.
来自物理世界。
61:17
Yeah, it's really two different things.
是啊，这真的是两码事。
61:19
Two very different things.
非常不同的两码事。
61:20
I mean, even theoretical physics,
我是说，甚至 theoretical physics，
61:21
like so far it's been theoretical computer science
就像到目前为止，一直是 theoretical computer science
61:22
building and math.
building 和 math。
61:23
Right, yeah.
对，是啊。
61:25
Maybe theoretical physics is next.
也许下一个就是 theoretical physics。
61:26
Yeah, yeah, yeah, yeah, yeah.
对，对，对，对，对。
61:29
Can I get us a sort of mental model
我能不能给我们弄一个大概的 mental model
61:31
of the levels of extraction that you can go?
关于你能达到哪些层级的 extraction？
61:34
So for example, in terms of automation,
所以比如说，在 automation 方面，
61:36
and I'm just on this theme again,
而且我又在讲这个主题了，
61:38
where we talked about the campaign,
就是我们聊到那个 campaign 的时候，
61:40
talked about individual essays and running tests
聊到单篇 essay 和跑测试，
61:43
and how their humans are bad at it.
以及他们的人类在这方面有多不擅长。
61:45
So we should stop humans from doing it.
所以我们应该阻止人类去做这件事。
61:46
Are there others?
还有别的吗？
61:48
So for example, one level higher than campaign
比如说，比 campaign 高一层，
61:50
could be a physics theory that you're testing
可能是你正在验证的一个物理理论，
61:55
or one level lower than campaign is what?
或者比 campaign 低一层的是什么？
61:59
And I like to think about it from that point of view,
而我喜欢从那个角度去想，
62:02
and then think about it from like,
然后再从类似这样的角度去想，
62:04
okay, well, this is an API call now.
好吧，那现在这就是一个 API call 了。
62:06
Like you never have to touch this again.
就像你以后再也不用碰这个了。
62:07
And this one is, no, this is still 90% human.
而这个呢，不，这个还是有 90% 是人类。
62:11
And maybe we can sort of draw the map
也许我们可以大致把地图画出来
62:12
of the territory that way.
用这种方式来标出这片领域。
62:15
Is there other levels?
还有别的层级吗？
62:17
I can tell you some levels.
我可以给你讲几个层级。
62:18
I feel like you asked a good question.
我觉得你问了一个好问题。
62:19
I don't have a very systematic answer,
我没有一个非常系统的答案，
62:21
but let's talk about some levels.
不过我们来聊聊几个层次。
62:22
So one level is the atomistic structure.
其中一个层次是 atomistic structure。
62:25
There's an abstraction, right?
这是一种抽象，对吧？
62:26
Like there is no perfect atomistic structure,
就是说不存在完美的 atomistic structure，
62:28
anything we do, but it's one approximation.
不管我们怎么做，它都只是一种近似。
62:31
Another one is the continue model.
另一个是 continue model。
62:33
So this one is like not atomistic anymore,
这个就不再是 atomistic 的了，
62:35
but it's a continuous mesh representing the material.
而是一个表示材料的 continuous mesh。
62:40
And then in a different dimension,
然后，在另一个维度上，
62:42
there's the thermodynamics.
还有 thermodynamics。
62:43
Like assuming you do this experiment forever,
就好比假设你把这个实验永远做下去，
62:46
what would be the final state?
最终会达到什么 final state？
62:48
And then there's another layer of abstraction,
然后还有另一层 abstraction，
62:50
which is kinetics.
也就是 kinetics。
62:51
Acknowledging that we're not doing this experiment forever.
要承认我们不会永远做这个实验。
62:53
So time matters.
所以时间就很重要。
62:55
So how quickly will that reaction happen or not,
所以那个反应会有多快发生，还是不会发生，
62:57
even if it's lower energy or not?
即使它的能量更低，或者并不更低？
62:59
So thermodynamics, kinetics, atomistic continuum,
所以 thermodynamics、kinetics、atomistic continuum，
63:02
what else, are there other levels of abstraction?
还有什么，有没有其他 levels of abstraction？
63:04
Well, I mean, I think like you're a point about like,
嗯，我的意思是，我觉得你是在说一个点，就像，
63:07
okay, there could be an overarching new theoretical advance
好吧，可能会有一个总体性的新 theoretical advance
63:10
that informs multiple campaigns, right?
能影响多个 campaigns，对吧？
63:14
It's kind of like as an investor,
这有点像，作为一个投资人，
63:16
I want to go like, here's the bottleneck, guys.
我想这么说，瓶颈就在这儿，各位。
63:19
We get everything else, I don't know.
其他的我们都能搞定，我也不知道。
63:22
I mean, from the beginning,
我是说，从一开始，
63:24
our hypothesis been that one of the big bottlenecks
我们的假设一直是，其中一个很大的瓶颈
63:26
is automatic characterization
就是 automatic characterization，
63:27
because it's not that hard to mix powders
因为把粉末混在一起并不难
63:30
to get to try stuff.
然后就能去试各种东西。
63:32
But if you can't characterize and analyze it
但如果你没法 characterize 和 analyze 它
63:34
and then decide what the next step should be intelligently,
然后明智地决定下一步该做什么，
63:37
you don't really benefit much from mixing powders randomly.
随机混合粉末其实没多大好处。
63:40
So we focused a lot on automatic characterization,
所以我们把很多精力放在 automatic characterization 上，
63:43
closing the loop with simulations,
通过 simulations 来 close the loop，
63:46
because that's how you like try a lot of things,
因为这样你才能尝试很多不同的东西，
63:48
make an informed decision
做出有依据的决定，
63:49
and then decide what to do the next day.
然后决定第二天该做什么。
63:51
Maybe there's a follow-up question.
也许还会有一个后续问题。
63:53
What is the, how do humans live in this loop?
这个，人类到底是怎么在这个 loop 里生活的？
63:57
Like, in what points do you put, like, drop the human in?
比如说，你会在哪些点把人放进去，或者说，把人丢进去？
64:00
I mean, I guess you could randomly drop in any part
我是说，我猜你可以随便塞到任何一个部分里
64:04
and then value us.
然后评估我们的价值。
64:05
But like, where do you decide to send your time?
但比如说，你会决定把时间花在哪儿？
64:08
And it seems like it might be all levels.
而且看起来这可能是所有层面上的事情。
64:09
Like you have, you have people in the lab
就像你有，你有实验室里的人
64:12
who are physically moving materials
他们在亲手搬材料
64:14
and then you also have scientists
然后你还有科学家
64:15
who are guiding decisions about what campaign
他们在指导关于做什么 campaign 的决策
64:17
and then you have humans who are,
然后你还有人类，他们……
64:20
what's that?
那是什么？
64:20
I think it's a question for me.
我觉得这是在问我。
64:21
100%, I think it's always shifting too.
100%，我觉得它也一直在变。
64:24
So I mean, I think even just using the guidance of campaigns
所以我的意思是，我觉得哪怕只是用 campaign 的指导
64:27
in early days, that was fully human-driven.
在早期，那也完全是由人驱动的。
64:29
Now increasingly, it's a mix of AI-driven and human-driven.
现在，它越来越是 AI-driven 和人工驱动的混合。
64:34
And then, Derek was saying early on,
然后，Derek 早先就说过，
64:36
we identified that characterization was a big bottleneck
我们发现 characterization 是一个很大的瓶颈，
64:38
in the early days.
在早期的时候。
64:39
It was human-driven.
那时候是人工驱动的。
64:41
And the scientists were very much overwhelmed
而科学家们当时真的不堪重负，
64:43
and the balance has shifted significantly towards AI-driven.
而这个平衡已经大幅转向了 AI-driven。
64:47
And this has allowed the scientists
这也让科学家们能够
64:49
who are previously spending all of their time
那些以前把所有时间都花在
64:51
doing these refinements to now elevate their work
做这些 refinements 上的人，现在把他们的工作提升
64:54
to something else.
到别的东西上。
64:56
But I mean, I think we're, it's really sort of a function
但我是说，我觉得我们，这其实真的有点像是
65:00
of time.
时间的函数。
65:01
It's always changing at each of these levels.
在每个层级上，它一直都在变。
65:04
Yeah, in terms of just general search,
嗯，就 general search 来说，
65:06
are there things that perform well in, let's say,
有没有什么东西在，比如说，
65:09
the physical world that don't perform well
在物理世界里表现不太好的那些
65:12
in the other worlds or vice versa?
在其他世界里，或者反过来？
65:14
So evolutionary search is pretty popular in LLMs.
所以 evolutionary search 在 LLMs 里挺流行的。
65:18
I don't imagine it works well here.
我觉得它在这里不太行。
65:21
So especially in simulations, people
所以尤其是在 simulations 里，人们
65:23
have been using evolutionary search for a long time.
已经用 evolutionary search 很久了。
65:27
That is pretty good at structure prediction, for example,
它在 structure prediction 上挺擅长的，比如，
65:30
finding low energy structures.
找到 low energy structures。
65:32
OK, that's what works.
好，这就是可行的办法。
65:33
And then the process engineers in the semiconductor industry
然后半导体行业的工艺工程师
65:36
will do DOE design or experiments.
会做 DOE 设计或实验。
65:38
And they'll often use some kind of zero-thorder,
他们通常会使用某种 zero-thorder，
65:41
like with its patient optimization or evolutionary search,
比如 patient optimization 或 evolutionary search，
65:44
I don't think they use RL, but similar idea, yeah.
我不觉得他们会用 RL，但思路类似，对。
65:47
You have to have some kind of algorithm.
你得有某种 algorithm。
65:48
Yeah, it's a figure.
对，那是一张图。
65:49
And all zero-thorder optimization
而且所有 zero-thorder optimization 都属于同一个敌人，对吧？
65:51
be of the same enemy, right?
对。
65:53
Yeah.
好，我想回到我之前问过的一个问题，
65:54
OK, I want to go back to a question I had earlier,
那是关于 scaling 的一个问题。
65:57
which was one of scaling.
所以我很想知道，你对 scaling 这个实验室的愿景是什么？
65:58
And so I'm curious, what is your vision for scaling the lab?
所以有不同方式。
66:02
So there's different ways.
我是从生物学领域过来的。
66:04
I'm coming from the biology world.
66:07
There's a lot of cool tricks you can do in biology.
在生物学里，你可以玩很多很酷的 tricks。
66:09
You can tag things.
你可以给东西打 tag。
66:12
You can use DNA sequencing for all sorts of interesting readouts.
你可以用 DNA sequencing 做各种有意思的 readouts。
66:16
And if you can map your complicated assay onto sequencing,
而且如果你能把你复杂的 assay map 到 sequencing 上，
66:19
you can now scale to millions or hundreds of millions
你现在就能 scale 到几百万或者几亿
66:22
or something.
之类的。
66:23
Can you do tricks like this?
你能玩这种 tricks 吗？
66:24
Like, ultimately, every assay has its own runtime.
就，说到底，每个 assay 都有自己的 runtime。
66:28
I'll have a favorite blog post.
我会有一篇最喜欢的 blog post。
66:30
I would want to quote in the show notes.
我会想在 show notes 里引用它。
66:32
But what does the run times look like for these labs?
但这些 lab 的 run times 看起来怎么样？
66:36
Is your scaling for materials just fundamentally linear
你们对 materials 的 scaling 是不是从根本上就是线性的？
66:41
in terms of you just need more resources to do more things?
也就是说，你只需要更多资源就能做更多事情？
66:46
Or can you paralyze things in clever ways
还是说你可以用一些聪明的方式把东西 parallelize？
66:49
and combine things and are there tricks there?
然后把东西组合起来，这里面有什么技巧吗？
66:52
Well, I think one way to think about it
嗯，我觉得思考这个问题的一种方式是
66:54
is the RL environment is not quite trivily run the experiment,
就是这个 RL environment 并不是那么随便就能跑这个 experiment，
67:00
wait for that agent to fully roll out
要等那个 agent 完整地 roll out
67:03
doing the calculations, going through all the instruments
把计算都做完，把所有 instruments 都走一遍
67:05
to the final outcome.
一直到最终结果。
67:07
We basically start producing data across all the different
我们基本上就开始在所有不同的
67:11
instruments and all these different campaigns.
instruments 和所有这些不同的 campaigns 上产生数据。
67:13
And one way you can begin to expand that for machine learning
而你要开始把这一套扩展到 machine learning 的一个方式，
67:16
is now reinforcement learning environments
就是 reinforcement learning environments 了
67:19
can be constructed based on, again, like,
可以基于，呃，就像，
67:22
dates for the experimental campaigns
实验活动的日期
67:24
or computational campaigns at that point.
或者那个时候的计算活动。
67:26
Or you can also start taking subsets of instruments.
或者你也可以开始取一部分仪器。
67:30
And you're like, OK, we're going to create reinforcement learning
然后你会说，OK，我们要创建 reinforcement learning
67:33
environments on the basis of this instrument alone.
environments，仅基于这个仪器。
67:36
Or maybe it's across these different instruments
或者也许是跨这些不同的仪器
67:38
to kind of get to some reward state.
来差不多达到某种 reward state。
67:42
So that's a way where this finite basket of data
所以这是一种方式，这有限的一篮子 data
67:45
when viewed in different ways can then
从不同角度去看时，就能
67:48
be expanded for training purposes.
为了 training 的目的被扩展出来。
67:51
So it doesn't get at the actual experiment run
所以它并没有触及实际的 experiment run
67:54
but from ML perspective.
但从 ML 的角度来看。
67:55
Maybe I can give you some analogues for the bio examples you gave.
也许我可以给你一些类比，来对应你刚给的那些 bio 例子。
67:58
So one thing people have tried is combinatorial sputtering.
所以人们尝试过的一件事就是 combinatorial sputtering。
68:01
And if you ever heard of this, that you
而如果你听说过这个，那你
68:03
have these sputtering targets and you
有这些 sputtering targets，然后你
68:06
create this compositional gradient.
就制造出这个 compositional gradient。
68:08
So when you look at the final product
所以当你看最终产品的时候
68:10
because there is gradient from each precursor,
因为每个 precursor 都会形成 gradient，
68:13
it creates different crystals at different spatial locations.
它会在不同的 spatial locations 生成不同的 crystals。
68:16
So in one go, you may be try hundreds, thousands of crystals.
所以一次就能试几百、几千个 crystals。
68:21
This is an example to your example.
这是给你那个例子举的一个例子。
68:24
And as far as I know, this isn't work very well so far,
而且据我所知，到目前为止这个还不太行，
68:26
because it turns out like there is diffusivity
因为结果发现，好像存在 diffusivity
68:29
so things move around.
所以东西会动来动去。
68:31
Another one that we thought about where we haven't done yet,
另一个我们想过、但还没做的，
68:33
but a lot of people talk about this
但很多人在聊这个
68:34
is superconductivity measurements are bottleneck
就是 superconductivity measurements 是个 bottleneck
68:37
because unlike XRD there's no high throughput
因为不像 XRD，没有 high throughput 的
68:39
superconductivity measurement
superconductivity measurement
68:40
and it can take like one hour per measurement.
而且一次测量可能要花一个小时左右。
68:42
So people thought about taking the different candidates,
所以人们想过把不同的候选材料，
68:45
mixing them all into one sample and putting it in.
把它们全混到一个 sample 里，然后放进去。
68:48
And if it has superconductivity, you know
如果它表现出 superconductivity，你就知道
68:51
that one of the sources that you can do
其中一个来源，你可以
68:52
a bit like how people are doing COVID testing too.
做得也有点像人们做 COVID 检测那样。
68:55
So their ideas like this exist.
所以这类想法是存在的。
68:57
Some of them work well, some of them don't.
有些效果很好，有些不行。
69:00
So you are I think are about to announce a big fundraise.
所以我觉得你快要宣布一笔大融资了。
69:03
How are you thinking about scaling?
你怎么考虑 scaling 的？
69:06
Like we've been doing our design of our labs for a while
就是我们做实验室设计已经有一阵子了
69:09
and I think the resources allow us to continue
我觉得这些资源能让我们继续
69:11
to scale up the labs significantly
大幅 scale up 实验室
69:14
but also the compute both from like the AI side
而且还有 compute，比如从 AI 这一侧
69:16
and the computational side.
以及 computational 这一侧。
69:18
But I think really importantly too,
但我觉得同样非常重要的是，
69:19
it's what are the new types of labs we're able to build.
关键是我们能建出哪些新类型的实验室。
69:22
So I think we've been proving out this loop
所以我觉得我们一直在验证这个 loop
69:26
in our initial labs here in Menlo Park,
在我们 Menlo Park 这里最初的实验室里，
69:28
but we'll be continuing to expand to new types of labs
但我们会继续扩展到新型的实验室
69:31
and repeat the same process.
并重复同样的流程。
69:33
Yeah, like the kind of lab we're trying to build,
对，就像我们正在尝试构建的那种实验室，
69:35
I think haven't been built before at this scale
我觉得以前还没有人以这种规模建过
69:36
or at this approach.
或者以这种做法建过。
69:38
So we have learned a lot from our own build up so far
所以到目前为止，我们从自己的搭建过程中学到了很多
69:41
and those lessons guide us for the next version
这些经验会指导我们做下一个版本
69:43
and the next version.
以及再下一个版本。
69:44
And then we can as you said, like increase the scale,
然后，就像你说的，我们可以扩大 scale，
69:48
the ambition, the quality of instruments,
提升野心，提高工具的质量，
69:51
some instruments are very expensive.
有些工具非常贵。
69:53
As we get a really good understand for what
等我们真正搞清楚什么
69:55
we get a lot of return from,
能给我们带来很多回报，
69:56
we can invest into those more.
我们就可以在这些方面投入更多。
69:58
And it makes hardware engineering
而且这让 hardware engineering
70:00
has become so core to it.
已经变得如此核心。
70:01
So for example, in the early days just for speed,
所以举个例子，早期只是为了速度，
70:03
we buy some off-to-shelf instruments.
我们会买一些现成的 instruments。
70:06
And as we kind of push the scale of the lab,
而随着我们不断把实验室的规模往上推，
70:08
you make your own.
你就自己做。
70:09
We have to make our own.
我们必须自己做。
70:10
So we realize, okay, for this instrument,
所以我们意识到，好吧，对于这个 instrument，
70:13
it's pretty fast.
它挺快的。
70:14
These components are actually pretty quick
这些 components 其实都挺快的
70:15
but this weighing machine actually becomes the bottleneck
但这个 weighing machine 反而成了 bottleneck
70:19
or there can be things pertinent to data quality
或者可能有一些跟 data quality 有关的问题
70:21
where well, the resting position of this robotic arm
就是，嗯，这个 robotic arm 的 resting position
70:24
actually is above the plate of previously mixed things.
其实是在之前混合过东西的那个 plate 上方。
70:28
So there's a contamination risk.
所以就有 contamination risk。
70:29
So in our hardware design, we're going to redesign it
所以在我们的 hardware design 里，我们准备把它重新设计一下
70:33
so that the resting position lies away from those things.
这样 resting position 就会远离那些东西。
70:37
These are the kind of subtle details
这些就是那种很微妙的细节
70:38
that allow us to kind of push the noise
让我们能够某种程度上把 noise 往下压
70:40
for down for our experimental campaign.
为我们的 experimental campaign 把 noise 压下去。
70:42
So basically take these learnings and scale it up.
所以基本上就是把这些学到的东西拿过来，然后 scale up。
70:46
You know, you ship your org chart.
你知道，你交付的其实就是你的 org chart。
70:48
You have many teams.
你有很多团队。
70:49
I think that we were surprised when we looked
我觉得，我们当时看的时候挺惊讶的
70:51
at your JAWS page and we were like,
在你的 JAWS 页面上，我们当时就在想，
70:53
okay, we don't think we fully map what periodic does.
好吧，我们觉得没完全搞清楚 periodic 到底是做什么的。
70:56
Maybe some organizing principles
也许可以有一些组织原则，
70:58
so to kind of make sense of the job page.
这样就能大概看明白这个招聘页面了。
71:01
So we hire for AI research and infrastructure,
我们招的是 AI research 和 infrastructure、
71:04
computational, experimental roles
computational、experimental 岗位，
71:07
and then hardware engineering roles
然后还有 hardware engineering 岗位，
71:08
and then product roles.
再然后还有产品岗位。
71:10
We want to achieve since this super intelligence, right?
我们想要实现的，因为这是 super intelligence，对吧？
71:12
And we feel like for that,
而且我们觉得为了那个目标，
71:14
we need the right chemistry expertise,
我们需要对的 chemistry 专业知识，
71:16
the right physics expertise, simulation theory
对的 physics 专业知识、simulation theory
71:19
and tin films, powder.
还有 tin films、powder。
71:21
So hardware engineering,
所以 hardware engineering，
71:23
but these seem like very diverse roles which they are,
但这些看起来是非常多元的角色，它们确实是，
71:25
but they're all actually coherent
但它们其实都是连贯的
71:28
in what they're trying to do,
在他们想做的事情上，
71:29
which is a sense of super intelligence.
也就是一种 super intelligence 的感觉。
71:30
And it also applies to the LLM researchers,
而且这也适用于 LLM 研究人员，
71:33
the infra, even the product engineers
infra，甚至产品工程师
71:35
because the product engineers kind of make sure
因为产品工程师某种程度上要确保
71:38
all the research gets made into a product
所有研究都能被做成产品
71:40
that the experimental center lab can use.
给 experimental center lab 用。
71:42
Yeah, so in order to do the end-to-end loop
对，所以为了跑通这个 end-to-end loop
71:44
with the physical world,
和物理世界打交道时，
71:45
it really requires your ability to have agency
这真的需要你有 agency
71:48
over the physical world to build these labs
能对物理世界施加影响，去建这些实验室，
71:50
to run the campaigns effectively,
有效地跑这些 campaigns，
71:52
to create AI systems against that,
去构建针对这一点的 AI 系统，
71:54
to make them good users of the computational tools.
让它们成为 computational tools 的优秀用户。
71:57
That's like the difficulty,
这就像是难点，
71:59
but also the opportunity of periodic.
但也是 periodic 的机会。
72:01
It's like this group of people
就像这群人
72:03
has just never been brought together
以前从来就没被聚到一起过
72:04
before it's irreducibly a multi-disciplinary problem.
这件事本质上就是一个 multi-disciplinary 问题。
72:07
Yeah, and one other guiding principle for us
对，而且对我们来说还有另一个指导原则
72:10
that is really important for us
这个原则对我们真的非常重要
72:11
is we want the people who are very good at what they do
就是我们希望那些非常擅长自己做事情的人
72:14
and experience to do it hands-on.
以及有经验的人，能亲自动手去做。
72:17
You know, like the modern life has gotten us
你知道，就像现代生活把我们变成现在这样了
72:18
to this place where, especially in academia,
到这样一个地步：尤其是在学术界，
72:21
when someone is really good at research in some area,
当某人在某个领域做研究特别厉害时，
72:24
we give them so much responsibility
我们会给他们那么多责任，
72:25
for grant writing, teaching,
负责 grant writing、教学，
72:27
that they stop having the time
以至于他们不再有时间
72:28
to do research in that area, hands-on.
去在那个领域亲手做研究。
72:31
But if you look back at like Bell Labs, IBM,
但如果你回头看 Bell Labs、IBM，
72:34
institutions that made really good progress,
那些取得过很大进展的机构，
72:37
it was really experienced people doing hands-on work,
真的都是经验丰富的人在亲手做具体工作，
72:39
like Bardeen was in the lab every day
比如 Bardeen 每天都在实验室里，
72:41
when though he's the theorist,
尽管他是理论家，
72:43
Alex Mueller was in the lab doing experiments
Alex Mueller 也在实验室里做实验，
72:45
even though he was the lab lead.
尽管他是实验室负责人。
72:46
So we try to do that here too.
所以我们在这里也试着这样做。
72:47
So we have some of the world's leaders in different fields,
所以我们这里有一些不同领域的世界顶尖人物，
72:50
but they're doing hands-on work.
但他们也在亲手做具体工作。
72:52
Do you want to brag or just call them out?
你是想炫耀一下，还是就想点名批评他们？
72:54
Oh, I would love to, yeah.
哦，我太愿意了，是啊。
72:55
So, you know, when I was doing my PhD,
所以，你知道，我读 PhD 的时候，
72:58
my favorite computational material scientist,
我最喜欢的 computational material scientist，
73:00
who's kind of from my age group is Morata and I call.
他跟我差不多算是同龄人，叫 Morata and I call。
73:04
He did his PhD with one of our advisors, Chris Wolverton.
他跟我们的其中一位导师 Chris Wolverton 读的 PhD。
73:07
And he is kind of like our computational
而且他差不多算是我们的 computational
73:10
material science expert, he's incredible.
material science expert，他特别厉害。
73:12
He's sometimes, you know, a real good experimentalist
他有时候，你知道，真的是个很棒的 experimentalist
73:14
even though he does simulations
尽管他是做 simulations 的
73:16
like he understands systems really well.
就是他真的很懂 systems
73:18
Our lab lead is Joe Czekowski,
我们实验室的负责人是 Joe Czekowski，
73:21
who is a professor at MIT,
他是 MIT 的教授，
73:22
but he's only lived to work with us full time.
但他只是全职和我们一起工作。
73:25
And Joe is this incredible physicist,
而 Joe 是个特别厉害的物理学家，
73:27
understands superconductivity really well,
对 superconductivity 非常了解，
73:28
but he also was a professor in Japan for a bit.
但他也在 Japan 当了一阵子教授。
73:31
So I learned synthesis and chemistry really well,
所以我把 synthesis 和 chemistry 学得非常好，
73:35
especially for a physicist.
尤其对一个物理学家来说。
73:36
We have Daniel Chica, who was a graduate student
我们有 Daniel Chica，他曾经是研究生，
73:39
in Mercury-Canada-Cities Group,
在 Mercury-Canada-Cities Group 里，
73:41
which is maybe the best-salt-state chemistry group
那可能是最好的 salt-state chemistry group，
73:43
in the world.
在世界上。
73:44
And Daniel was one of his,
而 Daniel 是他其中的一个，
73:45
like, synthesis experts.
比如，synthesis 专家。
73:46
So he's like a magician with his fingers.
所以他就像个手指特别灵活的魔术师。
73:49
I mean, there's so many examples like this.
我是说，这样的例子太多了。
73:50
No, I mean, I think like Dima Bhandao,
不，我是说，我觉得像 Dima Bhandao，
73:53
so he's been leading a lot of our AI and LLM efforts.
所以他一直在主导我们很多 AI 和 LLM 方面的工作。
73:56
He was the inventor of neural network attention.
他是 neural network attention 的发明者。
73:59
Oh, butto know, that butto know session.
哦，butto know，那个 butto know session。
74:01
Exactly, yeah, yeah, yeah.
没错，对，对，对。
74:03
Yeah, so very hands on in sort of getting into, like,
对，所以他特别亲力亲为，会亲自去钻研那些，比如
74:07
the traces, the training distribution,
traces、training distribution，
74:09
how to connect it to the lab.
怎么把它跟实验室连接起来。
74:10
The chemistry, he'll go deep.
chemistry 这块，他会钻得很深。
74:12
He was very deep.
他真的很深入。
74:14
He actually has, like, solid states synthesis books,
他居然还有，像是，solid states synthesis 的书，
74:16
like Atta's lab in Montreal, Ray Nakano.
比如 Montreal 的 Atta 实验室，Ray Nakano。
74:19
So he's tech lead, some of the operator work at OpenAI.
所以他是 tech lead，在 OpenAI 做一部分 operator 的工作。
74:24
He's in the lab.
他在实验室里。
74:25
So when we were doing some lab tours,
所以之前我们做那些实验室参观的时候，
74:27
we were initially like, oh, this new scientist,
我们一开始还觉得，哦，这个新来的科学家，
74:30
this new technician really looks like Ray.
这个新来的技师真的很像 Ray。
74:31
And no, it was literally, this was Ray,
但并不是，真的就是，这就是 Ray 本人，
74:34
like, actually in the lab with the scientists
就，真的在实验室里和科学家们一起
74:37
automating the machines.
把机器自动化。
74:38
And it's like, this is sort of the DNA
然后就像是，这算是某种 DNA
74:41
and the type of people we need to kind of pull this off.
以及我们需要的那种能把这件事搞定的人。
74:44
So we also saw that you had four deploy engineering roles.
所以我们也看到，你们有 forward deploy engineering 的岗位。
74:47
We happened to be spinning up a four deploy engineering
我们刚好在做一档 forward deploy engineering 的
74:49
podcast because there's so many engineers
播客，因为有很多工程师
74:52
that would do that role.
会去做那个岗位。
74:54
They may not necessarily know that there even
他们可能不一定知道，甚至
74:56
is a role for them at Periodic.
在 Periodic 有适合他们的岗位。
74:58
What is it?
那是什么？
74:59
And how do they work your customers?
那它们是怎么服务你们客户的？
75:01
So we've been building our own products,
所以我们一直在打造自己的产品，
75:04
our own tools, our own AI and computational for ourselves.
自己的工具，自己的 AI 和 computational，都是为我们自己服务的。
75:07
So we've been customer zero.
所以我们一直是 customer zero。
75:09
Now what we're doing is we're taking these same tools
现在我们正在做的，就是把这些同样的工具
75:11
to different industries.
带到不同的行业。
75:12
And we have a huge focus right now
而且我们现在有一个很大的重点
75:14
in the semiconductor industry.
是在 semiconductor 行业。
75:17
However, in order to do this, these are very private companies
不过，要做到这一点，这些公司都非常私密，
75:22
we need to be able to operate within some of the most
我们需要能够在一些最
75:24
secure environments.
安全的环境中运作。
75:26
And so our Periodic for Deployed Engineers and Researchers
所以我们的 Periodic for Deployed Engineers and Researchers
75:29
will actually be on site working and actually
实际上会到现场工作，并且真正
75:32
integrating our AI systems or computational systems
整合我们的 AI systems 或 computational systems，
75:35
to help our partners get to their goal
帮助我们的合作伙伴更快达到他们的
75:37
states more quickly.
goal states。
75:39
So basically, the tools, the know how
所以基本上，这些工具、这些 know-how
75:41
that we've built in doing our own materials discovery,
都是我们在做自己的 materials discovery、
75:44
materials engineering, we're now bringing to industry.
materials engineering 时积累出来的，现在我们把这些带到产业里。
75:47
And we think this is the most effective way
而且我们认为，这是最有效的方式
75:50
to get partners to their end states and to their goals
来让合作伙伴达到他们的最终状态、实现他们的目标
75:54
rather than just like throwing some technology
而不是就那样把某个技术
75:55
over the wall and telling them to figure it out.
隔着墙扔过去，然后让他们自己想办法搞定。
75:58
Then these four deployed engineers will also
然后这四名 deployed engineers 也会
76:00
do the inference locally, but then also
在本地做 inference，但之后也
76:03
can train on the data.
可以在 data 上 train。
76:05
So again, once we take our system that
所以再说一遍，一旦我们把我们那个能
76:07
understands these different areas and we deploy,
理解这些不同领域的系统部署下去，
76:10
we can make an expert on the customers and the partners data
我们就能基于客户和合作伙伴的 data 做一个专家，
76:13
so that they can own their own intelligence
这样他们就能拥有自己的 intelligence，
76:15
and have systems that understand this much more effectively
并拥有能更有效理解这些的系统，
76:18
than just like hitting some API or some untrained model.
而不是只是去调某个 API 或用一个没 train 过的 model。
76:24
And so this is a much expanded four deployment engineering
所以这个角色比典型的 deployment engineering 职位扩展得多，因为它真的需要非常深的 machine learning 专业知识。
76:29
role than typical because this really requires
你必须极其精确——从 data training 的角度、infrastructure 的角度，还有 physics 的角度都是如此。
76:31
some very deep machine learning expertise.
所以我们是在跟技术含量非常高的行业打交道。
76:35
You have to be incredibly precise from like a data
他们必须理解比如 chemistry 的不同领域。
76:38
training perspective, infrastructure perspective,
76:40
but also a physics perspective.
76:41
So we're working with highly technical industries.
76:45
They have to understand different areas of like chemistry
76:48
is material science, some of the devices.
是 material science，还有一些设备。
76:50
And so those are the roles we're building out right now.
所以这些就是我们目前正在组建的岗位。
76:53
Part of it is willingness to spend extended periods
其中一部分是愿不愿意长期
76:55
on site in Taiwan.
在 Taiwan 现场驻点。
76:57
OK, well, I know that kind of customer.
OK，嗯，我了解那种客户。
77:01
It's interesting that your monetization,
有意思的是，你们的变现，
77:03
that is the primary way that you're monetizing now.
这是你们现在主要的变现方式。
77:07
But it might change in a future.
但未来可能会变。
77:09
But like, you know, that's like a model
但就像，你知道，那就像一个模式
77:12
that I think people don't really get about a pure research lab
我觉得人们并没有真正理解纯研究实验室的这一点
77:16
because when you compare yourself to a like a Bell Labs,
因为当你把自己跟像 Bell Labs 这样的地方比时，
77:17
that's what people are thinking, right?
大家想的就是那个，对吧？
77:18
Which is, I think like a really great analog
我觉得，一个特别好的类比
77:20
would be software engineering.
就是 software engineering。
77:22
So in the early days of software engineering,
所以在 software engineering 的早期，
77:24
we have like, you know, GitHub copilot,
我们就有，就像，你知道，GitHub copilot，
77:26
then we had early versions of chat GPT.
然后我们有了早期版本的 chat GPT。
77:28
People were using these things as copilot
大家把这些东西当成 copilot 来用，
77:30
to help them get to their solutions.
帮自己找到解决方案。
77:32
And as the automation improved, now we
而随着自动化程度提高，现在我们
77:35
have things like codex and very few of our engineers
有了 codex 这类东西，而且我们的工程师里很少有人
77:37
are writing code the way they used to.
还像以前那样写代码了。
77:39
And we think a very similar thing could
而且我们认为，非常类似的事情可能
77:41
play out for periodic as well, where you
也会在 periodic 上重演，而你会
77:43
can have systems to accelerate the researchers,
可以有系统来加速研究人员，
77:46
the material scientists, materials engineers,
材料科学家、材料工程师，
77:47
the process engineers.
工艺工程师。
77:49
But as the autonomy, intelligence, and capability
但随着 autonomy、intelligence 和 capability
77:52
grows, you can begin to price outcomes.
不断增长，你就可以开始按结果定价。
77:56
So help me get to this kind of goal state.
所以帮我达到这样一种 goal state。
78:00
And I think that's a really interesting area for us.
而且我觉得这对我们来说是一个非常有意思的领域。
78:03
Yeah, I mean, we kind of feel like our best contribution
对，我的意思是，我们有点觉得我们最好的贡献
78:05
to solid state physics and science
对 solid state physics 和科学来说
78:07
could be if we made these tools and topics profitable.
如果我们让这些工具和话题变得有利可图，这也有可能
78:12
Similar to how chat GPT made CS and LLM majors
就像 chat GPT 让 CS 和 LLM 专业
78:16
way more popular in colleges before and after.
在大学里变得热门多了，之前和之后都是如此
78:19
We'd love to show that solid state physics, material science
我们很想证明，solid state physics、material science
78:22
research can make a big impact commercially.
研究可以在商业上产生巨大影响
78:24
And then they will attract more attention.
然后它们就会吸引更多关注
78:26
Young people will want to study physics, which will be a dream.
年轻人会想学物理，那真是梦寐以求的事
78:31
And you know, Bell Labs did make a huge commercial impact,
而且你知道，Bell Labs 确实带来了巨大的商业影响，
78:34
so they failed to commercialize some of their incredible
所以他们也未能把一些了不起的
78:36
advances.
进展商业化。
78:37
They of course did some research that just couldn't be
他们当然做过一些根本没法
78:40
commercialized like the cosmic background, but they did
商业化的研究，比如 cosmic background，但他们确实
78:43
benefit a lot from like the vacuum tube, connecting east coast
从像 vacuum tube 这类东西里受益匪浅，把东海岸
78:46
west coast by foam lines.
和西海岸用 foam lines 连接起来。
78:49
Yeah, and I think it's really also interesting to think
是啊，而且我觉得，想想也真的挺有意思
78:51
that technology and capital are incredibly intertwined.
技术和资本是极其紧密交织在一起的。
78:54
So if you look at what is the progress on chat bots for many
所以，如果你看看 chat bots 这么多
78:58
years, like imagine recounting, OK, what is the progress
年来的进展，比如想象一下回顾一下，OK，进展到底是什么，
79:01
from chat bots from, I'll say, 2010 to 2015?
从 chat bots 从，我想说，2010 到 2015 年？
79:06
We'd be kind of at a loss to say, over that five-year period,
我们大概会有点不知道怎么说，在那五年里，
79:09
how much did they improve?
它们到底进步了多少？
79:11
Whereas if you look at the period from 2021 to 2026,
而如果你看 2021 到 2026 年这段时间，
79:16
you know, it's night and day.
你知道，那简直是天壤之别。
79:17
And what happened was chat GPT and these other systems
然后发生的事情是，ChatGPT 和其他这些系统
79:20
were able to achieve a product market fit,
能够实现 product market fit，
79:22
and it changed the capital landscape entirely.
并且彻底改变了资本格局。
79:25
This changes the landscape for hiring, compute, data,
这改变了招聘、compute、data，
79:30
et cetera.
等等领域的格局。
79:31
The virtual cycle.
良性循环。
79:32
Exactly.
没错。
79:33
And success be success.
而成功会带来成功。
79:34
Exactly.
79:35
So technology, it's incredibly coupled to the capital and
79:38
those resources, and we want to achieve the same thing in the
79:40
physical world.
79:41
Yeah.
79:42
So one thing I've been seeing is there's this move from science
79:45
being funded from government grants and so on to VCs and
79:50
private funding.
79:51
Do you all plan on making anything you're doing open source,
你们打算把你们正在做的任何东西都做成 open source 吗，
79:55
like you're releasing data sets or actual models?
比如你们会发布 data sets 或者实际的 models？
79:58
Is this something like in the future, even, you know,
这是不是比如在未来，甚至，你知道，
80:00
there are many models which may be not your state of the art,
有很多 models 可能不是你们 state of the art 的，
80:05
but which could still be useful release?
但仍然可能是值得发布的？
80:07
Absolutely.
当然。
80:07
I mean, so we have a couple of things.
我是说，所以我们有几件事。
80:09
So we have been contributing to open source to kind of like
所以我们一直在为 open source 做贡献，有点像是
80:13
climate change, the materials project, code base, custodian,
气候变化、the materials project、code base、custodian，
80:17
their DFT runner, towards Sim.
他们的 DFT runner，面向 Sim。
80:20
It was actually created by one of the researchers here at
它其实是由这里的一位研究员创建的，就在
80:23
Vigit Gangan and he maintains it still, Jack's MD,
Vigit Gangan，而且他现在还在维护它，Jack's MD，
80:27
a Vigit maintains it.
一个 Vigit 在维护它。
80:28
So we do a lot of open source contributions.
所以我们做了很多 open source 贡献。
80:30
A scheduling, a mega-tron.
一个是 scheduling，一个是 mega-tron。
80:32
Yeah, so we've been very active contributors back into
是的，所以我们一直是非常活跃的贡献者，回馈到
80:35
open source.
open source。
80:36
And we also have an academic grant program where we give
而且我们还有一个学术资助项目，我们会给
80:39
academic gift grants to academic groups in universities
大学里的学术团队提供学术赠款
80:43
that we feel like are really advancing this direction
我们觉得这些团队真的在推进这个方向
80:46
towards since the superintelligence.
朝着 safe superintelligence 的方向。
80:48
So there's also been great to see.
所以看到这些也很棒。
80:49
I think the first paper from this funding is about to come out.
我想这项资助的第一篇论文快要出来了。
80:53
So it'll be exciting to just get more papers come off
所以能不断看到更多论文出来会很令人兴奋
80:56
like this, yeah?
就像这样，对吧？
80:57
OK, let's still man this say that you've successfully
OK，咱们来 steelman 一下，假设你已经成功
81:00
have 100% lab automation, which does everything
拥有 100% 的 lab automation，它能做所有事情
81:02
characterizes everything correctly.
能正确地 characterize 一切。
81:04
You have all of the fun agents, which can do everything.
你还有所有那些好玩的 agents，它们什么都能做。
81:07
I still am somewhat unclear about how you actually get
我还是有点不太清楚，你实际到底怎么才能
81:10
to superconductivity.
实现 superconductivity。
81:11
There's a lot of hard problems to solve.
还有很多难题要解决。
81:14
What is the path there when there is no sort of like theory
当根本没有什么类似 theory 的东西来做 model 时，那里的路径是什么
81:18
for model for most of the strongly correlated systems
对于大多数 strongly correlated systems
81:21
or high temperature systems?
或者 high temperature systems？
81:23
Yeah, so our view is that it's not hard
对，所以我们的看法是，这并不难
81:27
to think of chemical spaces that would host superconductivity.
去想出那些可能承载 superconductivity 的 chemical spaces。
81:31
What's really hard is to synthesize them.
真正难的是把它们 synthesize 出来。
81:33
That's why we're really emphasizing since the superintelligence.
这就是为什么我们真的很强调这一点，因为 superintelligence。
81:36
So maybe one example to give historically
所以也许可以举一个历史上的例子
81:38
is Alex Mueller, when he thought that
就是 Alex Mueller，当他想到
81:40
transmission metal oxides might be a good path to superconductivity,
transmission metal oxides 可能是通往 superconductivity 的一条好路，
81:43
he was actually trying to nucleate the nickel oxygen
他其实是在试着让 nickel oxygen nucleate
81:46
and some cations, and that didn't work.
和一些 cations，但没成功。
81:49
And then he tried cooperates just copper oxygens
然后他试了 cooperates，就是 copper oxygens
81:52
and some cations, and that worked.
和一些 cations，结果成功了。
81:53
And then he got a Nobel Prize.
然后他得了 Nobel Prize。
81:54
But then we know that later on turns on nickelase
但后来我们知道，之后会开启 nickelase
81:56
was a good idea.
那是个好主意。
81:57
I mean, nickel and copper are next to each other.
我是说，nickel 和 copper 是挨着的。
81:59
It was a waste of two ways, right?
这在两方面都是浪费，对吧？
82:01
So there are only so many 3D transmission metals.
所以，3D transmission metals 也就那么几种。
82:03
And then Harold Wang, one of our Stanford professors here
然后就是 Harold Wang，我们这儿的一位 Stanford 教授，
82:07
who's incredible, he realized he can make nickelase
他特别厉害，他意识到自己能把 nickelase 做成
82:10
and thin film form and show the superconductivity.
thin film 形式，并展现出 superconductivity。
82:13
So if we have a really good synthesis of intelligence
所以，如果我们有一个非常好的 synthesis of intelligence
82:17
in the lab, and we can really scale up these things,
在实验室里，而且我们真的能把这些东西 scale up，
82:20
I think we won't run out of ideas or the LLM won't run
我觉得我们不会想不出点子，或者 LLM 也不会想不出
82:22
out of ideas about what directions to try,
点子，不知道该试哪些方向，
82:26
like 3D transition metals, oxygen,
比如 3D transition metals、oxygen，
82:28
like there are very related ideas here.
比如这里有一些非常相关的思路。
82:30
Like many people have been trying like cobalt aids,
比如很多人一直在尝试类似 cobalt aids 的东西，
82:32
which is like instead of copper or nickel, you try cobalt.
这就像是，不用 copper 或 nickel，而是用 cobalt。
82:36
So I think that's super exciting.
所以我觉得这超级令人兴奋。
82:37
Like maybe one historical example to also study
比如说，也许还可以研究一个历史上的例子
82:40
is the Japanese group that discovered magnesium
就是发现了 magnesium dipolite 的那个日本团队
82:43
dipolite.
你也知道，那是在 ambient pressure 下、温度最高的 conventional superconductor
82:44
As you know, that's the highest temperature,
而且他们发现它，就是靠试了一大堆材料
82:46
ambient pressure, conventional superconductor.
我觉得他们试了 30,000 种不同的东西
82:48
And they discovered it by just trying a bunch of materials.
其中有 30 种似乎有很有意思的 superconductivity
82:51
I think they tried 30,000 different things.
82:54
30 of them seemed to host interesting superconductivity.
82:57
And MGB2 was one of them.
而 MGB2 就是其中之一。
82:59
But magnesium dipolite set on people's shelves
但 magnesium dipolite 一直摆在人们的架子上
83:02
as a precursor for decades before then.
作为 precursor，在那之前已经放了几十年。
83:05
And BCS theory had been invented back in 1957.
而 BCS theory 早在 1957 年就已经被提出来了。
83:08
So the way people found these materials
所以人们发现这些材料的方式
83:10
wasn't projected from theory.
并不是从理论推导出来的。
83:12
It wasn't because they couldn't synthesize until then.
也不是因为在那之前他们没法 synthesize。
83:15
It was just like they tried a bunch of things.
就只是他们试了一堆东西。
83:16
So now imagine if you have this perfect automation
所以现在想象一下，如果你有这种完美的自动化
83:19
that you're describing, it sounds like a dream.
就是你正在描述的那种，听起来就像做梦一样。
83:21
And we tried the 30,000, the Takahito group
而且我们试了那 30,000，Takahito group
83:24
tried in his career in a month.
在他职业生涯里一个月就试过。
83:28
We just really increased the surface area for luck.
我们真的把运气的接触面积大大增加了。
83:31
No, but it's very inspiring what you've done.
不，但你做的这些真的很鼓舞人。
83:33
Congrats on your success.
恭喜你成功。
83:34
I feel like you're creating, yeah, the modern like,
我感觉你在创造，嗯，那种现代的，就像，
83:36
play gone.
play gone.
83:37
Imagine recruiting must be like super easy for you.
想象一下，招聘对你来说肯定超级容易吧。
83:39
So I'm just like a little bit jealous,
所以我就有点小嫉妒，
83:40
but you've worked hard for it to get here.
但你能走到这一步也是付出了很多努力的。
83:43
So yeah.
所以，嗯。
83:44
Yeah, well, thank you so much.
是啊，嗯，非常感谢。
83:45
Yeah, it's a great talent for you both.
是啊，你们俩都很有才华。
83:46
Yeah, super fun.
是啊，超级好玩。
83:47
Yeah.
对。