Latent Space
Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub
00:00
00:05
Great to be here.
很高兴来到这里。
00:07
What an exciting morning,
真是个令人兴奋的早晨，
00:08
so many cool announcements.
有这么多酷炫的发布。
00:09
I think the future of bio-science is being
我觉得 bio-science 的未来
00:12
like announced right here.
就好像正在这里被宣布一样。
00:14
So this is the modeling session.
所以这是 modeling session。
00:17
We are all modelers,
我们都是 modelers，
00:19
so of course, it's natural for us to talk about data.
所以当然，我们聊 data 是很自然的。
00:22
So to start out with,
那么，先从这里开始，
00:24
I think one of the things that I like to think about when it comes to data
我觉得，说到 data，我很喜欢思考的一件事就是，
00:29
is like, is there a scaling,
就是，有没有一种 scaling，
00:32
like how do you scale data properly?
比如，你怎么才能正确地 scale data？
00:34
And this has brought me to this question about the bitter lesson,
而这就让我想到一个关于 bitter lesson 的问题，
00:37
but recastened the frame of data.
不过是以 data 的框架重新表述。
00:41
So the bitter lesson for those of you who are not AI people,
所以，对于你们当中不是 AI 圈的人来说，bitter lesson 说的是，
00:45
is the statement that methods that scale win eventually,
那些能 scale 的方法最终会赢，
00:50
if you can just scale enough it wins.
只要能 scale 得足够多，它就会赢。
00:52
And so my question I'll start with Sal here,
所以我的问题，我先从 Sal 开始，
00:54
is there a bitter lesson for data?
数据也有 bitter lesson 吗？
01:00
Yeah, I mean, for sure.
嗯，我是说，当然有。
01:01
I think that you obviously need the right data, right?
我觉得你显然需要正确的数据，对吧？
01:06
So I think one misconception of scaling laws is that scaling laws are everywhere
所以我觉得对 scaling laws 的一个误解是，scaling laws 无处不在
01:12
and they always exist.
而且总是存在。
01:14
I think a lot of the work is actually finding that scaling law, right?
我觉得很多工作其实是在找到那个 scaling law，对吧？
01:17
So a lot of what we really do is trying to figure out what's the situation
所以我们真正做的很多事，就是试着搞清楚，什么样的情况
01:24
where if you put more compute into it,
是你往里面投入更多 compute，
01:26
if you put more data into it,
投入更多 data，
01:27
you actually get a better result out.
就真的能得到更好的结果。
01:30
And the reason that's a great situation is then once that happens,
而这种情况之所以很棒，是因为一旦这发生了，
01:34
you can kind of just turn the crank, right?
你就可以有点像摇曲柄一样一直转，对吧？
01:35
Like becomes an engineering problem, which is something I like.
这就变成了一个 engineering problem，而这是我喜欢的那种事。
01:39
But that has to do with architecture,
但这跟 architecture 有关，
01:44
but it also very much has to do with data.
但这也在很大程度上跟 data 有关。
01:45
So if you don't have the data with the right information statistics
所以如果你没有带有正确信息统计的 data
01:49
to solve the problem that you want,
来解决你想要解决的问题，
01:51
which is I think probably pretty obvious,
我觉得这大概挺明显的，
01:53
like if you think about it for a second,
就比如你稍微想一下，
01:55
like you're not going to get a model that has the capabilities
你不可能得到一个具备你想要的能力
01:58
and the understanding of you want.
以及你想要的理解的 model。
02:00
So you can only really pull the information from the data that you have
所以你其实只能从你拥有的 data 里提取信息
02:04
and then use that to generate, like, to generalize beyond, right?
然后用它来 generate，就像，去 generalize beyond，对吧？
02:09
Yeah.
对。
02:10
So when it comes to data collection,
那说到 data collection，
02:12
how do you think about what types of data modalities are best?
你觉得哪几类 data modalities 最好？
02:16
And maybe a kind of related question is,
还有一个可能相关的问题是，
02:18
do you think that modellers have a tendency to work on problems
你觉得 modellers 是不是有一种倾向，去做那些
02:22
which are the data is available rather than maybe the data that,
有 data 可用的问题，而不是也许那些 data，
02:25
or the problem that solves your goal most,
或者最能解决你目标的问题，
02:28
maybe has the most impact for translational medicine?
也许对 translational medicine 的影响最大？
02:31
Oh, for sure.
哦，那肯定。
02:32
So I think that, well, I wouldn't speak for all modellers,
所以我觉得吧，嗯，我代表不了所有 modellers，
02:35
but I'm lazy, so I'm going to work with what's available.
但我比较懒，所以我会用手头能用的东西。
02:39
And I think there is actually a good side and a bad side to this.
而且我觉得这其实有好的一面，也有坏的一面。
02:45
And the good side is that, I think when you're doing machine learning
好的一面是，我觉得当你做 machine learning
02:51
or conventional machine learning,
或者传统的 machine learning 的时候，
02:53
you're really looking for the most pristine high quality data
你其实就是在找最干净、最高质量的 data
02:57
examples that you can find.
你能找到的那些例子。
02:59
But when you actually, if you're like me,
但当你真的，如果你像我一样，
03:02
you're actually someone who goes
你其实是那种会去
03:04
and you're kind of like rooting around in the back room
然后有点像在后屋里翻来翻去，
03:06
looking for two people, like, junk, like, what's there?
找两个人，像是，垃圾，像是，那儿有什么？
03:09
So like a concrete example of this is when we train,
所以，一个具体的例子就是，当我们训练的时候，
03:12
like, when we train a protein language model,
比如说，当我们训练一个 protein language model 时，
03:15
we train on metagenomic sequences,
我们是在 metagenomic sequences 上训练的，
03:17
which are not the highest quality data.
这些并不是最高质量的 data。
03:19
In fact, much of that data I can guarantee you
其实，我可以向你保证，那些 data 中很多
03:22
isn't even like a real whole protein.
甚至都算不上是一个真正完整的 protein。
03:25
And yet, that makes the performance of the model go up
但即便如此，这反而会提升 model 的 performance
03:28
for designing real proteins that work,
用来设计真正能起作用的 protein，
03:31
that make for understanding proteins that we know are things.
也有助于理解那些我们知道是真实存在的 protein。
03:35
And so that's like a positive side of it.
所以这算是它的一个积极面。
03:41
But that can lead you to like a negative place
但这可能会把你带到一个负面的境地。
03:43
where you say, well, let's just scale up the data
就是你说，咱们干脆把能轻松生成的数据
03:46
that we can generate easily.
直接 scale up 就好了。
03:49
And I think that's not the necessarily way to do it.
但我觉得这不一定是对的做法。
03:54
And I think that's one of the things
而且我觉得这正是
03:55
that's really exciting to me about what
我们今天聊到 BBI 时让我特别兴奋的地方之一。
03:57
we're talking about here today with the BBI.
它真的是直接去问，我们
04:01
It's really going out and saying, what
解决问题到底需要什么 data？
04:04
is the data that we need to solve the problem?
这就是我们解决问题需要的 data 吗？
04:07
It's the right resources.
是合适的资源。
04:09
But also the right community.
但也是合适的社区。
04:11
I think that one thing that we do, or we try
我觉得我们做的一件事，或者说我们尝试
04:15
and do it by all of anyway, is to work in the open,
而且反正都是大家一起做的，就是公开地工作，
04:18
work with the community, and move the whole community forward.
和社区合作，推动整个社区向前发展。
04:20
And I think that's a really critical piece for this
而且我觉得这对这件事来说是非常关键的一环
04:23
because if you just have people who are building models,
因为如果你只有一群在构建 models 的人，
04:27
we're going to lean towards the data that exists.
我们就会倾向于现有的 data。
04:32
If you just have people who generate the data,
如果你只有负责生成 data 的人，
04:36
they're going to lean towards the things that can be generated.
他们就会倾向于那些能被生成出来的东西。
04:38
But if you can work together as a community
但如果你能作为一个社区一起合作，
04:41
every step of the way in an open fashion,
一路上的每一步都以开放的方式进行，
04:44
then you can actually figure out what is the data that we need.
那你才能真正搞清楚我们需要的 data 是什么。
04:48
And then you can find that scaling law.
然后你就能找到那个 scaling law。
04:51
What do you think, Sal, a third or less than for data?
你觉得呢，Sal，data 是三分之一还是更少？
04:54
Yeah, so I think the, so I was that deep mind when
对，所以我觉得，那个，所以我当时在 DeepMind 的时候
05:00
Rich was with us and working and thinking about this idea
Rich 当时和我们在一起，一边工作，一边在琢磨这个想法
05:04
of the bit less than when Rich and I took the thing
就是那个 bit 比 Rich 和我拿走那个东西的时候要少一点
05:10
that I took from Rich's original sort of lecture
那个我从 Rich 最初那场类似讲座里得到的东西
05:14
that he actually gave at DeepMine was not
他其实是在 DeepMine 讲的那场，并不是
05:18
about the actual notion of whether data is useful
关于数据到底有没有用这个实际概念
05:22
or how should we think about machine learning.
或者我们该怎么看待 machine learning。
05:24
I think my sort of take was that he was
我觉得我的理解大概是，他其实是在
05:28
talking about something more conceptual.
讲一些更概念化的东西。
05:31
And the conceptual thing that he was talking about
而他当时谈到的那个概念性的东西
05:33
is sometimes when we're looking at problems,
是，有时候当我们在看问题时，
05:37
we think about solutions in a very religious way.
我们会以一种非常宗教式的方式去思考解决方案。
05:42
I'm a modeler or I'm a data generation person.
我是个 modeler，或者我是个做 data generation 的人。
05:45
And I think that is the bitter lesson
而我觉得，这就是那个 bitter lesson
05:48
that if you approach the problem with that mindset,
就是，如果你用那种心态去处理问题，
05:52
you might not succeed.
你可能不会成功。
05:54
The problem comes first.
问题本身是第一位的。
05:56
And you should be flexible in terms of your solution space.
而且你在 solution space 上得灵活一点。
06:00
You should try both things.
两种方法你都该试试。
06:01
You should see what you should try to understand
你得看看自己该试着去理解什么
06:04
what is the problem that you're trying to solve.
你到底想解决的是什么问题。
06:07
And if it requires modeling effort, do the modeling effort.
如果它需要 modeling effort，那就去做 modeling effort。
06:10
And if it requires the data collecting more data,
如果它需要 data、需要收集更多 data，
06:12
then collect data.
那就去收集 data。
06:14
But sort of the thing that happened in machine learning
但有点像 machine learning 里发生的那种事
06:20
in those times is people said, we are machine learning
那时候人们说，我们是 machine learning
06:24
researchers, the data set is there.
研究者，data set 就在那儿。
06:27
There is some training data, here's some test data.
有一些 training data，这儿有一些 test data。
06:31
And we will just optimize the model.
我们就直接 optimize 这个 model。
06:33
And that is broken.
但这是有问题的。
06:35
In the sense that if your eventual goal
也就是说，如果你最终的目标
06:39
is to solve the problem, then you have
是解决这个问题，那你就
06:41
to be able to look at both aspects of what
必须能够看到……的两个方面
06:45
goes into the process.
都会进入到这个过程里。
06:46
And what goes into the process is not just data or modeling
而且进入这个过程的不只是 data 或 modeling
06:52
is also about sort of expertise.
还涉及到某种专业知识。
06:56
So the example that I have is like with alpha fold,
所以我举的例子就像 AlphaFold 那样，
07:02
we looked at what was possible with current existing data
我们当时看的是，用当前现有的 data 能做到什么
07:09
because we did not have the core expertise
因为我们没有那种核心专业知识
07:12
of now sort of going to, or even the resources, by the way,
去进行现在这种尝试，或者说，顺便说一句，连资源都没有，
07:17
to sort of say, let's augment the PDB
去说，咱们来 augment 一下 PDB
07:21
by a significant sort of order of magnitude.
差了相当显著的一个 order of magnitude。
07:24
It was just not, I mean, the amount of sort of investment
那并不是说，我是说，那种投资的规模
07:27
that organizations and scientists across the world
全球各地的组织和科学家
07:30
had put in to construct that data was invaluable.
为了构建那些 data 而投入的，是极其宝贵的。
07:33
So there you have to sort of focus on,
所以在那方面，你就得有点把重点放在，
07:36
well, getting the biggest bang out of your buck
嗯，就是让每一分钱都花出最大的效果
07:39
for by investing it in modeling.
也就是通过把它投入到 modeling 里。
07:42
But in other areas, if you're sort of approaching, say,
但在其他领域，如果你有点像是在切入，比如说，
07:46
Selginomics, then we took the same approach
Selginomics，然后我们采用了同样的方法
07:49
and said, well, what can you do with Selbaji?
然后说，嗯，用 Selbaji 能做什么呢？
07:52
If you're looking at sort of Selginomics,
如果你看的是某种 Selginomics，
07:54
and there, it was very clear that after a bunch of work
在那里，经过一大堆工作之后，已经非常清楚
07:59
that the data was not there yet, to be able to build,
data 还没到位，还不足以去构建，
08:03
to go after that grand ambition of building the virtual cell.
去追逐构建 virtual cell 的那个宏大愿景。
08:07
And so it just brings people together to focus on
所以它只是把人们聚到一起，专注于
08:10
what are the actual challenges if you want to advance science
如果你想推动科学进步，真正要面对哪些挑战
08:15
rather than being religiously following
而不是教条式地追随
08:17
like advances on data generation or in modeling.
比如 data generation 或 modeling 方面的进展。
08:23
– Cool.
– 不错。
08:24
That's, I really like the answer.
这，我真的很喜欢这个回答。
08:25
So what do you think like the actionable takeaway is
所以你觉得，嗯，可落地的 takeaway 是什么？
08:29
from this, like, I mean, always define the problem first
从这里面，就是，我是说，永远先定义问题
08:33
and then figure out what solution space to search over.
然后再弄清楚该在什么样的 solution space 里搜索。
08:37
But more broadly, like, how should the community be thinking
但更宽泛地说，就是，社区应该怎么去思考？
08:41
about this as we go forward for the next generation
关于这件事，随着我们为下一代继续往前走
08:44
of basically trying to solve translational medicine?
基本上就是要去解决 translational medicine？
08:47
– Yeah, I think the advice that I give to anyone
– 对，我觉得我给任何人的建议
08:50
who is working, who is starting in the area
只要是正在这个领域工作、刚进入这个领域的人
08:53
is that think of yourself as a multidisciplinary person.
就是，把自己看成一个 multidisciplinary 的人。
08:57
Understand the problem first.
先理解问题。
08:59
Why are you working on this problem?
你为什么要做这个问题？
09:00
What are we trying to achieve?
我们到底想要达成什么？
09:02
And then think about what is it that will be needed,
然后要想想，到底需要什么，
09:06
whether it's modeling, whether it will be compute scaling,
到底是 modeling，还是 compute scaling，
09:08
whether it will be sort of data generation.
还是说 data generation。
09:11
So understanding of that problem sort of is extremely important.
所以，理解这个问题其实极其重要。
09:16
And of course, then you have to build sort of your expertise
当然，然后你还得建立起自己的专业能力
09:19
in the house.
在团队内部。
09:20
But, and there are also, it's not as if you don't have constraints.
但是，而且也有——并不是说你就没有限制。
09:24
Maybe there are constraints.
也许确实有限制。
09:25
Like, there is only a specific amount of data
就比如说，只有特定数量的 data
09:28
that you can sort of generate.
是你大概能 generate 出来的。
09:30
Or there is a specific model size
或者说，有一个特定的 model size
09:33
that you can afford to sort of train.
是你大概负担得起去 train 的。
09:36
Understand those constraints and stop and try to fail fast.
理解这些 constraints，然后停下来，尽量 fail fast。
09:41
And sort of look at what are the approaches
然后再大概看看有哪些方法
09:43
that are going to be feasible in the long term
长期来看会是可行的
09:45
in getting you to that intercept, the intercept level,
能让你到达那个 intercept，那个 intercept level，
09:50
the impact level that you are going after.
你要追求的那个影响力层级。
09:53
– Okay, cool.
– 好，酷。
09:54
So this actually leads me right into my next question,
所以这其实正好引出了我的下一个问题，
09:57
which is, when I look at sort of the evolution of modeling,
也就是，当我看 modeling 的演变时，
10:01
let's say, AlphaFold 2 was essentially a work of art.
比如说，AlphaFold 2 基本上就是一件艺术品。
10:05
Like, a bunch of very carefully handcrafted features.
就像，一大堆非常精心手工打造的 features。
10:09
There was a lot of thought, very intentional thought
背后有很多思考，非常刻意的思考，
10:11
done to every part of the solution.
投入到了这个解决方案的每一个部分。
10:14
And then, I think, some of Google or Afovets' work
然后，我觉得，Google 或 Afovets 的一些工作
10:20
has been, you know, kind of still stays in that space.
你知道，它们算是还一直留在那个领域里。
10:23
But it seems like the general consensus
但看起来，大家的普遍共识是
10:24
to move to more scalable, more general strategies.
转向更 scalable、更通用的策略。
10:28
Do you think that there's still a lot of,
你觉得，是不是仍然有很多，
10:30
like, artisanal craft solutions or problems that need that?
比如说，那种手工定制式的解决方案，或者需要那种做法的问题？
10:34
Or is our resources better spent on generally speaking,
还是说，总的来说，我们的资源最好花在，
10:39
you know, fixed resources, fixed money can go to compute,
你知道，固定的资源、固定的钱，可以投到 compute 上，
10:41
it can go to talent, it can go to data, you know,
它可以流向人才，也可以流向 data，你知道，
10:44
should we be focusing on scale first?
我们是不是应该先聚焦在 scale 上？
10:48
So my, again, sort of let's approach it from first principles.
所以，我呢，再说一遍，算是，咱们从 first principles 来想吧。
10:54
And think about, even so, like, when you think about,
然后想想，即便如此，就像，当你想到，
10:58
like, the arts and crafts of how do you construct a model
就像，你如何构建一个 model 的那套手艺和门道，
11:01
to be better, I think it was not an accidental thing, right?
让它更好，我觉得这不是偶然的事，对吧？
11:07
We did a lot of experimentation, but there was a vision behind it
我们做了很多实验，但背后有一个愿景，
11:10
that all this scientific intuition that came from biophysics
就是所有这些来自 biophysics 的科学直觉
11:15
and biochemistry, that those interactions,
还有 biochemistry，那些相互作用，
11:18
that sort of amino acid sort of residues are not just doing
那种 amino acid 的 residues 并不是只是在做
11:22
their own thing, they are being influenced by other residues.
自己的事，它们还受到其他 residues 的影响。
11:25
So let's bake that in.
所以咱们把这一点融进去吧。
11:26
Like, if you can, if you have learned something
比如说，如果可以的话，如果你学到了某些东西
11:30
from the scientific community, use it.
从科学界那里，就把它用上。
11:32
Use that information and try to sort of encourage the model
用那些信息，然后试着去鼓励 model
11:35
and sort of give it that unfair advantage that it has.
然后某种程度上给它那种它拥有的不公平优势。
11:42
And it does basically make the model much more data efficient
而且这基本上确实让模型变得 data efficient 多了
11:46
because it doesn't have to replicate everything.
因为它不用把所有东西都复制一遍。
11:49
I also have to sort of mention that curating good data
我还得稍微提一下，curating 好的 data
11:55
is in art in itself, right?
本身就是一门艺术，对吧？
11:57
It's not as if, like, you will say, well, put more data.
并不是说，你就来一句，好吧，多放点 data 就行了。
12:00
If you replicate the same amount of data, you're not going anywhere.
如果你复制同样数量的 data，那你哪儿也去不了。
12:04
So it's not just about big data.
所以这不仅仅是 big data 的事。
12:05
It's about good data and actually understanding
关键在于好的 data，以及真正理解
12:08
sort of the coverage of what data is necessary
算是，对于需要哪些 data 的覆盖范围
12:12
for making progress in the problem is, in fact,
为了在这个问题上取得进展，事实上，
12:14
I would say, a much interesting and much more challenging
我会说，是一个非常有趣、也更具挑战性的
12:18
problem in its sense.
问题，从某种意义上说。
12:20
Yeah, well, it looks like you have a thought, so do you think?
对，嗯，看起来你有想法，所以你觉得呢？
12:24
Yeah, I think, I mean, I very much agree.
对，我觉得，我是说，我非常同意。
12:28
I think as you, it really depends on, you know,
我觉得，就像你说的，这真的取决于，你知道，
12:32
as you're saying, the problem to be self.
正如你说的，问题本身。
12:34
So if you have kind of a smaller amount of data,
所以如果你手里的 data 量有点小，
12:38
then certainly having more inductive bias in the model
那 model 里有更多 inductive bias 肯定
12:42
is going to help you, right?
会帮到你，对吧？
12:44
And you actually need that to get the results
而且你其实需要它来得到结果
12:46
on smaller scales of data.
在更小规模的 data 上。
12:48
As you, you know, as you get more and more data, you know,
随着你，你知道，随着你的 data 越来越多，你知道，
12:53
you see that, like, sometimes the model
你会发现，就像，有时候 model
12:56
can find things that, like, you didn't necessarily know about
能找到一些，就像，你本来不一定知道的东西
13:00
and sometimes that inductive bias, if it wasn't exactly
而且有时候那种 inductive bias，如果它不是完全
13:03
correct, can hold you back, right?
正确的话，可能会拖你后腿，对吧？
13:06
Like, so there is definitely that, like, tipping point,
就像，所以肯定有那个，就像，临界点，
13:09
that seems better, but, you know, I guess, like, also,
那个看起来更好的，但是，你知道，我猜，而且，
13:12
I kind of object to your question, too, because I think
我其实也有点不认同你这个问题，因为我觉得
13:13
there actually is a lot of, I think there's a lot of craft
其实确实有很多，我觉得这里面有很多门道
13:16
actually to the scaling part of things as well.
其实在事情的 scaling 这部分也是一样。
13:19
So, you know, you need very, there's a lot of algorithmic
所以，你知道，你需要非常，这里面有很多 algorithmic
13:24
work that actually goes into taking these models
真正投入的工作，是把这些 models 拿过来
13:27
and training them on more data, putting more
然后用更多数据 train 它们，投入更多
13:29
compute in them, making them bigger, right?
compute 进去，把它们做得更大，对吧？
13:32
And we see that, like, not only is that
而且我们看到，就，这不仅仅是
13:34
from, like, an infrastructure perspective
从，就，infrastructure 的角度来看，
13:36
and, you know, making the models, like,
而且，你知道，让这些 models，就，
13:39
run inference faster, but, you know, we're actually seeing
跑 inference 更快，但是，你知道，我们实际上看到
13:42
that, like, things are going beyond transformers
就是，事情正在超越 transformers
13:45
to more bespoke architectures that can, well, I mean,
到更定制化的 architectures，它们可以，嗯，我的意思是，
13:48
I don't think we're in a post-transformer world
我不觉得我们处在一个 post-transformer 的世界里
13:51
in any way, shape or form, but, you know,
不管从哪个角度、以哪种形式，但，你知道，
13:53
we're modifying those architectures, right?
我们是在修改那些 architectures，对吧？
13:55
Like, in order to make them more fit for purpose
就是，为了把它们做得更符合实际用途，
13:58
and actually work better, even at, like,
而且实际上表现更好，哪怕是在，比如说，
14:00
internet scale data.
internet scale 的数据下。
14:01
So, you know, I think it's, that's kind of the project, right?
所以，你知道，我觉得这，这差不多就是这个项目，对吧？
14:07
Like, as more and more data comes in,
就像，随着越来越多的 data 进来，
14:10
it's not only, you know, helping to curate that data
你知道，这不仅是在帮忙整理那些 data
14:13
and select what is the next, like, batch of information
然后挑出下一个 batch，就像，信息
14:17
that we actually need.
到底什么才是我们真正需要的。
14:18
You know, it's more information than data.
你知道，这更多是信息，而不是 data。
14:20
How do you get that next batch of information
你怎么拿到下一个 batch 的信息，
14:22
that the models need?
是 models 需要的那种？
14:24
But then also, like, at every step of the way,
但然后还有，就像，在每一步上，
14:27
at every scale, what is the right architectures
在每个 scale 上，什么才是正确的 architectures
14:31
than to get the most out of that data?
才能从那些数据里获得最大价值？
14:35
Cool.
酷。
14:36
So, if you think about the state of protein structure
所以，如果你想想现在 protein structure
14:42
prediction right now, you know, the news will say,
prediction 的状况，你知道，新闻会说，
14:44
protein structure prediction has been solved.
protein structure prediction 已经被解决了。
14:47
But then, like, if you talk to my friends,
但然后，就像，如果你跟我的朋友聊，
14:50
they will all say, we have so much, like, left to do here.
他们都会说，我们这里还有好多，就像，还没做完的事。
14:54
So many things, like, function, dynamics,
好多东西，比如说，function、dynamics，
14:57
and then design are still, like, just wide open problems.
然后 design，也还都只是，那个，完全开放的问题。
15:01
It's been five this year since AlphaFold 2 was announced.
AlphaFold 2 发布到现在，今年已经五年了。
15:05
Progress has been made.
进展是有的。
15:06
But I'm curious, push me.
但我很好奇，你逼我一下。
15:08
What do you think the, like, is there another big leap
你觉得，那个，是不是还有另一次大飞跃
15:13
being made here?
正在这里发生？
15:13
Is there blockers and solving these problems?
解决这些问题是不是有 blockers？
15:17
Do we not have the right data yet?
难道我们还没有合适的 data 吗？
15:18
Do we not have the right algorithms or ideas?
难道我们还没有合适的 algorithms 或想法吗？
15:21
Or is this, like, something that we can see evolving
还是说，这有点像某种我们能看到它不断演进的东西
15:24
in the future, and it's just kind of a matter of time?
在未来，而这差不多只是时间问题而已？
15:28
Yeah, I think science, how it sort of operates
对，我觉得科学，它大概是怎么运作的
15:32
is basically by isolating something
基本上就是通过把某个东西隔离开来
15:35
and then making progress a step-by-step.
然后再一步一步地取得进展。
15:38
So when people sort of say the protein folding problem
所以当人们大概说到 protein folding problem 的时候
15:40
has been solved, like, at a conceptual level, yes,
已经解决了，就，在概念层面上，是的，
15:44
there might be sort of, yes, there have been some advances.
可能算是，对，已经有一些进展了。
15:49
But I think it's also, like, when you think about that narrative
但我觉得，还有就是，当你想到那种叙事的时候
15:56
of proteins building, being the building blocks,
关于 proteins 构建、作为 building blocks 这件事，
16:00
proteins are not blocks.
proteins 不是 blocks。
16:01
And they're not sort of, they don't act as blocks, right?
而且它们也不是，就，它们并不像 blocks 那样起作用，对吧？
16:05
I say proteins are the building blocks all the time,
我一直说 proteins 是 building blocks，
16:07
but actually, like, I don't believe in it, right?
但其实，就，我自己并不相信这个，对吧？
16:11
Proteins are extremely complex.
Proteins 是极其复杂的。
16:14
They are disordered.
它们处于 disordered 状态。
16:16
Their shape might change depending on the context.
它们的 shape 可能会随 context 变化。
16:21
And like, what AlphaFold did, and in fact, like John,
而且，就像 AlphaFold 做的，其实，就像 John，
16:25
John Jumper and I basically used to sort of discuss this,
John Jumper 和我以前基本上会讨论这个，
16:28
like, what are we trying to solve?
就像，我们到底想解决什么？
16:30
We don't know what is the actual true ground state
我们不知道 actual true ground state 到底是什么
16:35
that proteins take, and what is the actual distribution
proteins 会呈现的，以及实际的 distribution 是什么
16:38
of structure that proteins take.
蛋白质所呈现的 structure。
16:42
What we are trying to do is basically,
我们想做的事，基本上就是，
16:44
someone got a structure, deposited it in the PDB,
有人拿到了一个 structure，把它提交到了 PDB，
16:48
and we are trying to replicate and get the same structure.
而我们是在尝试复现，并得到同样的 structure。
16:52
That's what we did, right?
这就是我们做的事，对吧？
16:54
And it just so happens that it's useful.
而它恰好还很有用。
16:58
But that does not mean that we have understood
但这并不意味着我们已经理解了
17:01
all of protein dynamics.
protein dynamics 的全部。
17:03
So I think, like, at the top level,
所以我觉得吧，从顶层来说，
17:08
it's easier to communicate that we have made progress,
更容易传达出我们已经取得了进展，
17:12
but the scientists among us, and in this crowd,
但我们当中的科学家，还有在座这群人里的科学家，
17:15
the we know, right?
呃，我们心里清楚，对吧？
17:19
And I think that's the really important thing for us
而且我觉得，对我们来说真正重要的是
17:22
to understand and have that common ground
要理解这一点，并且有那样的共识
17:24
that a lot of work needs to be done.
那就是还有很多工作要做。
17:26
And we have made progress, and there's a lot to celebrate,
而且我们已经取得了进展，也有很多值得庆祝的地方，
17:30
but let's not stop the funding of protein structure
但咱们别停止给 protein structure
17:34
prediction and protein dynamics
prediction 和 protein dynamics 提供资助
17:36
because we are just getting started.
因为我们才刚开始。
17:39
So what do you think, the biggest blocker?
所以你觉得，最大的 blocker 是什么？
17:42
If there's one thing you could wave a magic wand
如果有一件事，你能挥一挥魔法棒
17:44
and say, like, we have more of point,
然后说，像是，我们有更多的 point，
17:46
that would accelerate one of those things,
那会加速其中一件事，
17:48
function, dynamics, or design,
function、dynamics，还是 design，
17:50
what would the magic wand be, or would you,
那根魔法棒会是什么，或者你会，
17:53
wave into existence?
一挥就让它出现？
17:54
Oh, I've basically been, so I'm a machine learning.
哦，我基本上一直是，所以我是做 machine learning 的。
17:58
My background is quite eclectic.
我的背景挺杂的。
18:01
I started as a security researcher
我一开始是一名 security researcher
18:03
then went into computer vision,
然后进入了 computer vision，
18:05
a Bayesian theory, and finally got into discriminative
Bayesian theory，最后进入了 discriminative
18:09
machine learning and deep learning, and so on,
machine learning 和 deep learning，等等，
18:11
and then AI for coding, and then finally science.
然后就是 AI for coding，最后才是科学。
18:14
So when you ask me that question, like,
所以当你问我那个问题的时候，怎么说呢，
18:17
my computer vision, the computer vision researcher
我的 computer vision，我内心的那个 computer vision researcher
18:19
in me was super excited about CryoEM micrographs.
对 CryoEM micrographs 超级兴奋。
18:23
I was like, what is this sort of PDB data?
我当时就想，这种 PDB data 是什么？
18:25
I should be working at the source, right?
我应该直接在源头做，对吧？
18:28
I should be looking at the CryoEM micrographs.
我应该直接去看 CryoEM micrographs。
18:30
I don't want those structures.
我不想要那些 structures。
18:32
They must be sort of missing out all the data.
它们肯定有点把所有 data 都漏掉了。
18:35
I should just directly operate at CryoEM micrographs.
我应该直接在 CryoEM micrographs 上操作。
18:38
Now, getting models that can scale at that level
现在，能得到能在那个层级上 scale 的模型
18:43
with the right amount of data and can extract all the dynamics
配上合适的 data 量，还能提取出所有 dynamics
18:46
and distributional information that is captured there
以及那里捕获到的 distributional information
18:49
would be an amazing thing.
那会是一件特别了不起的事。
18:52
I tried it, but it requires more work, right?
我试过，但这需要更多工作，对吧？
18:57
Still working on it, but you believe
还在做这个，不过你相信
18:58
that this is a route, this can give you.
这是一条路，能给你带来……
19:01
Yeah, I think at some point of time,
嗯，我觉得在某个时候，
19:02
maybe people better than me would take a stab at it,
也许比我更厉害的人会去试一试，
19:05
and we'll get some there.
然后我们会在那儿有些收获。
19:07
What do you think, Sal?
你怎么看，Sal？
19:09
Yeah, I think that it's really interesting
是啊，我觉得这真的很有意思，
19:11
because I think these models are, I mean,
因为我觉得这些 models 是，我是说，
19:16
they're quite useful, but they're not exactly the problem
它们挺有用的，但它们并不完全是那个问题
19:20
that most people want to solve.
大多数人想解决的问题。
19:22
They solve a very specific purpose,
它们解决的是一个非常具体的目的，
19:24
and then you can also use them to design new proteins, right?
然后你也可以用它们来设计新的 proteins，对吧？
19:29
Like, for example, which is not the thing
就像，比如说，这并不是那种
19:32
that you would kind of start with there,
你会一开始就在那里用的东西，
19:33
but so they're very useful.
不过话说回来，它们非常有用。
19:36
But I kind of think of the models that we build right now
但我有点会想到我们现在构建的 models
19:41
as you've got someone who's trying to understand
就像有人正试图弄明白
19:45
how a bicycle works, and you're modeling a spoke on it, right?
自行车是怎么运作的，而你在给上面的一根辐条做 modeling，对吧？
19:50
And you kind of have to, those models can get better
而且你多少也得这么做，这些 models 可以变得更好
19:53
and better over time, but a lot of what you really need to do
而且随着时间越来越好，但你真正需要做的很多事
19:56
is you need to move from models of spoke to wheels
是你得从辐条的 models 转到车轮
20:00
to maybe whole bicycles, because that's actually
再到也许整辆自行车，因为那其实
20:02
what people want to understand.
才是人们想弄明白的。
20:03
Actually, the contingency analogy
其实，是 contingency analogy
20:05
it seems like people want to use the model of the bicycle
看起来人们想用自行车的那个 model
20:08
to design this part for a pickup truck or something
来给一辆皮卡设计这个零件，或者什么
20:11
like that, right?
之类的，对吧？
20:13
But I think that as we put these models more into the particular context
但我觉得，随着我们把这些 model 更多地放进它们所处的特定 context
20:23
which they're in, right, then we'll actually
也就是它们所在的那个，对吧，然后我们其实
20:27
be able to learn more about these interactions on a broader sense.
就能从更广义的层面更多地了解这些互动。
20:32
And that's really, I think, where things are going to be going.
而且我觉得，这真的就是未来的走向。
20:36
But I agree with you.
但我同意你的看法。
20:37
I think we should keep trying to work on folding models,
我觉得我们应该继续试着做 folding models，
20:39
because I think they're going to keep getting better and better.
因为我觉得它们会变得越来越好。
20:43
So with regards to design, there's a famous crime
所以说到 design，有一句很有名的名言，
20:47
in quote, that which I cannot create, I do not understand.
引用一下：'我不能创造的东西，我就不理解。'
20:51
Now we're in a world where it's really easy to create things
现在我们处在一个创造东西非常容易的世界，
20:54
without understanding them at all.
却完全不需要理解它们。
20:56
So I'm curious what your feelings, your thoughts
所以我很好奇你的感受、你的想法，
21:01
are about how important is to actually have models
关于真正拥有 models 到底有多重要。
21:06
which can help humans understand things
这能帮助人类理解事物
21:09
versus magical black boxes, which
而不是 magical black boxes，它们
21:11
can effectively one-shot a picomole or binder or something like that.
能有效地 one-shot 一个 picomole 或 binder 之类的。
21:18
I mean, I think that we were designing things
我的意思是，我觉得我们当时在设计东西
21:23
with magical black boxes long before AI came around, right?
用的是 magical black boxes，远在 AI 出现之前，对吧？
21:26
So in some sense, I think that's still useful.
所以从某种意义上说，我觉得那仍然有用。
21:31
But I think that the understanding is really important.
但我觉得，理解真的非常重要。
21:34
And I think that's actually one of the big things
而且我觉得，那其实是其中很重要的一点
21:37
I think about a lot with AI.
我经常会思考很多跟 AI 有关的事情。
21:39
Because there's so much for these models to learn.
因为这些 models 要学的东西实在太多了。
21:43
And as intelligence is getting cheaper and cheaper and more available,
而且随着智能变得越来越便宜、越来越普及，
21:48
you can deploy it to learn more and more things
你就可以把它部署出去，让它去学习越来越多东西
21:50
about what's going on in the world.
关于这个世界上正在发生什么。
21:52
But then how do you actually pull that knowledge out of the machine,
但接下来，你到底要怎么把那些知识从机器里提取出来，
21:57
so to speak, right?
可以这么说，对吧？
21:58
And so it's something that I can understand.
所以这件事是我能理解的。
22:00
Maybe that's just my esoteric curiosity.
也许那只是我的一种比较冷门的好奇心吧。
22:03
I think there's so many things to learn.
我觉得要学的东西太多了。
22:05
I think lots of scientists really want to understand things
我觉得很多科学家真的就是想把事情搞明白，
22:08
and the endpoints are maybe not as important.
而 endpoints 也许并没那么重要。
22:10
But then we are here to solve translational medicine as a problem, right?
但我们在这儿，是要把 translational medicine 作为一个问题来解决，对吧？
22:17
Yeah.
是啊。
22:18
I mean, I think you do have to like, the more you dig into things,
我是说，我觉得你确实得，怎么说呢，你越深入去琢磨，
22:23
the more that you can find the right way to keep pushing them forward.
就越能找到合适的方法，继续把它们往前推。
22:28
So I think one thing that's really sailing it to me about this
所以我觉得，关于这一点，真正让我觉得有说服力的一点是
22:32
is I think the models actually have a lot more information in them than we know.
我觉得这些 models 里面实际包含的信息，比我们知道的要多得多。
22:37
So we've worked a lot on like interpretability, for example, for our models.
所以我们在 interpretability 上做了很多工作，比如说，针对我们的 models。
22:43
And you find a lot of information in there,
然后你会在里面发现很多信息，
22:47
which is I think people know that like protein language models
我觉得大家知道，就像 protein language models
22:50
like learn some notion of structure within their representations.
会在它们的 representations 里学到某种 structure 的概念。
22:55
But we find information about some functions.
但我们发现了一些关于某些 functions 的信息。
22:59
We find information about motions, right?
我们发现了一些关于 motions 的信息，对吧？
23:03
And so I think there's a lot in there still to be unlocked,
所以我觉得，这里面还有很多东西有待解锁，
23:07
even from the models that we have now, right?
甚至就是从我们现在有的这些 models 里，对吧？
23:10
And I think that's important for us to understand how do these models work
而且我觉得，理解这些 models 是怎么工作的，对我们来说很重要，
23:15
as we move up the, you know, as we raise the capability.
随着我们往上走，你知道，随着我们提升能力。
23:19
Like at the end of the day, like you expect like a world model
就像说到底，你会期待有一个 world model
23:23
to come out of like trying to compress all this information into the model, right?
是从试着把所有信息压缩进 model 里的过程中产生出来，对吧？
23:28
Like how does the model do its job?
就像这个 model 到底是怎么完成它的工作的？
23:29
How does it design a protein?
它又是怎么设计一个 protein 的？
23:31
Well, you know, it's compressed all the information from evolution
嗯，你知道吧，它把来自进化的所有信息都压缩了
23:36
into this like model, right?
塞进这么一个像 model 一样的东西里，对吧？
23:38
Like so in addition to it, you know,
所以呢，除了它能，你知道，
23:41
being able to spit out something which is useful to us,
吐出一些对我们有用的东西之外，
23:44
there's certainly something to learn just by looking at what's inside.
光看看它里面有什么，肯定也能学到些东西。
23:48
Yeah.
对。
23:49
What do you think, Prushmeet?
你觉得呢，Prushmeet？
23:50
Design versus understand.
设计 vs 理解。
23:53
So, so I have a sort of a different take on this
所以，所以我对这件事有一个稍微不同的看法
23:59
in the sense that I think when we think about models,
我的意思是，我觉得当我们思考 models 的时候，
24:05
I think some level of understanding is necessary.
我觉得某种程度的理解是必要的。
24:09
So I'll say what level of understanding is necessary?
那我会说，需要什么程度的理解才算必要呢？
24:12
So alpha-fold actually is not perfect, right?
所以 alpha-fold 其实并不完美，对吧？
24:16
It's not perfect.
它并不完美。
24:17
It's a 90GT alpha-fold to receive was 90GT at that time on that sort of on that set.
它是一个 90GT 的 alpha-fold，当时在那种……在那个集合上，它收到的是 90GT。
24:26
But even if it was 95GT,
但即使它是 95GT，
24:29
but the PLDDT score was completely uncalibrated.
但 PLDDT score 完全没做 calibration。
24:35
Who would trust it?
那谁还会信它？
24:36
If it would just magically sort of give good answers,
如果它就是能神奇地给出还不错的答案，
24:39
but suddenly tell you here's answer, very, very confident
但突然告诉你，这就是答案，非常非常自信
24:43
and you will be working on it for the next one here
然后你接下来还得继续做下一个，
24:46
and finding out it was completely wrong.
结果发现它完全错了。
24:48
The calibration of the uncertainty measure was extremely important.
uncertainty measure 的 calibration 极其重要。
24:53
So in that sense, we do understand
所以从这个意义上说，我们确实明白
24:57
and we actually made a lot of progress in understanding
而且我们实际上在理解方面取得了很大进展
25:01
how does alpha-fold to behave.
也就是 alpha-fold 是如何表现的。
25:04
Now that's different from how did it work and find the solution.
现在，这跟它是如何运作并找到解决方案是不一样的。
25:10
And there I agree with Sal, that there's a notion of interpretability.
在这一点上我同意 Sal，存在一个 interpretability 的概念。
25:16
Why did it work?
它为什么有效？
25:17
Why did it give this answer?
它为什么给出了这个答案？
25:20
And I think if you think about it, alpha-fold 2 was at the highest level interpretable,
而且我觉得，如果你仔细想想，alpha-fold 2 在最高层次上是 interpretable 的，
25:26
because if you look at how, and this is its ability to generalize,
因为如果你看看它是怎么做到的，而这正是它 generalize 的能力，
25:34
and we did not sort of discover it before launching.
而且我们其实并不是在发布之前就发现了这一点。
25:36
When we launched alpha-fold 2 and made the way it's available,
当我们发布 alpha-fold 2，并让它以这种方式开放使用的时候，
25:41
people found out that it was a great,
人们发现它其实是一个很棒的、
25:45
disordered protein predictor.
disordered protein predictor。
25:47
It could figure out which elements of the protein are disordered.
它能弄清楚 protein 的哪些部分是 disordered 的。
25:52
So it just shows that it generalizes.
所以这就说明它能 generalize。
25:54
Now interpretability asks the question interpretable by whom?
那么 interpretability 问的问题是：对谁来说 interpretable？
26:00
If you are seeing interpretable by a human rational system
如果你是从人类理性系统的角度来看 interpretable 的话
26:04
with the cognitive and computational limitations of the human brain,
考虑到人脑的认知和计算局限，
26:10
then no, alpha-fold 2 is not interpretable.
那么不是，alpha-fold 2 并不是 interpretable 的。
26:14
But if you are asking, is alpha-fold 2 available sort of interpretable
但如果你问的是，alpha-fold 2 是不是可用，或者对一个更大、更复杂的 model 来说，
26:18
to a much larger and more sophisticated model as to how it works?
在某种意义上算得上 interpretable，能理解它是怎么运作的？
26:23
Maybe it is, we just don't get it, right?
也许它是，只是我们搞不懂，对吧？
26:26
But I think as users of alpha-fold 2,
但我觉得，作为 alpha-fold 2 的用户，
26:30
we do need to understand what is it able to do
我们确实需要明白它能做什么，
26:33
and what it is not able to do.
以及它不能做什么。
26:34
So understanding the behavioral characterization
所以，理解 behavioral characterization
26:37
and the strengths and limitations of the model.
以及这个 model 的优势和局限。
26:39
And I think that is extremely important.
而且我觉得这极其重要。
26:41
And we can't just be using these models
而且我们不能只是用这些 models
26:44
without having that behavioral characterization,
却没有那个 behavioral characterization，
26:47
because otherwise it will sort of rather than help
因为不然的话，它与其说会帮上忙
26:50
being helpful, harm us, right?
不如说会害了我们，对吧？
26:53
So your take is that interpretability, strictly speaking,
所以你的观点是，严格来说，interpretability
26:56
isn't necessary, but a way of ensuring
不是必需的，而是一种确保
26:59
the model is trustworthy so humans can make actionable decisions
model 可信，从而让人类能做出可执行决策的方式
27:02
is the thing that people should be focusing on.
才是人们应该关注的事情。
27:04
Exactly.
没错。
27:05
And I also think that interpretability
而且我也觉得 interpretability
27:07
is also in the eyes of the beholder,
也是见仁见智的，
27:09
like who is interpreting it?
比如，到底是谁在解读它？
27:10
If it is a human scientist who isn't trying to interpret
如果是一个并不想解读的人类科学家
27:13
how the model is sort of going about
model 大概是怎么着手的
27:16
and sort of how will it make this prediction?
以及它大概会怎么做出这个 prediction？
27:19
That's a different question from another much larger sort
这跟另一个规模大得多的
27:23
of LLM, if you give some of these large frontier models
LLM 的问题不一样，如果你让这些未来大型 frontier models 中的一些
27:27
of the future access to the activation layers
能接触到 activation layers
27:30
and say, OK, tell me, can you predict what alpha-fold will do?
然后说，好，告诉我，你能预测 alpha-fold 会做什么吗？
27:34
Maybe they will be able to predict.
也许它们就能预测。
27:36
And they will come up with a theory of how actually alpha-fold 2
而且它们会提出一套理论，解释 alpha-fold 2 实际上是如何
27:39
was interpreting and producing these results.
当时正在解读并生成这些结果。
27:44
OK, that's the two-minute warning.
好，这是两分钟警告了。
27:46
So I am one last question for both of you, quick one.
那我就问你们两位最后一个问题，很快。
27:50
So there's a real chance that AI will make dramatic improvements
所以 AI 真的有可能带来巨大的改善
27:53
in human health in the next immediate future.
在人类健康方面，就在很近的将来。
27:57
I like quantitative predictions.
我喜欢定量预测。
28:00
Best guess, how long until we start seeing AI results
大胆猜一下，还要多久我们才能开始看到 AI 的成果
28:04
in the clinic?
在临床中？
28:06
Vishni, do you want to start?
Vishni，你想先开始吗？
28:09
I mean, the point is, AI results in the clinic.
我的意思是，重点是，AI 在 clinic 里的成果。
28:13
I think it's ill-posed to it's not a AI is being used today
我觉得这么问有点 ill-posed，并不是说今天 AI 没被用上
28:21
already in every part of the drug discovery process.
已经用在 drug discovery process 的每一个环节里了。
28:26
So if you, from that perspective, it's already there.
所以如果你从这个角度看，它其实已经在那儿了。
28:30
But if you are sort of asking me the question,
但如果你有点是在问我这个问题，
28:32
when will we see that 10x acceleration or 100x acceleration
我们什么时候才能看到那种 10x 加速或者 100x 加速
28:37
in the timelines, then the other sort of question
在时间线上，那就是另一类问题了
28:41
is what are we accelerating?
那，我们到底在加速什么？
28:43
Is it sort of doing lead optimization?
是类似于做 lead optimization 吗？
28:45
Is it the target discovery?
是 target discovery 吗？
28:47
Is it sort of the preclinical or the talks sort of work?
是偏 preclinical 的那种，还是 talks 那种工作？
28:52
So I think over the next sort of few years
所以我觉得，在未来大概几年里
28:59
and which is why this effort that we announced today
而这正是为什么，我们今天宣布的这项努力
29:02
is extremely important, we need to tackle
极其重要，我们需要去解决
29:06
some of these hard challenges of biology.
生物学里的一些硬核挑战。
29:09
And only then will we be able to get these true unlocks
只有到那时，我们才能真正获得这些
29:13
of acceleration that dramatically transforms how drug
加速的真正解锁，它会极大地改变 drug discovery
29:17
discovery will happen.
将会发生的方式。
29:19
So AI will be there or like in the clinic
所以 AI 会出现在那里，或者说像在临床里
29:24
and will be impacting sort of things
并且会影响到那种
29:27
that go into the clinic all the time.
一直进入临床的东西。
29:29
But like the actual sort of larger acceleration
但就像那种实际的、更大规模的加速
29:32
that will be unlocked only with a better understanding
只有在对它有更好的理解之后，才会被解锁
29:36
of the biological models that this effort
那些 biological models 的，这项努力
29:40
is trying to sort of create.
算是想要创造出来。
29:43
All right, thanks.
好，谢谢。
29:43
So a real quick question, if we were almost out of time,
那快速问一个问题，如果我们快没时间了，
29:45
did you?
你问了吗？
29:47
Yeah, I mean, I'm tempted to like literally put on my bio hat
对，我是说，我有点想真的戴上我的 bio 帽子
29:51
to answer this question.
来回答这个问题。
29:52
But I think it's like, it's really important
但我觉得吧，这真的挺重要的
29:57
to think about this as you were saying.
像你刚才说的那样去思考这件事。
30:00
What does it mean to push the field forward?
推动这个领域向前发展，到底意味着什么？
30:03
The thing that we really want to see
我们真正想看到的事情
30:05
is those outcomes affected.
是那些结果受到影响。
30:07
And so when will it be that a drug that was made entirely
所以，什么时候会出现一种药，它完全由
30:12
by AI, by an AI or something?
AI、由某个 AI 或类似的东西制造出来？
30:14
I don't know, it's hard to know that.
我不知道，这很难说。
30:17
But I do think we're going to see rapid progress here
但我确实觉得，我们会在这里看到快速进展。
30:21
very, very quickly because all these tools
非常非常快，因为所有这些工具
30:23
are being used and stuff like that.
都在被使用，还有诸如此类的东西。
30:25
And going back to your point about the basic research,
再回到你关于 basic research 的观点，
30:31
we need basic understanding.
我们需要基本的理解。
30:33
One thing I learned a long time ago in my career,
我很久以前在职业生涯中学到的一件事，
30:36
actually back at Google, is sometimes it's easier
其实是在 Google 的时候，就是有时候更容易
30:42
to approach a problem by looking at what
去处理一个问题，方式是看
30:44
is it going to take to make a 10x improvement rather than a 10%
要做成 10x improvement 而不是 10%，需要付出什么
30:49
improvement.
改进。
30:50
And that's not because that's necessarily an easier path.
而这并不是因为那一定是一条更容易的路。
30:54
It's because it allows you to take a broader view
而是因为它能让你有更宏观的视角，
30:58
and see some solutions that you haven't been approaching.
看到一些你之前一直没在尝试的解决方案。
31:02
And what is actually the path to do it?
那真正实现它的路径到底是什么？
31:04
And you would go back and look from first principles.
然后你会回过头，从 first principles 的角度去审视。
31:07
How are we going to do this?
我们要怎么做这件事？
31:08
How are we going to really push this?
我们要怎么才能真正推动这件事？
31:10
So I think both the approaches that both approaches the 10%
所以我觉得，不管是 10% 的方法，还是 10x 的方法，都很有价值，
31:20
and the 10x are valuable, and we should do both.
而且我们应该两种都做。
31:23
I'm just very happy that at BioHob, we have a very beautiful
我只是特别高兴，在 BioHob，我们有一个非常美好
31:30
and lofty mission statement to cure all disease.
而且崇高的使命宣言，那就是治愈所有疾病。
31:33
And if you want to do that, you really
如果你想做到这一点，你真的
31:36
need to go and figure out what is the 10x approach.
得去弄清楚什么是 10x 的方法。
31:39
And so I think it's a really great opportunity
所以我觉得，能去经历这一切，
31:44
to be able to go through that.
真的是一个很棒的机会。
31:46
Awesome.
太棒了。
31:47
Thank you both.
谢谢你们两位。
31:47
Thank you for being in the literal hot seat.
谢谢你们真的坐上了这个 hot seat。
31:49
Thank you.
谢谢。
31:51
Thank you.
谢谢。