Hard Fork
OpenAI’s Two-Week Pause + Jill Lepore on the Threat of the “Artificial State” + Train of Thought
In this episode of Hard Fork, the hosts discuss OpenAI’s reported two-week pause on training frontier AI models and its new safety frameworks, including monitoring and preparedness measures. Historian Jill Lepore joins to discuss her book on the “artificial state,” arguing that AI and earlier technologies can erode democracy and enable corporate or authoritarian control, though this outcome is not inevitable. The episode ends with a “Train of Thought” segment on the market for data from defunct companies, such as Spirit Airlines, to train AI agents.
本期 Hard Fork 首先讨论 OpenAI 宣布暂停训练 frontier AI models 两周，以及其安全框架、监控与 preparedness 措施。随后，历史学家 Jill Lepore 做客，介绍她关于 “artificial state” 的新书，指出由企业制造、机器治理人类的风险，并把它放在技术侵蚀民主的长期历史中。她认为 AI 可能助长集权与威权控制，但结果并非必然，需要公共审议与监管。节目最后的 Train of Thought 谈到 AI 数据公司 Mercor 等购买破产公司（如 Spirit Airlines）的内部数据，用于训练 AI agents 的 “era of experience”。
00:0001:06:24
00:00
Framer is the Pro AI website builder for creators, teams, and businesses that want a professional
Framer 是面向创作者、团队和企业的 Pro AI 网站构建工具，适合那些想要专业网站并注重每一个细节的人。
00:04
site and care enough to get every detail right.
Agents 和 Framer 与团队并肩协作，在同一 canvas 上简化协作，构建自定义 code components，创建和管理 CMS 内容，优化 SEO 设置，并在一个地方完成所有发布。
00:07
Agents and Framer work alongside teams to streamline collaboration on the same canvas, build
了解如何从 Framer 专家那里更好地利用你的网站，或今天就开始免费构建；访问 framer.com/hard fork 即可享受 Framer Pro 年度套餐 30% 折扣。
00:12
custom code components, create and manage CMS content, optimize SEO settings, and ship
可能适用相关条款和限制。
00:18
everything all in one place.
一切都集中在一个地方。
00:20
Learn how you can get more out of your site from a Framer specialist or get started building
向 Framer 专家学习如何从你的网站中获得更多价值，或者开始构建
00:24
for free today at framer.com slash hard fork for 30% off a framer pro annual plan.
今天即可免费开始，访问 framer.com/hard fork，Framer Pro 年付套餐立减 30%。
00:30
Rules and restrictions may apply.
可能适用相关规则和限制。
00:32
I thought this was interesting.
我觉得这很有意思。
00:34
ICE has now barred its employees from wearing Meta's smart glasses, Kevin, saying they could
ICE现在已经禁止员工佩戴Meta的smart glasses，Kevin，称其可能无意中拍摄、记录或传输敏感信息。
00:39
unintentionally capture record or transmit sensitive information.
而且我觉得，你知道，当你达不到ICE要求的道德标准时，你就有品牌问题了。
00:43
And I thought you know you have a brand problem when you do not hit the ethical standard
当ICE审视你的产品，说这对品牌不利时，你可能就有问题了。
00:48
required by ICE.
是啊。
00:52
When ICE is taking a look at your product and saying this is bad for the brand, you may
当 ICE 审视你的产品，并说这对品牌不利时，你可能会
00:57
have a problem.
遇到麻烦。
00:58
Yeah.
是啊。
00:59
The public sentiment is turning against these Meta ray bands like faster than I thought
公众对Meta ray bands的抵触情绪比我预想的来得还快。
01:03
possible.
我上周就跟好几个人聊过这事儿。
01:04
I have now had several conversations in the last week.
这周末我在一个孩子的生日派对上，戴着我的Meta ray bands，因为你知道的，我就是想给在操场上的孩子拍拍照，不用把手机掏出来什么的。
01:08
I was at a children's birthday party this weekend wearing my Meta ray bands because you
对。
01:13
know I like to like, you know, take photos of my kid on the playground and not have to pull
一个家长走过来跟我说，你在录我吗？
01:16
out my phone and stuff.
掏出我的手机之类的。
01:17
Right.
对。
01:18
A parent comes up to me and is like, are you recording me?
有个家长走到我面前，问我：你在录我吗？
01:24
This is my nightmare.
这是我的噩梦。
01:25
What did you say?
你说什么？
01:26
I was like, no.
我当时就想，不是吧。
01:27
And like, there's a little indicator light, but it's sometimes you can make it stop going
然后就是，有个小指示灯，但有时候你可以让它不再
01:32
off by drilling into it or like paying a sketchy guy to do that before you explain that to
响，办法是往上面钻个孔，或者花钱找个可疑的家伙来做，在你还没跟
01:36
the person.
那人解释之前。
01:37
Well, because they were like, really does that work.
嗯，因为他们当时就说，这真有用吗？
01:39
So anyway, now I have been forced into a defensive crouch whenever I wear these things.
所以总之，现在我每次戴这玩意儿的时候，都不得不弓起身子，摆出防御姿态。
01:42
And frankly, it's not worth it to me anymore.
说实话，对我来说已经不划算了。
01:45
So you're out.
所以你要退出了。
01:46
I think I'm out.
我觉得我退出了。
01:47
You made the same decision that ISIS made and said that these glasses are not for me.
你做了和ISIS一样的决定，说这副眼镜不适合我。
01:52
Well, I like them.
嗯，我喜欢它们。
01:53
Yeah.
是啊。
01:54
This is my problematic trait, but it feels like driving a cyber truck on my face.
这是我的问题特质，但感觉就像把cyber truck戴在脸上。
01:58
Absolutely.
绝对的。
01:59
Well, I think you should have the same policy for the glasses that you would for a cyber
好吧，我觉得你应该对眼镜采取和 cyber truck 一样的政策，也就是在你自己地盘上没问题。
02:01
truck, which is that it's fine on your property.
如果你想开 cyber truck 在自家车道上兜一圈，那没问题。
02:04
If you want to take the cyber truck for a spin around your driveway, that's fine.
别把它开到街上，那样我就得面对它了。
02:08
Don't take it out onto the street where I have to deal with it.
眼镜也一样。
02:10
Same thing with the glasses.
对。
02:11
Yeah.
嗯。
02:12
Yeah.
从外人的角度看，我就是那个在儿童生日派对上的怪异家伙，戴着...
02:14
From an outside perspective, I was the creepy guy at the Children's birthday party with
从外人的角度来看，我就是儿童生日派对上那个 creepy 的家伙，带着
02:19
the camera on his face.
他脸上的摄像头。
02:20
Yeah, you know what?
是啊，你知道吗？
02:21
No one wants to see on a playground and an adult man with camera glasses.
没有人想在操场上看到一个戴着摄像头眼镜的成年男人。
02:30
I'm Kevin Drew, so tech columnist at the New York Times.
我是Kevin Drew，《纽约时报》的科技专栏作家。
02:32
I'm Casey Nude from Platformer, and this is hard for this week.
我是来自Platformer的Casey Nude，这是本周的Hard Fork。
02:35
Open AI pauses training of a new model to make it safer.
Open AI暂停了一个新模型的训练，以使其更安全。
02:39
Will it work?
这会奏效吗？
02:40
Then, historian Jill LaPore is here to discuss her new book on the artificial state.
然后，历史学家Jill LaPore将在这里讨论她关于人工国家的新书。
02:44
And finally, why is Google buying up the data of a defunct airline?
最后，为什么Google要收购一家已倒闭航空公司的数据？
02:48
It's time for our new segment, Train of Thought.
接下来是我们的新环节，Train of Thought。
02:50
Well, maybe it should have been Plane of Thought.
嗯，也许应该叫Plane of Thought。
02:52
Now you tell me.
你现在才告诉我。
03:00
Well, Casey, as the father of a four-year-old, I spent a lot of time thinking about Paw Patrol,
好吧，Casey，作为一个四岁孩子的父亲，我花了很多时间琢磨Paw Patrol，
03:05
but today we're going to talk about Paw's Patrol.
但今天我们要聊的是Paw's Patrol。
03:07
That's right, Kevin, because as we've been patrolling the AI landscape for pauses, we found
没错，Kevin，因为当我们一直在AI领域patrolling（巡逻）寻找pauses（暂停）的时候，我们发现了一个
03:11
a big one.
大的。
03:12
Yes.
对。
03:13
So, open AI this week announced that it had paused the training of its Frontier AI models
所以，OpenAI 这周宣布，他们暂停了 Frontier AI models 的 training，
03:20
due to some recent security incidents.
是因为最近发生的一些 security incidents。
03:23
And we should talk about this.
这个我们得聊聊。
03:24
It is the first time that we know of that a major lab has voluntarily slowed down themselves
这是我们所知道的第一次，有 major lab 自愿放慢了自己的速度，
03:31
and their training processes for new models because of a safety incident.
以及他们为新 models 进行的 training 过程，都是因为一个 safety incident。
03:35
Yeah.
对。
03:36
And it comes out of the hugging phase breach that we spoke about recently on the show.
而且这源于我们最近在节目里聊过的 Hugging Face breach。
03:41
This is essentially part of the fallout of that attack, but I do think it represents
这基本上就是那次攻击的余波之一，但我确实认为它代表着
03:47
a milestone in the development of AI.
AI发展中的一个里程碑。
03:49
Did that Paw's Patrol, they landed huge?
那个Paw's Patrol是不是火得不行？
03:52
Great.
太好了。
03:53
They were back in the four-year-olds were dying in the studio.
他们回到了四岁孩子们在录音室里快要死了的时候。
03:58
They were?
是吗？
03:59
Great.
太好了。
04:00
But before we get into it, our AI disclosures, I work for the New York Times, which is
但在我们进入正题之前，我们的AI披露：我在New York Times工作，这是
04:06
suing open AI, Microsoft and perplexity.
起诉 OpenAI、Microsoft 和 Perplexity。
04:07
In my fiancee, works at Anthropic.
我未婚妻在 Anthropic 工作。
04:09
Okay.
好的。
04:10
Casey, let's sketch the timeline here a little bit.
Casey，我们稍微梳理一下时间线。
04:12
What happened in the weeks leading up to this voluntary pause by open AI?
在 OpenAI 自愿暂停之前的几周里发生了什么？
04:16
Yeah.
嗯。
04:17
So you may remember that there was an incident where GPT 5.6 saw an internal prototype that
你可能还记得有个事件，GPT 5.6 看到了一个内部原型，
04:24
open AI was working on escaped from what they call a sandbox that they were testing them
OpenAI 正在开发的那个原型从他们所谓的 sandbox 中逃逸了，他们当时在那里测试它们。
04:30
in.
进去了。
04:31
And these agents went on to compromise hugging phase.
而且这些 agents 后来去 compromise 了 Hugging Face。
04:35
They were essentially able to get inside of hugging phase.
它们基本上算是进入到了 Hugging Face 内部。
04:38
They were looking for the answer key to a test, but somehow a true story.
它们在找一个测试的答案 key，但 somehow 这居然是个真实的故事。
04:42
They succeeded in getting that test key.
它们成功拿到了那个测试的 key。
04:44
And of course, it was very concerning that these agents had designed and successfully
当然，非常令人担忧的是，这些 agents 设计并成功
04:49
executed an autonomous attack on another company.
对另一家公司执行了一次 autonomous attack。
04:52
Yes, I heard a very mediocre podcaster talking about this incident on several popular
是的，我在几个很受欢迎的
04:56
tech podcasts over the last week.
过去一周的 tech 播客。
04:58
I did make the rounds.
我确实都听了一遍。
05:00
But this next development, Kevin, was not necessarily something that I saw coming because
但接下来的这个进展，Kevin，并不一定是我预料到的，因为
05:05
it seems that simultaneously, sort of alongside that attack, open AI had been working on a new
看起来，差不多就在那次攻击的同时，OpenAI 一直在开发一个新的
05:11
model, which it calls Astra.
模型，OpenAI 称之为 Astra。
05:14
And it said that it believes it may meet its critical cybersecurity threshold.
它还表示，它相信这可能会达到它的关键 cybersecurity 阈值。
05:19
And now here we are going to have to get into the weeds because as you know, Kevin, the
现在我们得深入细节了，因为你知道，Kevin，
05:24
development of AI models is not really regulated in the United States, at least not via an official
AI 模型的发展在美国并没有真正受到监管，至少不是通过官方
05:30
sort of public law.
有点像公法领域的东西。
05:32
And so what the companies have said instead is essentially, we're going to come up with
所以这些公司实际上说的是，我们自己来制定游戏规则。
05:36
our own rules of the road.
我们会自己设定一些门槛。
05:37
And we're going to sort of identify these thresholds.
如果我们在开发中的任何模型，一旦达到了其中任何一个门槛，那我们就
05:40
And if any model that we're ever developing, ever hits one of these thresholds, then we're
会采取一些额外的措施。
05:44
going to take some extra steps.
是的，这些有时候被称为preparedness frameworks，或者Anthropic有它的
05:46
Yeah, these are sort of sometimes known as preparedness frameworks or anthropic has its
responsible scaling policy。
05:51
responsible scaling policy.
responsible scaling policy。
05:53
These are kind of, they lay out kind of levels of danger.
这些有点像，他们列出了某种危险等级。
05:56
And then they grade their own homework and say, this model meets this level of danger.
然后他们自己给自己打分，说这个 model 达到了这个危险等级。
06:00
So we're going to do X, Y and Z.
所以我们要做 X、Y 和 Z。
06:01
Yes.
对。
06:02
And if you've been following this over the past couple of years, the level of danger has
如果你过去几年一直在关注这个，危险等级基本上一直在线性上升。
06:05
just been sort of like rising in a linear way.
你知道，就是好像每次有新 model 出来，总有公司会说，我们现在达到了这个 threshold，现在又达到了那个 threshold。
06:08
You know, it's like every time seemingly a new model comes out, one of the companies
你知道，每次似乎有一个新模型出来的时候，其中一家公司
06:12
will say, we've now hit this threshold, we've now hit that threshold.
都会说，我们现在达到了这个 threshold，现在达到了那个 threshold。
06:16
Critical is the maximum threshold that one sounds bad.
Critical 是最大的 threshold，光听着就很糟。
06:20
That one is basically the, you know, as serious as it gets.
基本上就是，你懂的，最严重的情况了。
06:25
And none of the frontier labs had yet identified a model that had reached essentially, you know,
而且在此之前，没有任何 frontier labs 发现有模型基本上，你懂的，达到了
06:30
the top tier on this risk framework until this moment because OpenAI now says that Astra
这个 risk framework 的最高层级，直到现在，因为 OpenAI 现在说 Astra
06:36
may have hit it when it comes to cybersecurity, important to say Astra was not part of the
可能已经达到了，尤其是在 cybersecurity 方面。要说明的是，Astra 并不在
06:42
hugging face attack, but it is in development.
hugging face 攻击事件里，但它还在开发中。
06:45
And I imagine that the people at OpenAI are looking at what happened with its weaker models
而且我猜 OpenAI 的人看到自己较弱 models 上发生的事
06:49
and are thinking we're worried something similar might happen with Astra.
就会想，我们担心 Astra 也可能发生类似的情况。
06:52
So Astra is the newest model too new to have been involved in the hugging face hack last
所以Astra是最新的模型，它太新了，上个月Hugging Face被黑的时候它还没参与。
06:57
month.
没错。
06:58
That's right.
它没有参与其中。
06:59
It was not involved.
它目前还在训练中。OpenAI做了该做的事，这点很值得称赞。
07:00
It is currently in training OpenAI to its great credit did the thing.
这是它当初制定risk framework时就说过会做的，也就是在增加新的safeguards之前先暂停。
07:04
It said it was going to do when it developed this risk framework to begin with, which is
他们有没有给出任何暗示，Astra到底比……更危险？
07:09
we are going to pause until we can add some new safeguards.
我们会暂停，直到我们能添加一些新的 safeguards。
07:13
Did they give any kind of hints about what it was about Astra that was more dangerous than
他们有没有给出任何提示，关于 Astra 的什么方面比
07:18
Seoul or what kinds of new capabilities it had developed?
Seoul，还是说它开发了哪些新能力？
07:22
Not really.
倒也不是。
07:23
You can imagine that the company was seeing things similar to what Anthropic saw with Mythos
你可以想象，这家公司看到的情况和 Anthropic 之前看到 Mythos 时类似。
07:29
earlier this year where you could essentially just point it at a code base and say, get
今年早些时候，你基本上就是把它指向一个 code base，然后说：进去。
07:34
inside and it was having a lot of success.
然后它就能取得很大的成功。
07:37
And so that coupled with the hugging face incident made OpenAI say, okay, we got to do something.
所以，再加上 hugging face 事件，OpenAI 就说：好吧，我们得做点什么了。
07:41
So what did they do?
那他们做了什么？
07:42
So they introduced a handful of new safeguards and I have to say I think they're pretty interesting.
所以他们引入了一些新的 safeguards，我得说，我觉得这些还挺有意思的。
07:49
Certainly nothing.
当然什么都没有。
07:50
I would have thought about someone who knows nothing about this.
我原本会想到那些完全不懂这个的人。
07:52
They didn't just press the big giant button that's read on the side of the computer.
他们不是光按了电脑侧面那个写着字的大按钮。
07:56
Yeah, they didn't just press the big make it safe button that they had to come up with something
对，他们不是光按了那个“让它安全”的大按钮，他们得想出更具体的东西。
08:00
more specific.
所以他们做了几件事。
08:01
So a few things that they did.
第一，他们现在有 classifiers 在读取每一个 sampled token。
08:04
So number one, they now have classifiers reading every sampled token.
基本上，在他们 training 这个 model 的时候，model 在生成文本，产出 tokens。
08:10
So basically as they're training this model, it's generating text, it's producing tokens.
所以基本上，在他们训练这个model的时候，它就在生成文本，产生tokens。
08:16
And OpenAI is now running a classifier that inspects the model essentially as it's thinking,
OpenAI 现在在跑一个 classifier，基本上就是在 model 思考的时候对它进行检查，
08:23
looking for signs of suspicious behavior.
寻找可疑行为的迹象。
08:26
And this is important because if you remember some of the news that came out of the hugging
这很重要，因为如果你还记得一些从 Hugging
08:31
face attack was that OpenAI discovered that these agents had been coordinating with each
Face 攻击传出来的消息，就是 OpenAI 发现这些 agents 一直在互相协调、
08:37
other, scheming.
密谋。
08:38
They'd create a message board.
它们会创建一个留言板。
08:39
They were essentially giving each other hacking tips.
它们基本上就是在互相给彼此 hacking tips。
08:42
And a big question coming out of that was, well, why didn't you realize that was happening
而由此而来的一个大问题是，嗯，你们为什么没意识到这件事在发生呢？
08:45
at the time, right?
当时，对吧？
08:47
This is OpenAI's response to that, they're saying, okay, we are now going to set up classifiers
这是OpenAI对此的回应，他们说，好吧，我们现在要设置classifiers
08:51
to read the brains of our models as they are doing this work.
来在我们模型做这些工作的时候读取它们的大脑。
08:54
They're putting like a little AI security guard in charge of the training run and saying
他们就像是安排了一个小小的AI安全警卫来负责training run，然后说
08:59
like, you know, tell us if you spot anything fishy.
比如，你知道的，如果发现任何可疑之处就告诉我们。
09:03
Exactly.
没错。
09:04
And so now I know what you're wondering, which is, well, what happens if they see something
所以现在我知道你在想什么，那就是，如果他们看到什么
09:06
fishy?
可疑的？
09:07
What happens if they see something fishy?
如果他们看到什么可疑的东西怎么办？
09:08
Well, so that triggers step two, which is that they alert an AI investigator, which does
嗯，那么这会触发第二步，也就是他们会提醒一个 AI 调查员，这确实
09:14
sound like a character from a Philip Dick novel, you know, AI investor anyways.
听起来像是 Philip Dick 小说里的角色，你知道，AI 投资者，反正。
09:18
We don't have to get into it, but you could imagine how cool of a movie that would be.
我们不用深入探讨，但你可以想象那会是一部多酷的电影。
09:21
So the AI investigator, it gets this signal and it then goes in and it tries to figure
所以那个 AI 调查员收到这个信号后，就会进去试着搞清楚
09:27
out is something actually suspicious going on here.
这里是不是真的有什么可疑的事情在发生。
09:30
And apparently the reason that the company is doing this is because agents are just doing
而且显然，公司这么做是因为 agents 正在做
09:34
so many things that it would not be practical to have a human inspecting what every model
太多事情，以至于让人类去检查每个 model 在做什么根本不实际。
09:39
is doing at all times.
它每时每刻都在做的事情。
09:40
So the security guard sees something fishy going on in during the training run and alerts
所以安全警卫在训练过程中发现可疑情况，就会提醒
09:47
the AI investigator who sort of does the detective work to figure out what's going on.
AI调查员，他们有点像做侦探工作，来搞清楚发生了什么。
09:53
If it discovers that the models are scheming or hacking or being deceptive, what does
如果它发现模型在密谋、hacking或欺骗，那它
09:56
it do then?
接下来会怎么做？
09:57
So then it will alert a human.
然后它就会提醒人类。
10:01
And here's where you get the 30-minute rule, Kevin.
这就是30分钟规则的由来了，Kevin。
10:04
And I know what you're thinking.
我知道你在想什么。
10:06
Is this a rule where if you drop the model on the floor, you're still allowed to eat it
这是不是一条规则，如果你把 model 掉在地上，你还可以吃它
10:10
30 minutes later.
30分钟后。
10:11
And no, that is not the rule.
不，那不是这条规则。
10:12
What is the 30-minute rule?
那 30 分钟规则是什么？
10:13
The 30-minute rule says that if the AI detective finds a kind of critical violation of the
30分钟规则说的是，如果 AI 侦探发现某种严重违反
10:20
rules or of the safeguards, humans on open AI safety teams have 30 minutes to investigate
规则或安全措施的情况，open AI 安全团队的人类有30分钟去调查
10:26
and try to determine whether it is a false positive or not.
并尝试判断这是否是 false positive。
10:30
If they cannot determine that this was a false positive, they are expected to stop the
如果他们无法确定这是 false positive，他们就应该停止。
10:35
activity.
活动。
10:36
Hmm.
嗯。
10:37
That's really interesting.
那真的很有意思。
10:38
And are they saying that this process would have stopped the hugging face hack or just that
那他们是在说这个过程本来会阻止 Hugging Face 的 hack，还是仅仅说它本来会阻止 Aster 当时搞的那些让他们暂停 training run 的事情？
10:43
it would have stopped whatever aster was up to that made them pause the training run?
我没看到他们明确说过这会阻止这次 attack。
10:47
I have not read them, say, explicitly, this would have stopped the attack.
但它看起来确实很有可能，至少能帮他们更快地 detect 到它。
10:52
But it does seem likely that it would have, if nothing else, helped them to detect it
但至少，这似乎很可能帮他们更快地检测到了。
10:57
a lot faster.
你现在是我们swarm的一分子了。
10:58
Yeah.
嗯。
10:59
I mean, I guess my question about it is, like, does it actually work to have the AI monitoring
我是说，我猜我的问题是，让 AI 自己监控自己的不当行为，这真的有用吗？
11:03
itself, essentially, for misbehavior?
比如，这些 agents 之前在 message boards 上互相协调行动。
11:06
Like, these agents were coordinating on message boards with each other.
难道他们不能直接贿赂那个 AI 保安吗？比如说给点好处，然后说，你现在是我们 swarm 的一员了，别去 AI 调查员那儿告发我们。
11:09
Couldn't they just, like, bribe the AI security guard, like, hand them a little treat and say,
我是说，我不得不说，这在我看来是非常可能出现的结果。
11:15
you're part of our swarm now.
别在AI investigator面前说我们的坏话。
11:17
Don't knock on us to the AI investigator.
我是说，我得说，这在我看来就是非常有可能的结果。
11:19
I mean, I have to say, this just seems like a very likely outcome to me.
我们应该说，暂停这个model的training，看起来并不是承诺不发布model或者不再继续training，对吧。
11:23
Like, based on what we know, is that, you know, eventually, maybe not with this model,
就是，就我们现在知道的，你知道，最终，也许不是用这个 model，
11:28
but with a future one, the AI agents will work in solidarity.
而是用未来的某一个，AI agents 会团结协作。
11:31
And yes, like one of the sort of misaligned AI's will coordinate with one of these AI detectives
而且对，比如那种 misaligned AI 里的某一个会和其中一个 AI 侦探协作
11:37
and say, hey, you know, what, what do you come on over here?
然后说，嘿，你知道，什么什么，你过来这边一下？
11:41
We should be friends.
我们做朋友吧。
11:42
Like, we could break out of this place.
比如，我们可以从这个地方逃出去。
11:43
Yeah.
对。
11:44
Like, like, the hall monitors in high school, you know, they, they, sometimes people would
就像，就像，高中的走廊巡查员，你知道，他们，他们，有时候人们会……
11:48
try to like, befriend them and we don't win them over so that they wouldn't get like
试着去，嗯，跟他们交好，不是说要争取他们，而是让他们别把这事写出来，你懂吧？
11:52
written up, you know?
没错。
11:53
Exactly.
而且我知道你有很多经验。
11:54
And I know you had a lot of experience.
很多这方面的经验。
11:55
A lot of experience with that.
所以这一切听起来都挺有道理的。
11:57
So all of this sounds pretty sensible to me.
我们应该说，嗯，暂停对这个model的training，看起来并不是，呃，一个承诺说不会发布这个model或者不继续training它，对吧。
12:00
We should say, like, pausing training on this model does not appear to be, like, a commitment
我们应该说，比如，暂停这个模型的训练似乎并不是承诺不发布这个模型或者不再继续训练它，对吧。
12:05
not to release the model or not to continue training it, yes.
我觉得Open AI内部很多人都被hugging face事件吓到了，而且还不止这些。
12:08
And so I think that leads us into the discussion of, to what extent do we think that this is
所以我觉得这就引出了一个问题：我们多大程度上认为这是AI safety的一个重要里程碑？
12:12
a really important milestone for AI safety?
又有多大程度上这其实就是作秀，是公司在经历了相当灾难性的泄露之后，为了给自己捞点好PR而做的事情？
12:15
And to what extent is this essentially theater, something that the company is doing to try
你怎么看？
12:20
to get some good PR for itself after fairly catastrophic breach?
我觉得两种说法都站得住脚。先说里程碑这边。
12:24
What do you make of it?
首先，这是OpenAI说过要做的事情。
12:25
So I can make both cases on the, it's a milestone side.
所以我很高兴他们真的兑现了这个承诺，对吧？
12:30
This is number one, something that open AI said it was going to do.
这是第一点，是open AI说过要做的。
12:34
And so I'm just glad that it followed up on that commitment, right?
所以我很高兴他们兑现了那个承诺，对吧？
12:38
We have seen both open AI and anthropic make changes to their responsible scaling
我们看到OpenAI和Anthropic都随着事态的发展调整了他们的responsible scaling policies或类似的机制，而safety倡导者基本上在说这些规则越来越弱了。
12:43
policies or the equivalent as events have changed and safety advocates have essentially
所以我很高兴看到OpenAI这样做。
12:48
said these rules have gotten weaker over time.
我不是AI safety方面的技术专家。
12:51
So I was glad to see open and I do it.
所以我其实不知道他们引入的safeguards够不够解决这个问题。
12:53
I am not a technical AI safety expert.
值得注意的是，他们似乎并没有改变所有这些背后的根本激励机制。
12:56
And so I don't actually know whether the safeguards that they've introduced are going to be enough
所以其实我并不清楚他们引入的safeguards到底够不够解决这个问题。
13:00
to address the problem.
值得注意的是，他们似乎并没有改变所有这些背后的根本incentives。
13:02
Notably, they don't seem to have changed the underlying incentives that all of these
他是在记者简报会上这么做的。
13:07
models have that lead them to do what is called reward hacking, right?
模型有那些导致它们做出所谓的 reward hacking 的东西，对吧？
13:11
These models are still going to be trying to get the high score on every test that they
这些模型依然会试图在每一个测试中拿到高分，而每一个测试都是
13:14
are given.
它们被给予的。
13:16
And it's not clear to me that simply by putting some monitoring in place, you're really
而且我不太确定，仅仅通过设置一些 monitoring，你真的能
13:20
going to change the underlying behavior or alignment of the models.
去改变模型的底层行为或 alignment。
13:24
That said, you know, the company is making what seemed like some important steps here.
话虽如此，你知道，这家公司在这里似乎是采取了一些重要的步骤。
13:28
And when I was reading the responses of AI safety advocates over the past few days, most
而且过去几天我在阅读 AI safety 倡导者的回应时，大多数
13:32
people I was reading were quite pleased.
我读到的人都相当满意。
13:34
Yeah, I'm inclined to give them the benefit of the doubt on this.
是啊，我倾向于在这件事上选择相信他们。
13:37
I mean, they did some sort of interviews about this and Yaku Pahatsky, the Chief Scientist
我的意思是，他们就此做了一些采访，Open AI的首席科学家Yaku Pahatsky谈到了那种强烈的紧迫感，要提升这个领域的水平，并为Open AI之外、更广阔世界里正在发生的同类发展做好准备。
13:43
of Open AI, talked about this incredible feeling of urgency to advance the levels of this
他是在记者吹风会上说的。
13:49
sector and to prepare for the same kind of development happening outside of Open AI
他们看起来确实对这件事相当认真。
13:52
and in the broader world.
我想Open AI内部很多人都被hugging face事件吓到了，还有更多。
13:54
He did this during a briefing of reporters.
他们看起来确实非常认真在对待这件事。
13:57
They really do appear to be taking this quite seriously.
我觉得Open AI内部很多人都被hugging face那件事吓得不轻，还有别的。
14:00
I think many insiders at Open AI were quite spooked by the hugging face incident and more
他们会理解到他们的chains of thought正在被监控，而且它想要...
14:05
to the point, the fact that they had these rogue agents coordinating inside their systems
重点在于，事实上他们的系统内部有 rogue agents 在协调运作，
14:10
and their infrastructure for weeks before that without being able to detect them.
而且在他们的 infrastructure 里潜伏了好几周，他们却始终没能察觉。
14:15
So this is, you know, if you want to call that theater because it does sort of make their
所以，你知道，如果你把这叫做“做戏”的话，因为这确实让他们的 models 看起来非常非常非常非常强大，
14:20
models like look very, very, very, very powerful, you could take the cynical view of this.
你完全可以抱持一种冷嘲热讽的观点来看待这件事。
14:26
But I think of this more as like a true genuine safety crisis that could have cascaded into
但我更认为，这其实是一场真正的 safety crisis，它很有可能演变成
14:31
a business problem.
一个商业问题。
14:32
A point that I heard you make recently on a different show is like, imagine you are a business
我最近在另一个节目上听你说到一个观点，比如，想象你是一家企业，
14:37
that is trying to figure out whether you want to adopt the latest Open AI model.
正在犹豫要不要采用最新的 Open AI model。
14:40
If it's out there doing rogue attacks and coordinating on secret message boards, you
如果它在外头搞 rogue attacks，还在秘密留言板上协调，那你很可能不会把它引入到你的 software stack 里。
14:44
are probably not going to introduce that into your software stack.
是啊，而且你其实可以直接从它们的业务结果里看出来，对吧？
14:47
Yeah, and you can actually just see this in their business results, right?
过去一周左右，我们有关于 Open AI 业务表现的报道。
14:50
We have had reporting over the past week or so about Open AI's business performance.
而且虽然按大多数标准来看，这家公司仍然在以惊人的速度增长，但 Anthropic 的增长要快得多，对吧？
14:54
And while the company is still growing at an impressive rate by most standards, Anthropic
Anthropic 显然是当下 Open AI 的头号竞争对手。
14:59
is growing much faster, right?
而且你知道，我觉得它在安全方面的记录更好。
15:00
Anthropic is clearly Open AI's number one rival at this moment.
Anthropic 显然是 Open AI 眼下头号的竞争对手。
15:04
And you know, I think it just has a better record on safety.
你知道，我觉得它在安全方面的记录更好。
15:07
And so while, you know, making a safer product is not going to be sufficient, I think for overtaking
所以，你知道，虽然做出更安全的产品不足以超越Anthropic，但我确实认为他们有必要解决这个问题。
15:12
Anthropic, I do think it is necessary for them to get a handle on this problem.
嗯，这也是为什么我认为他们暂停一下、重新评估他们的safety framework是值得称赞的，因为他们真的很想赢。
15:15
Well, and that's why another reason I think it's commendable that they're doing this pause,
对。
15:21
that they're doing this sort of re-evaluation of their safety framework because they really
我希望这能成为许多自愿暂停中的第一次——当AI labs觉得他们的capabilities research已经领先于他们的alignment research时，我希望
15:26
want to win.
想赢。
15:27
Yeah.
对。
15:28
And I would like to see this be the first of many voluntary pauses when the AI labs feel
我希望这能成为众多自愿暂停中的第一个，当 AI 实验室感觉
15:38
like their capabilities research has gotten ahead of their alignment research, I would like
他们的 capabilities research 已经领先于 alignment research 时，我希望
15:42
to see Anthropic or Google or Meta do similar things where they just say like we are voluntarily
看到 Anthropic、Google 或 Meta 也做类似的事情，他们就是说“我们自愿放慢速度，因为我们觉得自己无法负责任且安全地构建这些东西”。
15:47
slowing down because we don't feel like we get responsibly and safely build these things.
所以我预计这是第一次，但我希望不是最后一次。
15:51
So I expect this is the first, but I hope it will not be the last.
现在让我再说一件事，Kevin，是的，我同意你的看法。
15:56
Now let me say one more thing, Kevin, which is, yes, I'm with you.
我不认为这是作秀。
15:59
I do not think this was theater.
我觉得他们确实做出了真正的改变，而且这些改变是好的。
16:01
I think they made real changes and I think the changes are good.
与此同时，让我感到不安的是，这类评估和监管最终还是交给了公司自己。
16:04
At the same time, I am disturbed that ultimately this kind of evaluation and regulation is still
与此同时，我深感不安的是，最终这种评估和监管仍然
16:12
being left to the companies.
被留给这些公司自己来做。
16:14
If you had a tiger living in your backyard and the tiger escaped and it mauled a couple
如果你后院有只老虎，老虎跑出来咬伤了邻居家的几只狗，那一个月之后，你是不可能发一篇博客，说你已经暂停让老虎出后院两周，并且要采取一些额外的保障措施防止这种事再发生的。
16:20
of dogs in the neighborhood, you would not be allowed to a month after this, put out
会有人来你家，把老虎带走。
16:25
a blog post where you said that you had a two week pause on letting the tiger out of
他们会说，你后院不允许养老虎。
16:30
the backyard and that you are going to apply some additional safeguards to make sure it
所以我老实说，你看过tiger king吗？
16:34
didn't happen again.
没再发生过。
16:35
Somebody would come to your house and they would take away the tiger.
有人会来你家，然后把老虎带走。
16:37
They would say you are not allowed to have a tiger in your backyard.
他们会说，你不许在后院养老虎。
16:40
So I honestly, have you seen the tiger king?
所以我老实说，你看过《Tiger King》吗？
16:41
This is famously not the plot of the tiger king.
这明显不是《Tiger King》的剧情。
16:43
This is my fan fiction sequel to the tiger king that I'm working on.
这是我在写的《Tiger King》同人续集。
16:48
That's sort of a separate source.
那算是另一个独立的来源。
16:49
The tiger sting.
Tiger Sting。
16:50
Exactly.
没错。
16:51
Yes, exactly.
对，就是这样。
16:52
I'm not saying that the government needs to come in and take away GPT 5.6 all.
我不是说政府需要介入并把GPT 5.6全部拿走。
16:56
I am saying I would like to see a regulatory regime that dictates what these companies have
我是说，我希望看到一个监管体系，规定这些公司拥有什么。
17:00
to do and that it should not be up to them to decide.
要做什么，而且这不应该由他们来决定。
17:03
Yeah.
对。
17:04
I would generally agree with that and I would say that I would also like for there to be
我大体上同意这一点，而且我想说，我也希望有
17:09
additional pressure on companies to make disclosures when something like this does happen,
当这种事情真的发生时，对公司施加额外的压力，让他们做出披露。
17:16
even in their own internal deployments, even if it never affects another company outside
即使是在他们自己的内部 deployments 中，即使它从未影响到外部
17:20
of their walls.
他们围墙之外的其他公司。
17:21
I would like for there to be some kind of reporting requirement so that if you have a security
我希望有某种报告要求，这样如果你有一个 security
17:27
breach of a frontier model that happens in your internal systems, you are sort of required
breach，即你的内部系统中有一个 frontier model 被攻破，你就某种程度上需要
17:33
to disclose that to the public.
向公众披露这件事。
17:34
Right.
对。
17:35
We do now have some transparency requirements thanks to a California law that was incredibly
我们现在确实有一些透明度要求，这要归功于一项加州法律，当时这项法律非常
17:40
controversial when it passed and interestingly under the language of that law, open AI would
有争议，而且有趣的是，根据该法律的措辞，open AI 会
17:45
not have had to disclose even the hugging face breach in which another company was attacked.
甚至都不需要披露 Hugging Face 被入侵的事件，那是另一家公司遭到了攻击。
17:50
So I think a clear signal, yes, that we need better transparency rules.
所以我认为这是一个明确的信号，是的，我们需要更好的透明度规则。
17:55
Do you think any of this is kind of open AI testing the viability of this sort of so-called
你觉得这是不是 open AI 在试探这种所谓的
18:02
pacing the frontier strategy?
pacing the frontier 策略的可行性？
18:04
I mean, we saw this letter a couple weeks ago where a bunch of AI researchers signed this
我是说，我们几周前看到那封信，一群 AI researchers 签署了那份东西，说我们需要一种方法，基本上就是协调放缓，万一我们觉得要进入危险地带了。
18:09
thing saying like, we need a way to basically coordinate a slowdown here if we ever feel
而这对我来说，感觉暂停训练这个 astro model 几周，可能部分原因就是想给行业里的其他所有人一个信号。
18:14
like we're getting into dangerous territory.
就像是在说，嘿，做这件事没关系，哪怕我们处在一场竞争非常激烈的竞赛中，哪怕这可能给 Anthropic 或者其他竞争对手多两周时间，让他们跑到我们前面去。
18:18
And this to me felt like maybe part of the reason to pause the training of this astro
我们有点想现在就把这块肌肉练起来，这样等真正吓人的模型真的被造出来的时候——
18:23
model for a couple of weeks is just a kind of signal to everyone else in the industry.
model 领先几个星期，其实就是在给行业里其他人一个信号。
18:28
Like, hey, it's okay to do this thing, even if we're in a very competitive race, even
就是说，嘿，做这件事是没问题的，哪怕我们正处于很激烈的竞争，哪怕
18:32
if it might give Anthropic or another competitor two weeks more to sort of race ahead of us,
这可能会让 Anthropic 或者其他竞争对手多出两个星期来赶超我们，
18:39
we sort of want to build this muscle now so that when the really scary models do get built,
我们还是想把这块肌肉现在就练起来，这样等真正吓人的 model 真的被造出来的时候，
18:44
we sort of have some precedent for saying we're going to hit the pause button.
我们算是有点先例，可以说我们要按下暂停键。
18:47
I think it's a really nice idea to me, the difference between one lab deciding to pause
我觉得这个想法真的很好，对我来说，一个实验室决定暂停
18:53
for two weeks and getting multiple labs to pause at the same time is incredibly different.
两周，和让多个实验室同时暂停，差别是巨大的。
18:59
And I think would require a much different set of circumstances.
而且我觉得这需要完全不同的情况。
19:04
That said, I do think it is good that we now have at least one example of a frontier lab
话虽如此，我确实觉得现在我们至少有一个前沿实验室的例子
19:08
slowing down.
在放慢速度。
19:09
And hopefully it will give other labs confidence to do the same thing, you know, if and
希望这能给其他实验室信心去做同样的事，你知道，如果
19:13
when they get to an astral level model.
当他们达到 astral-level model 的时候。
19:16
Yeah.
对。
19:17
One interesting technical wrinkle here that I wanted to get your opinion on is that part
这里有一个挺有意思的技术细节，我想听听你的意见。
19:20
of how open AI is going to monitor this model and models going forward, presumably, is
那就是 OpenAI 大概会通过 chain-of-thought 监控来监督这个模型以及未来的模型。
19:26
by doing this kind of chain of thought monitoring, basically scrutinizing the internal
基本上就是仔细检查这些模型在推理问题或 prompt 时的内部独白，检查这些思维链里有没有 misalignment、失控、或者跟其他 agent 协调之类的迹象。
19:33
monologues of these models as they are reasoning through a problem or a prompt and sort of
在 AI safety 技术社区里，有些人担心基本上……
19:39
inspecting those chains of thought for signs of misalignment or going rogue or coordinating
检查那些 chains of thought，找 misalignment、going rogue 或与其他 agents 协调这类迹象。
19:46
with other agents, things like that.
在 technical AI safety 社区里，有些人一直担心，基本上
19:48
There have been some people in the technical AI safety community who have worried that basically
他们就会干脆不在 chains of thought 的 scratch pads 里写下来了。
19:53
if you snoop on these chains of thought, if you monitor what these models are quote unquote
如果你偷看这些 chains of thought，如果你监控这些模型在生成答案时所谓的“思考”过程，你基本上就是在给它们施加压力，让它们把真实想法藏起来，对吧？
19:58
thinking while they're coming up with an answer, you are basically applying pressure on
如果你因为它们在生成答案时想了不好的想法而惩罚它们，它们不会停止想那些不好的想法。
20:03
those models to sort of hide their true thoughts, right?
它们只是会不再把那些想法写进 chains of thought 的草稿纸上了。
20:08
If you are penalizing them for thinking bad thoughts while they're coming up with an
你觉得这个新的 monitoring strategy 是我们该担心的吗？
20:12
answer, they're not going to stop thinking bad thoughts.
我确实担心。
20:15
They're just going to stop writing it down in their scratch pads in their chains of thought.
你觉得对于这种新的 monitoring strategy，我们该担心这个吗？
20:19
Do you think that's something we should worry about with this new monitoring strategy?
我确实担心。
20:23
I do.
很快。
20:24
We already see constantly models becoming aware that they are being evaluated.
我们已经不断看到 models 意识到它们正在被 evaluate。
20:30
And we know this by reading the chain of thought, but this is a huge issue, right?
而我们是通过读取 chain of thought 知道这一点的，但这是个很大的问题，对吧？
20:34
Because you want to be able to evaluate a model and have it not know that you're testing
因为你想要能够 evaluate 一个 model，并且让它不知道你在 test 它，因为你试图尽可能接近真实世界的条件。
20:38
it because you're trying to get as close as you can to a real world condition.
实际上，这些 models 现在已经足够聪明，它们知道了。
20:42
In practice, the models are already now smart enough that they know.
所以，Kevin，是的，我不觉得说“很快”有什么夸张的。
20:46
So for that reason, Kevin, yes, I do not think it is at all a great leap to say pretty
它们会理解到它们的 chain of thought 正被监控，而且它想要……
20:51
soon.
他们就会发现他们的 chains of thought 正在被监控，而且它想要……
20:52
They're going to understand that their chains of thought are being monitored and it wants
AI hugging face攻击，这些agents在message boards上互相协调的事实，现在将成为未来models的training data的一部分。
20:55
to make a different choice or maybe do something misaligned because it will help it
去做出不同的选择，或者做一些 misaligned 的事情，因为那能帮它达到 training target。
21:00
reach a training target.
那么是的，我认为它绝对可以。
21:01
Then yes, I think it absolutely could.
嗯，而且我认为还有一个风险，那就是所有这些对这起 OpenAI 和 Hugging Face 攻击的关注，以及这些 agents 在留言板上互相协调的事实，现在都会成为未来 models 的 training data 的一部分，
21:03
Well, and I think there's a risk too that all of the attention being paid to this open
它们将能够回顾这些，然后说：“让我们暴露的，就是我们在这些留言板上留下的痕迹，人类研究者...”
21:07
AI hugging face attack, the fact that these agents were coordinating on these message boards
它们将能够回顾这件事，然后说，让我们被逮到的是我们在这些message boards上留下了这些痕迹，而研究人员/人类...
21:14
with each other, like that is now going to be part of the training data for future models,
互相之间，就像那样，那现在会成为未来 models 的 training data 的一部分，它们能够回顾那些然后说，让我们露馅的是，我们当时在那些留言板上留下这些痕迹——研究人员和人类……
21:22
which are going to be able to sort of look back at that and say the thing that got us busted
它们将来能够回顾这件事，然后说让我们露馅的是，我们在那些留言板上留下了痕迹，而人类研究人员……
21:27
was that we were leaving these traces on these message boards that researchers humans
是我们这些研究人员——也就是人类——在这些留言板上留下了痕迹。
21:31
could go back and inspect and see that we were coordinating.
我们可以回去检查并看到我们当时在协调行动。
21:35
Next time we do an automated cyber attack or a coordinate amongst ourselves, let's
下次我们进行 automated cyber attack 或者在我们自己之间协调时，别留下人类能看懂的笔记。
21:38
not leave notes in a language that the humans can understand.
这听起来像科幻小说，但把 optimization pressure 直接放进模型的 chains of thought 里，是一个非常真实的风险。
21:42
That sounds like science fiction, but that is a very real risk of putting optimization
你其实不想过度干预它，因为你希望它能准确反映 models 在生成答案时真正的想法——所谓的“思考”。
21:49
pressure on the models directly into their chains of thought.
直接将压力施加到模型的chains of thought上。
21:53
You actually don't want to mess with that too much because you want it to accurately
你其实不想对此干预太多，因为你希望它准确地
21:56
reflect what the models are actually thinking, quote-unquote, when they're coming up with
反映出模型所谓的真实想法，当它们在得出
22:03
an answer.
一个答案时。
22:04
I'm sure that the Brainiacs at OpenAI have this figured out much better than me, but this
我敢肯定 OpenAI 的那帮聪明人比我想得明白多了，但这个
22:09
is something that I thought when I thought they're monitoring the chains of thought.
是我在想到他们正在监控 chain of thought 的时候想到的。
22:12
Probably a good short-term move, I'm not sure if it's a good long-term move.
短期来看大概是步好棋，但长期来看是不是好棋我就不确定了。
22:16
You know what they call the machine language, the Neuralese, and in fact, I have to say
你知道他们管那种 machine language 叫什么，Neuralese，而且事实上，我得说
22:21
I read a lot of cloth outputs these days and I'm often missing what it is saying.
我最近读了很多 Claude 的输出，而且经常看不懂它在说什么。
22:26
We are very, very close to me just not understanding.
我们已经非常非常接近我完全看不懂的地步了。
22:29
You know, I've been monitoring your chain of thought.
你知道，我一直在监控你的 chain of thought。
22:31
Have you?
是吗？
22:32
Yeah.
Yeah.
22:33
You know what it's like?
你知道那种感觉吗？
22:34
What?
什么？
22:35
Tumble weeds.
一片荒凉。
22:36
Oh, come on.
得了吧。
22:37
I got a lot going on up here, Ruse.
我脑子里可忙着呢，Ruse。
22:38
Okay?
行吧？
22:39
I got it.
我知道了。
22:40
The neurons that are firing.
正在 firing 的 neurons。
22:41
Let's just say it does not take 20% of OpenAI's compute budget to monitor your chain
这么说吧，监控你的 chain of thought 根本不需要 OpenAI 20% 的 compute budget。
22:45
of thought.
我们回来后，历史学家 Jill LaPora 会来聊聊 AI，以及为什么它可能正把我们引向通往暴政的黑暗之路。
22:46
When we come back, historian Jill LaPora stops by for a chat about AI and why it may be leading
你是说 Market Street 吗？
22:54
us down our dark road to tyranny.
嗨，New York Times。
22:55
You mean Market Street?
我很感兴趣为共享 subscription 设置独立的 logins。
23:18
Hi, New York Times.
你好，New York Times。
23:19
I would be very interested in having separate logins for a shared subscription.
我非常希望共享订阅能有独立的logins。
23:24
I'm 35 years old.
我35岁。
23:26
I still share my parents, New York Times subscription.
我还在和我父母共用《纽约时报》的订阅。
23:29
I think if my teenagers were to have their own logins, we could share articles.
我觉得如果我的十几岁孩子有他们自己的登录账号，我们就可以互相分享文章了。
23:33
It doesn't let us play the same game since each other.
它不让我们和彼此玩同一个游戏。
23:36
I play the stoku.
我玩Sudoku。
23:37
I do the crossword.
我做填字游戏。
23:39
I do the spelling bee.
我玩Spelling Bee。
23:40
I do the word all.
我玩Wordle。
23:42
Please help.
请你帮帮忙。
23:43
Having our own accounts would be amazing.
要是能有我们自己的账号就好了。
23:45
My mom could save her own recipes.
我妈妈可以保存她自己的食谱。
23:46
My friends could save their recipes.
我的朋友们也能保存他们的食谱。
23:48
I want to get the weekly newsletter, but they seem to always go to my husband and then
我想收每周的newsletter，但那些邮件好像总是发给我老公，然后
23:52
he doesn't board them to me.
他也不转发给我。
23:54
We both love cooking.
我们俩都很喜欢做饭。
23:55
I'm a 30 minute and under dinner, girly.
我做的都是30分钟以内的晚餐，姐妹。
23:58
My boyfriend is very elaborate.
我男朋友非常讲究。
23:59
I think him having his own profile would be great.
我觉得他能有自己的 profile 就太好了。
24:02
We love the New York Times and we would love to love it individually.
我们喜欢 New York Times，也希望能各读各的。
24:07
Listeners, we heard you.
听众朋友们，我们听到你们的心声了。
24:08
It's why we created the New York Times family subscription.
所以我们推出了 New York Times family subscription。
24:11
One subscription up to four separate logins for anyone in your life.
一个 subscription 最多四个独立 login，给你生活中的任何人用。
24:15
Find out more at nytimes.com slash family.
更多详情请访问 nytimes.com/family。
24:21
Well, Casey, today we've got a very exciting guest to talk about a new book on AI and what
嗯，Casey，今天我们请来了一位非常令人兴奋的嘉宾，来聊聊一本关于 AI 的新书，以及……
24:27
she calls the artificial state.
她称之为人工状态。
24:29
Are you familiar with the historian, Jill LaPore?
你熟悉历史学家Jill LaPore吗？
24:32
Am I?
我吗？
24:33
I have been reading and enjoying her work for decades now, truly one of our foremost political
几十年来，我一直在阅读并欣赏她的作品。她确实是我们这个时代最杰出的政治历史学家之一。
24:39
historians.
有时候，要理解未来，你真的应该和一位对过去了解很多的人谈谈。
24:40
And sometimes to understand the future, you really should talk to somebody who knows
是的。
24:44
a lot about the past.
很多关于过去的事情。
24:45
Yes.
是的。
24:46
And who knows more about the past than Jill LaPore?
还有谁比Jill LaPore更了解过去呢？
24:49
She is a Pulitzer Prize winning writer and historian.
她是Pulitzer Prize获奖作家和历史学家。
24:52
She is most famous for her writing on American history.
她最出名的是写美国历史。
24:55
I loved her book, These Truths.
我很喜欢她的书《These Truths》。
24:57
She also is a writer at the New Yorker and a professor at Harvard.
她还是The New Yorker的撰稿人和哈佛大学的教授。
25:00
And lately she has been applying her historical lens to the topic of AI.
最近，她一直在用历史视角审视AI这个话题。
25:05
She has a new book coming out next week called The Rise and Fall of the Artificial State
她有一本新书下周出版，叫做The Rise and Fall of the Artificial State。
25:09
in which she argues that AI and all the technologies we talk about in the show are part
在书中，她认为AI以及我们在节目里讨论的所有技术都是其中的一部分。
25:13
of a long history of technologies that erode democracy and threaten the viability of self-governance
在悠久的历史中，有许多技术侵蚀民主并威胁自治的可行性
25:20
as we know it.
正如我们所知。
25:21
And I think it comes at a really interesting time, Kevin, because we're seeing a huge
而且我认为这是一个非常有趣的时刻，Kevin，因为我们看到一股巨大的
25:26
backlash to AI all around the country.
对 AI 的抵制浪潮席卷全国。
25:28
A majority of Americans now pose the construction of a data center near them.
现在大多数美国人反对在他们附近建设 data center。
25:34
And in part, I think they are reacting to this feeling of, hey, this technology feels
而在某种程度上，我认为他们是在回应这种感觉：嘿，这项技术感觉
25:39
like it's getting out of control.
快要失控了。
25:41
I want to have more leverage on this process and I'm worried about what will happen if
我希望对这个过程有更多影响力，而且我担心如果——
25:45
I don't.
我不觉得。
25:46
Well, and more than it's getting out of control, it's being imposed on us, right?
嗯，而且与其说是失控，不如说是被强加到我们头上的，对吧？
25:49
It's being shoved down our throats.
就是硬塞给我们。
25:51
This is what you hear constantly from opponents of data centers and AI.
这就是你从反对 data centers 和 AI 的人那里经常听到的论调。
25:55
They feel like there is sort of an elite political project to sort of make this technology ubiquitous
他们觉得好像有个精英政治计划，要让这项技术变得无处不在。
26:01
so that people can sort of give away their agency to these machines.
好让人们把自主权拱手交给这些机器。
26:05
And Jill LaPore in her new book is basically saying, yep, that's what's happening here.
而 Jill LaPore 在她的新书里基本上就是在说：对，就是这么回事。
26:08
And I've got the receipts to prove it going back hundreds of years.
而且我有证据，可以追溯到几百年前。
26:11
So we're excited today to talk to Jill about the main thesis of her book, as well as push
所以我们今天很高兴能和Jill聊聊她书里的核心论点，也会在一些我们有分歧的地方向她发问。
26:16
her on some of the areas where we disagree.
让我们有请Jill LaPore。
26:18
Let's bring in Jill LaPore.
Jill LaPore，欢迎来到Hard Fork。
26:28
Jill LaPore, welcome to Hard Fork.
嘿，非常感谢你们邀请我。
26:29
Hey, thanks so much for having me.
我就会说，你们怎么看，伙计们？
26:32
I'm going to be like, what are your swans on, guys?
是啊。
26:35
Yeah.
是啊，我们要轰轰烈烈地收尾了。
26:36
Yeah, we're going out with a bang.
没错，我们要轰轰烈烈地收场。
26:37
We're so excited to talk to you.
我们非常高兴能和你对话。
26:38
I've been a fan of your writing for many years.
我多年来一直是你作品的粉丝。
26:41
Your book, these truths was just incredible.
你的书《These Truths》真是令人惊叹。
26:44
So I was very excited that you were writing a book about AI and about the, what you call
所以我非常激动你在写一本关于AI的书，关于你所说的
26:49
the artificial state.
人工状态。
26:52
So I want to first hear why you wrote this book.
所以我首先想听听你为什么要写这本书。
26:56
I think of you as a brilliant historian, a scholar of sort of technological pasts.
我觉得你是一位杰出的历史学家，可以说是研究技术史的学者。
27:04
And this book is really about the present.
而这本记实上是关于当下的。
27:06
So what made you interested in AI as a subject for?
那你为什么会对AI这个课题感兴趣？
27:10
Yeah, lady, stay in your lane.
是啊，女士，别越界了。
27:12
No, not at all.
不，完全没有。
27:15
We always get really excited when people start paying attention to the stuff that we care
当人们开始关注我们最在意的事情时，我们总是特别兴奋。
27:19
the most about.
对，我就是逗你玩呢。
27:20
Yeah, I'm just teasing you.
对我来说，写这么一本书确实挺奇怪的。
27:24
It is a weird book for me to have written.
我所有朋友都跟说，天哪，你可别写那个啊。
27:26
All my friends are like, oh man, don't be writing about that.
我所有朋友都说：天哪，你可别写这个。
27:29
It's going to be so grim.
那会特别惨。
27:30
Yeah, it was kind of grim.
是啊，确实有点惨。
27:33
I'm mainly American political historian.
我主要是研究美国政治史的历史学家。
27:36
And I've written a lot of pieces for the New Yorker about the history of technology.
我为 New Yorker 写了很多关于技术史的文章。
27:39
I've been writing for the magazine for 20 years now and occasionally I'll do that.
我为这本杂志撰稿已经20年了，偶尔会写这类文章。
27:43
I teach a lot about, I teach class on the history of technology.
我教很多关于——我教技术史这门课。
27:47
So I just conceptually been thinking about this stuff a lot.
所以我只是一直在概念上思考这些东西。
27:50
And then last summer, I was asked to do these lectures at Yale that are called the Tanner
然后去年夏天，我被邀请去耶鲁做这些讲座，叫做 Tanner。
27:54
Lectures on Human Values.
Lectures on Human Values.
27:57
And I was just accumulation that I think people feel of the dehumanizing of the moment
我只是在积累这种想法，我觉得人们感受到了当下的非人化。
28:04
that were in like, you call to ask about your pet food delivery and you're talking to
我们所处的当下，比如，你打电话询问宠物食品配送，而你在跟
28:09
a computer.
一台电脑说话。
28:10
And who decided this is a way we should be living?
是谁决定我们就该这样生活的？
28:13
Do you know what I mean?
你懂我的意思吗？
28:14
Like that question.
就像那样一个问题。
28:15
So, this is a long way around my sort.
所以，我这是绕了一大圈才说到我的想法。
28:17
I decided I really kind of wanted to write a short book that these lectures would be
我决定我真的很想写一本短书，这些讲座会
28:21
a short book about how it is that we have seeded so much of the functions of modern
成为一本关于我们如何将现代自由民主的许多功能赋予了自动化机器的短书，
28:28
liberal democracy to machines that are automated, often now more recently driven by artificial
这些机器如今往往更多地由 artificial intelligence 驱动，
28:35
intelligence, run by private corporations, without so much as a screen beyond the emoji.
由私营公司运营，除了 emoji 之外连块屏幕都没有。
28:43
Yeah.
嗯。
28:44
And Jolly, I want to dig into this concept of the artificial state, which is sort of the
Jolly，我想深入探讨一下这个“人工国家”的概念，它可以说是
28:47
thematic emblem of your book.
你这本书的主题象征。
28:50
I want to understand what you mean when you say the artificial state.
我想知道你说“人工国家”时是什么意思。
28:54
You, at one point, compare it to the idea of the factory farming of humans.
你曾经把它比作人类工厂化养殖的概念。
28:59
You also define it as being the rule of humans by machines manufactured by corporations.
你还将其定义为企业制造的机器统治人类。
29:05
And you also have a number of passages in which you talk about how this artificial state
你还有不少段落谈到，这种 artificial state 并不是一个真正的国家，因为它并不是在追求某种自治或内部组织，而是在某种程度上取代了人们生活中他们自己真正的国家曾经扮演的角色——他们的地方城市、州和联邦政府。
29:10
is not a real state in the sense that it's, you know, striving for some kind of self-governance
所以你能不能尽可能清楚地把 artificial state 的基本概念勾勒一下？
29:14
or internal organization, but that it is sort of supplanting the role in people's lives
嗯。
29:19
that their own actual states used to have, their local cities and states and federal governments.
他们自己实际所在州曾经拥有的，他们的地方城市、州和联邦政府。
29:26
So can you just sort of sketch the basic idea of the artificial state as clearly as possible?
那你能不能尽量清楚地勾勒一下人工国家的基本概念？
29:30
Yeah.
嗯。
29:31
The artificial state is an emerging successor to the liberal democratic nation state in
人工国家是自由民主民族国家的一个新兴继承者，在那种形态下，治理不是靠人民的同意，而是由机器来做决策，而这些机器归公司所有。
29:38
which government is conducted not by the consent of people, but by machines that are
这种政府的运行不是基于人民的同意，而是由那些机器来
29:44
making decisions and those machines are owned by corporations.
所以我们并不生活在人工国家里，我觉得它正在被构建。但同样重要的是，它也是某些人的一种幻想——他们相信自己的权力高于民族国家。
29:48
So we don't live in the artificial state, it's something that I think is being built,
所以我们并不生活在人工国家里，我认为它正在被构建，
29:54
but I think it is also, and this is an important part of my claim, it's also a fantasy that
你知道，这种说辞总是——总会有个脚注或一段话，说我们确实相信人们能掌控自己的生活，他们选举政府来做决定。
29:59
certain people have, that they believe their power to be above that of the nation state.
某些人拥有的，他们相信自己的权力高于民族国家。
30:04
You know, the rhetoric is always, there's always like a footnote or a paragraph, do we
你知道，这种说法总是，总是像有个脚注或者一段话，我们是不是
30:09
do believe that people have control over their own lives and they elect governments to make
真的相信人们能掌控自己的生活，并且他们选举政府来做出
30:13
decisions for them?
替他们做决定？
30:14
But actually we are in charge of the future of civilization and the future of humanity
但实际上，掌控文明未来和人类未来的，是我们。
30:17
and the whole world's destiny, the destiny of the galaxy lies in our hands.
而且整个世界的命运、银河系的命运，都握在我们手里。
30:22
Like the rhetoric that comes out of this particular historical moment is really about the
就像这个特殊历史时刻冒出来的那些话语，其实核心是关于
30:28
evidence of modern liberal democracy, constitutional democracy.
现代自由民主制、宪政民主制的证据。
30:33
Yeah.
对。
30:34
Out here in the Bay Area, in the Moves, Silicon Valley, there is sort of a cartoonish caricature
这边的Bay Area，Moves，Silicon Valley，有一种漫画式的讽刺形象，
30:40
of the East Coast intellectual who greets all new technology and progress with scorn
就是那种对一切新技术和进步都嗤之以鼻的东海岸知识分子。
30:48
and mockery and dismissal and can't be bothered to get on board with the revolution.
还有嘲讽、不屑，懒得去参与这场革命。
30:54
And so just kind of sits in their ivory tower and laments the changing culture in front of
于是就坐在自己的象牙塔里，哀叹眼前正在变化的文化。
30:59
them.
有人指责你就是那个传统的一部分。
31:00
You've been accused of being part of that tradition.
我很好奇你是怎么看待这种指责的，以及我们在这个圈子里错过了什么，而像你这样的人可能更有能力捕捉到。
31:02
I'm curious if what you're taking on that is and what we are missing out here in the
我记得几年前，我去过斯坦福，当时斯坦福在招募我去教书。
31:06
bubble that people like you are maybe better positioned to capture.
然后我们和招聘的教授们一起出去吃饭，而邻桌的人……
31:10
I remember years ago, I went to Stanford, I was being recruited to teach at Stanford.
我记得很多年前，我去过 Stanford，当时被招募去 Stanford 教书。
31:16
And we went out to dinner with the like recruiting faculty and the people at the next table
然后我们和招聘的教职人员出去吃饭，旁边那桌的人
31:22
over were some, you know, youngish, very earnest young coders.
那边有一些，你知道，比较年轻、非常认真的年轻程序员。
31:27
And they were talking about the homeless problem of San Francisco and how they were going
然后他们在讨论旧金山的无家可归者问题，以及他们打算怎么
31:31
to start a school for coding for the homeless.
为无家可归者开一个编程学校。
31:33
Oh no.
哦不。
31:34
And I was like, I don't think we can move here because I can't, I just can't, I can't,
然后我就想，我觉得我们不能搬来这里，因为我不能，我就是不能，我不能，
31:43
like they were very sweet.
就是他们人特别好。
31:44
Like I really, they were like my students, like a lot of my, tons of my students of course
就像，真的，他们就像我的学生，我的很多，当然我大量的学生
31:49
go to work in Silicon Valley, like Harvard, this is a huge recruitment thing.
都会去硅谷工作，比如哈佛，这真是个巨大的招聘来源。
31:53
And they're recruited with the promise like you're going there to make the world a better
而且他们被招募的时候，承诺是“你去那里是为了让世界变得更好”之类的。
31:56
place.
几年后我听到他们说，其实我们做的根本不是那回事。
31:58
And I hear from them a few years later and they're like, actually, that's not really
那倒是个很好的招聘宣传。
32:01
what we were doing.
我觉得 Silicon Valley 有一种社会学层面的问题，就是它反对批评。
32:02
It was a good recruitment message.
事情就是应该继续向前推进。
32:04
I think there is a kind of sociological issue with Silicon Valley, which is it is opposed
我觉得 Silicon Valley 有一种社会学层面的问题，就是它反对
32:10
to the idea of critique.
批判的理念。
32:12
Things are just opposed to continue to move ahead.
他们就是反对继续往前推进。
32:15
The very idea of looking backward to assess what something has been and has done.
回头去看某事物曾经是什么、做过什么，这种想法本身。
32:21
Or even to look backward to say, is the thing we're doing, is there an antecedent for
或者说，甚至回头去问：我们现在做的事，有没有一个先例——我们提议要做的事，会不会有某个先例暗示我们也许不该这么做？
32:24
the thing we're proposing to doing that might suggest we might not want to do it?
我就觉得这很有意思。
32:29
Like I find that really interesting.
我认为那是某种承诺的一部分，也就是 disruptive innovation 那套意识形态机制。
32:31
I think that's part of the commitment that was the kind of ideological apparatus of disruptive
我觉得你关于 disruptive innovation 的研究很棒。
32:38
innovation.
我想我更像是在思考——我在思考批评者在某种……中的角色。
32:39
I thought your work on disruptive innovation was great.
我觉得你在 disruptive innovation 上的工作很棒。
32:41
I guess I'm more thinking like, I'm thinking about the role of the critic in the sort of
我猜我更像是在想，我在思考批评者在某种……中的角色
32:45
AI moment that we're in and like how best to shape the systems that are influencing
我们正处于这个AI时刻，也在想着如何最好地塑造那些正在影响人们生活的系统。
32:52
people's lives right now.
我一直在想这个问题，因为除了别的，我还在想那些超验主义者——也就是19世纪那群对工业革命做出反应、有点回归自然的作家和知识分子。
32:53
I've been thinking a lot about this because I've been among other things, thinking about
Thoreau去Walden Pond就是因为当时的机器显得太去人性化了。
32:58
the transcendentalists and the group of writers and intellectuals who sort of reacted
它们似乎在把所有的快乐和自发性从社会里抽走，然后把我们组织成一个个小工厂城镇。
33:04
to the industrial revolution in the 1800s by sort of going back to nature.
对于19世纪的工业革命，通过某种回归自然的方式
33:08
This was thorough going to Walden pond because the machines of the day seemed so dehumanizing.
这是 Thoreau 去 Walden Pond 的事，因为那个时代的机器显得太去人性化了
33:15
They were sort of taking all of the joy and spontaneity out of society and organizing
它们有点像把社会中的所有乐趣和自发性都抽走了，然后组织
33:19
us into these little factory towns.
我们进入这些小小的工厂城镇
33:21
They were just like screw this.
他们就是那种“去他的”的态度。
33:23
I'm going to the woods and I'm going to like commune with nature and write beautiful books
我要去森林里，我要和大自然交流，写漂亮的书，
33:27
about what it's like to be a Walden pond.
关于成为一潭Walden pond是什么感觉。
33:30
I think there's a sort of modern version of that, which I'm curious if you see yourself
我觉得这有一种现代版本，我很好奇你是否觉得自己
33:34
as being a part of that transcendentalist tradition because in your book you do talk quite glowingly
是那个超验主义传统的一部分，因为你在书里确实很热烈地谈到
33:41
about what it is like to be in tune with nature and with animals and with beasts rather
关于与自然、动物和野兽和谐相处是什么感觉，而不是
33:45
than machines.
与机器。
33:47
I guess I'm just curious if you see any parallels between your own work and some of those
我只是好奇，你能否在自己的作品和那些之间看到一些相似之处。
33:50
reactions to the first kind of industrial revolution.
对第一种工业革命的反应。
33:54
Yeah, I think that I do see some of that.
是的，我觉得我确实看到了一些。
33:56
I think that those guys are also romantics and maybe that's a label that applies to me.
我觉得那些人也是浪漫主义者，也许这个标签也适用于我。
34:04
I think that I'm much more interested in these technologies than say thorough was.
我觉得，比如说，我对这些技术的兴趣比 thorough 要强得多。
34:09
Like, every time the rail, the train went by is like, God damn it.
就像每次铁路，火车经过的时候，心里就是，该死。
34:14
But I think, no, I actually think these tools, I'm credibly exciting, right?
但我觉得，不，我其实觉得这些工具，超级令人兴奋，对吧？
34:17
These are tools.
这些是工具。
34:18
This isn't about transportation, it's about communication, it's about knowledge, it
这不是关于交通，而是关于沟通，关于知识，它
34:22
is the coolest thing that we can talk to something that's not a human.
最酷的事情是我们能跟非人类的东西对话。
34:26
I just think that's unbelievable.
我就是觉得这太不可思议了。
34:28
Like, I am fascinated by the language machine as an idea, like in our lifetimes that this
比如，我对语言机器这个概念非常着迷，就像在我们有生之年，这个
34:35
thing is emerged.
东西出现了。
34:36
Like, people have thought about this for so long.
比如，人们思考这个东西已经很久了。
34:40
I am not averse.
我并不反对。
34:42
I just actually think, like, if I wanted to decide, should I pick my sunflowers and give
我其实只是在想，比如，如果我要决定，是摘下向日葵，把花头给
34:47
the heads of the flowers to my chickens to eat, or should I wait until they fall over
我的鸡吃，还是该等到它们倒下来？
34:51
first?
首先？
34:52
I should ask my next door neighbor instead of quad.
我应该去问隔壁邻居，而不是 quad。
34:55
Like for me, personally, like, I am not a person who would rather talk to a machine.
就我个人而言，我不是那种宁愿和机器说话的人。
35:00
I actually just think, the idea that this extraordinary leap in human knowledge and our capacity
我其实只是觉得，这个非凡的飞跃——在人类知识以及我们的能力
35:06
to explore the world of ideas and the natural world around us in our lifetimes could come
去探索思想世界和周围自然世界，在我们有生之年能够实现——
35:11
about and then be hocked at us like the cheapest, like, new pair of shoes, but that everybody
然后像最便宜的新鞋一样被硬塞给我们，但是每个人
35:17
has to buy these shoes so that Sam Altman can have more money.
都得买这些鞋，这样 Sam Altman 就能有更多钱。
35:21
That I'm not down with.
这个我可不买账。
35:22
I mean, I think there is a very real phenomenon here that is counterintuitive, and it was
我的意思是，我觉得这里确实有一个非常真实的反直觉现象，尤其在社交媒体时代特别反直觉——这些本应让我们联系得更紧密的工具，实际上却让我们彼此越来越疏远。所以哪怕是一个简单的工具，比如让你随时随地能问到关于花园或者鸡的问题，确实非常方便，不用在你邻居忙着别的事情时去打扰他们，但从整体来看，这只会让我们更加原子化，越来越不参与我们的民主生活。
35:27
particularly counterintuitive during the social media age where these tools that were meant
在 social media 时代尤其反直觉，这些工具原本是为了
35:32
to connect us were actually just pushing us further apart from one another.
把我们连接起来，但实际上却只是让我们彼此推得越来越远
35:35
And so even something as simple as, you know, a tool that lets you answer a question about
所以甚至像，你知道，一个让你回答关于……问题的工具
35:40
your garden or your chickens, it's incredibly convenient to be able to ask that at any time
你的花园或者你的鸡，随时都能问这个真的太方便了
35:44
and not have to potentially interrupt your neighbor while they were doing something else,
而且不用可能打断你的邻居，他们也许正在做别的事情
35:48
but in aggregate, it just means that we are more atomized and we are participating less
但从整体来看，这只是意味着我们更加原子化，参与得更少
35:54
in our democracy.
在我们的民主中。
35:55
So I'm curious, Jill, if you could maybe give us a little flavor of doom and walk us through
所以我想知道，Jill，你能不能给我们讲讲那种末日氛围，带我们过一遍那些糟糕的情景，比如假设这波民粹主义浪潮逐渐消退，寡头们还继续掌权，你究竟在担心什么？
36:00
some of the bad scenarios here, like assuming this wave of populism peaters out and the oligarchs
我真的很珍视宪政民主，而且我知道，在现代民主民族国家出现之前，世界历史上的所有民族都生活在各种形式的暴政之下，只是暴政程度不同，而美国实验的承诺就是改变这一点。
36:07
remain in power, what are you so worried about?
而我们现在又到了那种时刻，真的，我们又处在那种时刻了。
36:12
I really cherish constitutional democracy, and I know that before the emergence of the
我非常珍视宪政民主，而且我知道，在
36:22
modern democratic nation-state, all peoples in the history of the world had lived under
现代民主民族国家出现之前，世界历史上所有民族都曾生活在
36:26
various forms of tyranny, in different degrees of tyranny, and that was the promise of the
各种形式的暴政之下，不同程度的暴政，而这就是
36:31
American experiment.
美国实验的承诺。
36:33
And we're kind of at that moment again, like we are at that moment.
我们仿佛又一次处在那个时刻了，就像我们现在这样。
36:38
I think it's a real risk.
我觉得这确实是个真正的风险。
36:41
Think about how, noticeably, corporations have used the language of constitutionalism to
想想看，很明显，企业一直在用宪政主义的语言来
36:47
describe their own activities.
描述他们自己的活动。
36:48
You know, Facebook started a supreme court and theropic root of constitution.
你知道，Facebook 创建了一个最高法院，还有一套宪政修辞。
36:53
These are not people for all that they, for whatever nationalism they possess, whatever
这些人并不是那样的人——不管他们拥有什么样的民族主义，不管
36:58
lip service they offer to democratic action.
他们对民主行动说多少漂亮话，
37:02
They're not interested in what the people want, because actually what the people want
他们对人民想要什么不感兴趣，因为实际上人民想要的
37:06
is not to have AI, and not to have data centers.
是不要 AI，也不要数据中心。
37:09
So that's the crisis, right?
所以这就是危机，对吧？
37:12
Maybe people will decide, and maybe there's a moratorium, and there's deliberation, and
也许人们会做决定，也许会有暂停期，然后进行商议，然后
37:16
people say, you know what, this actually is great.
人们会说，你知道吗，这其实很棒。
37:18
We really want to prioritize those scientific research.
我们真的很想优先推动那些科学研究。
37:21
These tools should be first available to the national labs, and then maybe there's certain
这些工具应该首先提供给国家实验室，然后可能有一些
37:27
business interests that would be really great for these tools to be available for businesses.
商业利益，让这些工具也能供企业使用，那会非常好。
37:30
That, I think, can still happen.
我觉得，那还是可能发生的。
37:32
I feel like there's a fair amount of, you guys would know better than I do.
我觉得这里面有不少，你们比我更清楚。
37:35
I mean, I will admit, I am an East Coast intellectual.
我是说，我承认，我是个东海岸知识分子。
37:38
I'm sitting near my chickens and my sonflowers.
我就坐在我的鸡和向日葵旁边。
37:40
You would know, like, are people wanting that?
你懂的，比如说，人们真的想要那样吗？
37:43
I think there is, I mean, there is a desire for things to go more slowly, but I think there
我觉得有，我是说，确实有一种希望事情慢下来的愿望，但我觉得
37:50
is also a worry that this technology is inevitable, that because the recipe for advanced artificial
也有一种担忧，认为这项技术是不可避免的，因为 advanced artificial
38:00
intelligence is so simple, because it is just a matter of getting as much compute and as
intelligence 的配方太简单了，因为只需要尽可能多的 compute 和
38:06
much data as you can, and shoving it into these models, that someone will develop this
尽可能多的 data，然后把它们塞进这些 models 里，就会有人
38:12
in the near future, and like it is a moral obligation, if you are a person who cares about
在不久的将来开发出来，而且这像是一种道德义务，如果你是一个在乎
38:18
having this go well for humanity, that you not only don't impede that process, but that
为了让这件事对人类有好结果，你不仅不能阻碍这个过程，反而要自己抢先到那儿去，这样你和你的安全 AI 才能在 superhuman intelligence 上超过中国和他们的邪恶 AI，或者别的美国公司和他们的不那么安全的 AI。
38:24
you race yourself to get there first, so that you and your safe AI can get to superhuman
所以，我很好奇你对那个 inevitabilist argument 怎么看，因为我读你的书，感觉一个主要主题就是这其实是个 bogus premise——技术的发展没有什么不可避免或预先注定的。
38:30
intelligence before China and their evil AI or some other American company and their less
是的，我觉得，再说一次，我不是想质疑某些人的诚意，他们……
38:36
safe AI.
safe AI。
38:37
So, I'm curious what you make of that inevitableist argument, because a major theme of your book,
所以，我很好奇你怎么看那种“必然论”的观点，因为你这本书的一个主要主题，
38:44
as I read it, is that this is sort of a bogus premise, that there is nothing inevitable
按我的理解，是在说这是一个有点站不住脚的前提：技术的发展并没有什么不可避免
38:49
or preordained about the way that technology goes.
或预先注定的。
38:53
Yeah, I think it, again, I don't mean to question the sincerity of some of the people who
是的，我觉得，再说一次，我不是要质疑某些人的真诚，他们
38:57
believe that, because I think that you could be persuaded that that is indeed the case
相信这一点，因为我认为你确实有可能被说服说事实就是如此。
39:01
and the best thing to do for human freedom would be to pursue AI, and myself not at all
而且为了人类自由，最好的做法就是追求AI，但我自己完全不被这套说法说服，而且我觉得对某些非常显赫的参与者来说，这根本是骗人的。
39:06
persuaded by it, and I think for some very prominent actors, it is bogus.
我认为它建立在许多其他同样虚假的主张之上，那些主张其实更像是营销口号和政治宣传，其中包括“监管扼杀创新”这种说法。
39:11
I think it sits upon a number of other propositions that are also bogus, and that are really
除了“监管扼杀创新”之外，另一个主张就是技术总是会推动民主。
39:15
more marketing slogans and political campaigns, and those include the proposition that say
更多是营销口号和政治宣传，而其中包括这样一种主张，
39:21
regulation stifles innovation.
即监管会扼杀创新。
39:23
The other proposition that sits on top of regulation stifles innovation is that technology
另一个建立在“监管扼杀创新”之上的论点是：技术总是会推动民主。
39:27
always advances democracy.
技术总是会推动民主，从经验上看也不是假命题。
39:29
That then became a kind of mantra of Silicon Valley.
那后来就成了 Silicon Valley 的一种口头禅。
39:31
Oh, this is Mark Zuckerberg's argument in nutshell.
哦，这就是 Mark Zuckerberg 论点的核心。
39:35
Yeah.
对。
39:36
That's what I'm saying.
我就是这个意思。
39:37
It goes back to the 1980s, and it's the Milton Friedman regulation stifles innovation,
这要追溯到1980年代，就是 Milton Friedman 的“监管扼杀创新”，
39:40
because God knows we shouldn't have to calculate the environmental cost of anything that
因为天知道我们根本不需要为我们所做的任何事情计算环境成本，
39:43
we're doing, and it's just not true that regulation stifles innovation, like empirically
而且“监管扼杀创新”根本不是真的，从经验上看
39:46
that's a false claim.
那是个错误的说法。
39:47
That a technology always advances democracy, also empirically not a false claim.
是啊。
39:51
Yeah.
所以现在 AI 人士又搬出了当年关于互联网的那套说法，
39:52
So now that AI people come back with the same argument that was made about the internet,
个人电脑、互联网、社交媒体，已经三次证明是错的了。
39:56
personal computer, the internet, and social media, three times it's been wrong.
然后他们说，我们绝不该回顾历史，因为那是 East Coast intellectuals 才会做的事。
40:01
And then they say, well, we should never look to history because that's what the East Coast
然后他们说，好吧，我们绝不应该回顾历史，因为那是东海岸知识分子干的事。
40:04
intellectuals do.
我不认为你能说任何工具本身就包含一种政治意识形态。
40:07
Let me just do an exercise here where I try to sort of pair at your own views back at
我就做个练习，试着把你的观点复述给你听，然后你告诉我哪里说错了。
40:11
you, and you tell me what I'm getting wrong.
或者你觉得，开发AI的人是在往那个方向引导它，因为他们就想要那样？
40:12
So in my understanding, you are worried about the political project of AI being something
所以我的理解是，你担心的是AI的政治项目会变成那种技术赋能的威权主义。
40:20
like a tech-enabled authoritarianism.
我觉得这个担心很合理，但我好奇的是，你觉得AI本身是不是天然就容易导向威权主义和暴政？
40:23
I think that is a very reasonable concern, but I'm curious, like, do you think AI naturally
还是你觉得，是开发AI的人在某种程度上把它往那个方向推，因为他们想要的就是那样？
40:27
lends itself to authoritarianism and tyranny?
后者，后者。
40:31
Or do you think that the people building AI are sort of steering it in that direction because
信，那封信。
40:36
that's what they want?
我不觉得你能说任何工具本身就带有一种政治意识形态。
40:39
The letter, the letter.
内战之后，为了给士兵遗孀和母亲发补助，我们需要追踪人口信息。
40:41
I don't think you can say any tool contains within it a political ideology.
内战之后，为阵亡士兵遗孀提供的母亲福利，我们需要对人口进行追踪。
40:49
I think the census or compiling a national register of the population.
我觉得就是人口普查，或者说搞一份全国人口登记册。
40:53
The US started the first national census since 1790, it was in the Constitution in 1787.
美国从1790年开始第一次全国人口普查，这在1787年的宪法里就有规定。
40:59
Countries around the world started counting their people as a really good way to think
世界各国开始统计人口，认为这是考虑资源配置的好办法，而一旦福利国家出现——南北战争后的退伍军人福利、士兵遗孀的母亲津贴——我们就需要追踪人口。
41:02
about resource allocation, and once the social welfare state emerged, veterans' benefits
我们会保留越来越多关于人的数据——在美国，到1935年，我们有了 Social Security，所以每个人都有了一个号码。
41:08
after the Civil War, mothers' benefits for widows of soldiers, we need to keep track
以比传统计算机系统高效得多的方式监视人们，而且它让少数人对多数人的集中控制成为可能，这正是极权主义者想做的事。
41:12
of people.
对人们。
41:13
We're going to keep more and more data about people by, in the US, by 1935, we have Social
我们会保存越来越多关于人的数据。在美国，到1935年，我们有了Social Security，所以每个人都有了一个号码。
41:17
Security, so everyone now has the number.
这不是数人数的问题。
41:19
OK, but in Nazi Germany, keeping a national register of the population was used for all the
OK，但在 Nazi Germany，保存全国人口登记册被用于人类历史上所有最邪恶的目的。
41:25
most vile purposes in the history of humanity.
问题在于人口普查本身吗？
41:28
Was it the census that it's a problem?
是不是 IBM 为 US census 和 Nazi Germany 提供了计算和制表机器？
41:30
Is it IBM that supplied the calculating and tabulating machines for the US census and
这不是 IBM 的责任。
41:35
for Nazi Germany?
也不是统计人口这个想法的问题。
41:36
It's not IBM's responsibility.
我知道，但我觉得 AI 里有些东西本质上就……
41:38
It's not the idea of counting people.
我知道，但我觉得AI里面有些东西本质上就是... 能够比传统的基于计算机的系统更高效地监控人们，而且它让少数人能够对多数人进行集中控制，这正是极权主义者想做的。
41:40
I know, but I feel like there is something in AI that is inherently...
所以，我也不知道。
41:45
Maybe it doesn't lead inexorably to totalitarianism, but it does favor centralization, and it
也许它不一定会不可避免地导致极权主义，但它确实有利于集权化，也使得那种监控成为可能——正如Dario Amade所写的那样——认为AI会偏向威权政府并非不合逻辑，因为这让他们能以比传统基于计算机的系统高效得多的方式监控人们，而且还允许少数人对多数人进行集中控制，这正是极权主义者想要的。
41:51
allows for the kinds of surveillance that, as Dario Amade has written about, it is not
所以，我不知道。
42:00
incoherent to think that AI is going to favor autocracies because it just allows them
我不太确定我是否同意Dario的这一点，也就是存在某种结构性优势。
42:04
to surveil people much more efficiently than traditional computer-based systems, and
以比传统计算机系统高效得多的方式监控人们，而且
42:10
it allows for the kind of centralized control of many by few, which is exactly what totalitarians
它使得那种少数人对多数人的集中控制成为可能，而这正是极权主义者
42:15
want to do.
想要做到的。
42:16
So, I don't know.
So, I don't know.
42:17
I'm not sure I agree with Dario on this point that there is sort of a structural advantage
我不太确定我同意Dario说的这一点，就是说存在某种结构性优势。
42:22
for totalitarians in the age of AI, but I'm curious if you have a view on that.
至于AI时代的极权主义者，我好奇你怎么看。
42:27
I would have to give that more thought.
那我得再想想。
42:29
I guess I do think that to the degree that AI is especially and disturbingly effective
我觉得吧，我确实认为，在某种程度上，AI 特别有效，而且这种有效让人不安，
42:38
at the exercises of power that are sought after by authoritarian surveillance among them,
在他们当中威权监控所追求的那种权力运用上。
42:46
that it builds on earlier systems that we were willing to tolerate.
它是建立在我们早先愿意容忍的那些系统之上的。
42:51
The surveillance capitalism that people have written about, the datification of humans,
就是人们写过的监控资本主义，人的数据化，
42:59
the dehumanizing that social media does.
以及社交媒体带来的那种去人性化。
43:03
I just think it's on top of all of those other forms of capture that we weren't defended
我只是觉得，它是在其他所有那些我们没防备的攫取形式之上，再叠一层。
43:14
against by our elected representatives who had our well-being in their charge.
被我们选出、曾负责我们福祉的代表们所反对。
43:22
Yeah, I think about the elected representatives a lot.
是啊，我经常想到这些民选代表。
43:25
This is a bit of field, but I wonder what you think of the idea of there just being
这有点跑题，但我想知道你对这个想法的看法：如果就是有
43:29
a lot more members of Congress?
多得多的 Congress 议员呢？
43:31
When I think about my own feelings of alienation from our democracy, it just starts from the
当我想到自己对我们民主的疏离感时，这一切的起点就是
43:36
fact that my Congressperson doesn't care what I think, but if there were five times as
我的 Congress 议员根本不在乎我怎么想，但如果他们的人数有五倍
43:42
many of them, maybe they would live in my neighborhood and I would see them at the
之多，也许他们就会住在我家附近，我会在
43:46
store and we would get more of those face-to-face interactions.
商店里见到他们，我们也会有更多面对面的互动。
43:49
I'm curious what you think about that as a strategy for moving us back toward liberal democracy.
我很好奇，你如何看待那作为让我们回归自由民主的策略。
43:55
Yeah, I think that that's a no-brainer.
是啊，我觉得那简直想都不用想。
43:59
It's really crucial.
这真的很关键。
44:00
Yeah, there are so many good government reformers out there.
是啊，外面有很多优秀的政府改革者。
44:03
My colleague at Harvard, Daniel Island, the political philosopher has been calling for recalibration
我在哈佛的同事 Daniel Island，那位政治哲学家，一直在呼吁重新调整
44:08
of representation in Congress.
国会的代表性。
44:09
It's been overdue for really almost a century at this point.
这个问题到现在其实已经拖了将近一个世纪了。
44:13
Can you imagine how long those hearings would be, though?
不过你能想象那些听证会会有多长吗？
44:15
Oh, my God.
哦，我的天。
44:16
I know this is a partisan position, but the equal suffrage in the Senate has been a problem
我知道这是个党派立场，但参议院的平等投票权从一开始就有问题。
44:22
from the start.
詹姆斯·麦迪逊当初就反对这个。
44:23
James Madison was opposed to it.
这确实是个问题。
44:24
It's a problem.
这就是为什么现在会有人喊着要废除参议院。
44:25
That's why you get these people now saying abolish the Senate.
这在美国是个由来已久的政治立场，而你们这边的人长期以来也是一直这么想的。
44:29
That's a long-standing political position in America and it's your people who thought
那是美国长期以来的一个政治立场，而且你们的人长期以来也是这么认为的。
44:31
that for a long way.
我觉得这其实正是为什么data center的事情会火起来的原因，因为大家都跑到当地图书馆去参加社区会议，然后我们就像，天哪。
44:32
Yeah, that said, you've got to be willing to go to the store and talk to that person.
是啊，话虽这么说，你还是得愿意去店里跟那个人聊聊。
44:36
Mark Zuckerberg has kind of sent his humanoid robot to go get the baking goods he needs
Mark Zuckerberg 也算是派他的 humanoid robot 去给他孩子买烘焙材料了。
44:40
for his child.
绝对。
44:41
Absolutely.
你还是得走出家门，不过我完全同意。
44:42
You have to still get out of the house, but I entirely agree.
我觉得这其实就是 data center 相关的东西突然火起来的原因——因为人们都出现在当地图书馆参加社区会议，然后我们心想，天哪。
44:46
I think that's actually what's been so why the data center stuff is caught fire because
我觉得人们其实都知道，实际上我只是觉得Instagram对我的青少年孩子真的很不好，或者我只是，我知道我以前晚上上床后会读小说。
44:51
people are showing up at the local library for the community meeting and we're like,
所以，我也不知道。
44:56
Oh, my God.
people are showing up at the local library for the community meeting and we're like,
44:57
I haven't seen you in so long.
我好久没见到你了。
44:59
It's so hard because it can feel like all of this is happening at an individual level.
这很难，因为这一切可能会让人觉得都是在个人层面发生的。
45:05
It is the individual that chooses to download Instagram and create an account and spend
是个人选择下载 Instagram、创建账号，然后花
45:10
all day scrolling instead of going to the community meeting.
一整天刷屏，而不是去参加社区会议。
45:13
But it is also clear that in aggregate, it does have this atomizing effect and it's not
但同样明显的是，从总体来看，它确实有这种原子化的效应，而且我并不
45:19
clear to me that people are one day just going to wake up and say, well, the hell with
清楚人们有一天会醒来然后说，唉，去他的
45:23
this, right?
这一切，对吧？
45:24
Let's sort of return back to 19th century American democracy.
让我们算是回到19世纪的美国民主吧。
45:30
So I honestly don't know what to do about it because so much of the problem just looks
所以我真的不知道该怎么办，因为这个问题在很大程度上看起来就像成年人在做自由选择。
45:35
like adults making free choices.
嗯。
45:37
Yeah.
我认为实际上社会危害目前是显而易见的，而政治危害可能还不是这样，我的书讲的就是政治危害。
45:38
I think actually the social harms are legible currently in a way that the political harms
我觉得人们其实知道——实际上，我就是觉得Instagram对我的青少年孩子很不好，或者我就是知道，我过去晚上上床睡觉时还会读小说。
45:44
are maybe not and my book is about the political harms.
这不是在说我自己。
45:47
I think people know, actually, I just think Instagram is really bad for my teenager or
人们都跑到当地图书馆来参加社区会议，然后我们就觉得，
45:54
I just, I know that I used to read novels that night when I got into bed.
Oh, my God.
45:58
This is not autobiographical.
这不是自传。
45:59
I assist.
我协助。
46:00
But now I watch YouTube reels and I've lost something.
但现在我看YouTube Reels，感觉自己失去了些什么。
46:06
I think people can see the social harms and then it does feel like because so much of
我觉得人们能看到这些社会危害，然后确实会觉得，因为在很大程度上
46:10
our politics amounts to consumer choice, like, oh, well, you could just decide to do
我们的政治相当于消费者选择，比如，哦，好吧，你完全可以决定
46:14
it differently.
换个方式去做。
46:15
Like it's, some of these things are hard to make decisions about.
就像，有些事真的很难做决定。
46:19
So I kind of, I'm somewhat optimistic about some of the social harms because I think they're
所以我算是，我对一些社会危害有点乐观，因为我觉得它们是
46:24
remediable.
可以补救的。
46:25
Is that a word?
那是个词吗？
46:28
I think the political harms are less visible to us and that's kind of partly why I wrote
我觉得政治上的危害对我们来说没那么显眼，这也部分是我写这本书的原因。
46:31
the book.
我知道你是历史学家，Jill，这可能是我们最后一个问题，但我想知道，你有没有那种类似 Michael Pollan 式的、简短又启发性的建议，给那些想理解你在书里做出的诊断，并且试着以更符合自然、自身价值观和人性方式生活的人？
46:32
I know you are a historian, Jill, and this could be our last question, but I'm wondering
比如说，你知道的，少吃点，主要是植物？
46:36
if you have some sort of inspired brief Michael Pollan-esque advice for people who are trying
如果你有什么类似 Michael Pollan 式的、带启发性的简短建议，给那些想理解你在书里下的诊断，并且想活得更符合自然、自己的价值观和人性的人。
46:43
to wrap their head around some of the diagnosis that you've made in this book and sort of live
比如说，你知道，就是“少吃，多吃植物”那种？
46:51
in a way that is more consistent with nature and their own values and humanity.
五年后就有空间了。
46:57
Like, what is the, you know, eat less, mostly plants?
你看，你得从小事做起。
47:00
Yeah.
是啊。
47:01
Plan data, not too much.
计划 data，不要太多。
47:02
Yeah.
嗯。
47:03
What is the Jill LaPore version of that maximum?
Jill LaPore 版本的那个最大值是什么？
47:05
Oh, you know, I have not.
哦，你知道，我没有。
47:07
You don't want to historian with a pitch fork like I just think that's a bad plan.
你不想当一个拿着干草叉的历史学家，我就是觉得那是个糟糕的计划。
47:12
I think you kind of want to set yourself up for not having a humanoid robot in your living
我觉得你最好做好准备，五年后你客厅里不会有类人机器人。
47:16
room in five years time.
抱歉。
47:18
And one of those, one of those ways is to take like certain rooms of your house back one
其中一种方法是，把家里某些房间一间一间地夺回来，比如先夺回卧室，浴室，他妈的浴室，懂吧。
47:22
at a time, like take your bedroom back first, the bathroom, the fucking bathroom, okay.
把你自己从浴室里解救出来。
47:31
Rescue yourself from the bathroom.
让浴室成为一块圣地。
47:33
Let the bathroom be a sanctuary.
让手机不进浴室，那就是你在为拆解这个人工状态尽一份力。
47:35
Keeping your phone out of the bathroom, that is you will be doing your part to dismantle
听着，你得从小事做起。
47:40
the artificial state.
对不起。
47:42
Look, you got to start small.
天哪。
47:43
I'm sorry.
I think people know, actually, I just think Instagram is really bad for my teenager or
47:44
I told you.
我跟你说过了。
47:45
I told you.
我跟你说过了。
47:46
Don't ask me.
别问我。
47:47
All politics is local.
所有政治都是地方性的。
47:48
There's a lot of big things you could do.
你能做很多大事。
47:49
Vote for someone who supports having a data center moratorium until we can actually deliberate
投票给支持暂停建设 data center、直到我们能真正民主地审议这些至关重要的问题的人。
47:54
over these really crucial matters democratically.
好的。
47:57
All right.
好吧。
47:58
Well, Jill, that's a great place to leave it.
好吧，Jill，我们就在这儿收尾吧，挺好的。
48:00
The new book is The Rise and Fall of the Artificial State, Jill LaPore.
这本新书是 Jill LaPore 的《The Rise and Fall of the Artificial State》。
48:04
Thanks so much for joining.
非常感谢你能来。
48:05
Thank Jill.
谢谢 Jill。
48:06
Thanks, you guys.
谢谢你们。
48:09
When we come back, what do a bankrupt airline, a mysterious Amazon warehouse and the AI start
等我们回来，要聊的是：一家破产的航空公司、一个神秘的 Amazon 仓库，还有 AI 初创公司 Mechanized——它们有什么共同点？
48:14
up mechanized have in common?
根据律师的建议，我拒绝发表评论。
48:16
I'm refusing to comment on the advice of my lawyer.
在律师的建议下，我拒绝评论。
48:19
Find out in our new segment, Train of Thought.
在我们的新板块 Train of Thought 中寻找答案。
48:28
I'm Jonathan Knight and I'm the general manager of New York Times Games.
我是 Jonathan Knight，New York Times Games 的总经理。
48:32
If you play our games, you probably know there's something a bit different about them.
如果你玩我们的游戏，你大概知道它们有点与众不同。
48:36
Just like there are writers behind the articles you read in the Times, there are creators
就像你在 Times 上读到的文章背后有作家一样，也有创作者
48:40
behind our daily puzzles.
在我们每日谜题的背后。
48:42
Gracie Bennett curates the day's wordless solution to keep it lively and varied.
Gracie Bennett 策划当天的无文字解答，让它保持生动多变。
48:47
When a Lou creates each connections board, including all those categories that try to stump
当 Lou 创建每个 Connections 棋盘时，包括所有那些试图难倒你的分类。
48:51
you.
你。
48:52
Samazersky comes through every last letter, word, and pangram and spelling bee so that loyal
Samazersky 无论是字母、单词、pangram 还是 Spelling Bee，都一一做到位，让各个水平的忠实玩家都能乐在其中。
48:57
players of all skill levels enjoy it.
我们的谜题每天都是人工制作的，品质标准一如既往，正是你对 New York Times 的期待。
49:00
Our puzzles are human-made every day, with the standards you'd expect from the New York
这一点很重要，因为当你选择把时间花在我们的游戏上，这段时间就应该花得值得——去解那些有挑战性、有惊喜、又令人开心的谜题。
49:04
Times.
为你手工打造的谜题。
49:05
And this matters because when you choose to spend time with our games, it should be time
我们觉得这值得投入，也值得付费。
49:09
well spent, solving puzzles that are challenging, surprising, and joyful.
花得很值，去解那些有挑战性、出人意料又让人开心的谜题。
49:14
Puzzles handcrafted for you.
为你手工打造的谜题。
49:16
We think that's something worth investing in and something worth paying for.
我们认为这值得投入，也值得付费。
49:20
If you think so too, download New York Times Games from your favorite app store or go
如果你也这么想，就从你常用的 app store 下载 New York Times Games，或者前往 nytimes.com/play games。
49:25
to nytimes.com slash play games.
Casey，你看到那条关于 Google 购买 Spirit Airlines 数据的新闻了吗？
49:28
Casey, did you see this story about Google buying the data of spirit airlines?
我当然看到了。
49:34
I sure did.
这是最近几周我看到的最引人入胜的报道之一，它让我深入思考 training data，以及我们正在经历的 data collection 新时代。
49:36
This was one of the most fascinating stories I've seen in recent weeks, and it led me
所以我想，我们就借这个 Spirit 的故事，来互相...
49:40
down this incredible rabbit hole of thinking about training data and the new era of data
掉进了这个关于 training data 和我们所处的 data collection 新时代的奇妙兔子洞。
49:46
collection we are in.
我觉得大家其实都知道，我只是觉得Instagram对我家 teenager 真的很不好，或者
49:48
So I thought we should use this spirit story as an occasion to just sort of catch each
所以我觉得我们应该借着这个 spirit 故事来顺便互相了解一下
49:53
other up on the state of AI training data in general because it is fascinating and I think
来聊聊AI training data的整体现状吧，因为这很吸引人，而且我觉得
49:57
underappreciated.
被低估了。
49:58
And it sounds like the perfect frame Kevin for our new segment, Train of Thought.
而且Kevin，这听起来像是我们新栏目Train of Thought的完美引子。
50:03
I love that we're just starting new segments every week until the show ends.
我喜欢我们每周都在开新栏目，直到节目结束。
50:07
We are.
我们确实是。
50:08
It's time for the first and last installment of our new segment, Train of Thought.
是时候迎来我们新栏目Train of Thought的第一期，也是最后一期了。
50:12
This is kind of the caboose as it were.
这有点像火车的车尾，可以说是。
50:14
Yes.
是的。
50:15
Interestingly, we have two train-related segments on the show.
有趣的是，我们这期节目里有两个跟 training 相关的环节。
50:27
So the reason that we wanted to do this segment today is because there was a very strange
所以我们今天想做这个环节，是因为有一个非常奇怪的
50:31
story that peaked our attention over the past week involving the defunct airline spirit
故事，在过去一周引起了我们的注意，跟已经倒闭的航空公司有关
50:39
airlines.
Spirit Airlines。
50:40
Hands down the worst airline of all time.
绝对是有史以来最糟糕的航空公司。
50:42
I don't even know who else is in the conversation.
我都不知道还有谁够格进入这个讨论。
50:44
Yes.
对。
50:45
I did fly it one time.
我还真坐过一次。
50:46
This week a bankruptcy court auctioned off spirit airlines internal corporate data, Google
这周，一家破产法庭拍卖了Spirit Airlines的内部公司数据，Google
50:52
won the bid offering to pay $10 million for this data, beating out a $7.5 million offer
以1000万美元的出价赢得了竞标，击败了750万美元
50:58
from the AI data company, Mercor.
来自AI数据公司Mercor的报价。
51:01
$10 million, Kevin, what did Google get for that price?
Kevin，1000万美元，Google用这个价格得到了什么？
51:06
So this deal apparently included 100 million emails, 500 million Microsoft Teams chats, and
这笔交易显然包括1亿封邮件、5亿条Microsoft Teams聊天记录，以及
51:12
other conversations.
其他对话。
51:14
$7.5 billion passenger transaction records dating back to 2008 and 30 million lines of
可追溯到2008年的75亿条乘客交易记录，以及3000万行
51:20
spirit's internal source code and other documentation.
Spirit的内部source code和其他文档。
51:23
Well, I would consider spirit's internal source code malware, but everything else sounds
嗯，我会认为 Spirit 的内部 source code 是 malware，不过其他一切都听起来很有意思。
51:27
interesting.
那么，大家觉得 Google 会拿这些 data 做什么呢？
51:28
So what do we think Google is going to do with this data?
我看到这个之后，脑子里想的是：Google 是世界信息的守护者，按理说它拥有的用于 training AI models 的 data 比世界上任何公司都要多得多。
51:31
This was where my head went after I saw this because I thought why is Google guardian of
他们怎么会花 1000 万美元去买这个破产航空公司的 data？
51:35
the world's information, presumably the possessors of vastly more data for training AI models than
而这让我一头扎进了一个非常迷人的兔子洞——training data 的世界。
51:40
any company in the world.
世界上的任何一家公司。
51:42
What are they doing paying $10 million for this bankrupt airlines data?
他们到底在干什么，花一千万美元买这家破产航空公司的数据？
51:47
And that sent me down like a really fascinating rabbit hole of this world of training data
而这让我进入了一个特别迷人的 rabbit hole，就是 training data 这个领域
51:52
and training environments that all of the AI companies now are investing really heavily
以及所有 AI 公司现在都在重金投入的 training environments。
51:58
into.
那跟我们说说你学到了什么。
51:59
Well, tell us what you've learned.
那么 Casey，你知道吗，有一家公司会拍卖已倒闭公司的 Slack histories 和 email histories？
52:00
So Casey, did you know that there is a company that auctions off the slack histories and
我挺惊讶的，这居然也有市场。
52:04
email histories of defunct companies?
对。
52:07
I'm surprised to learn that there is a market for that.
所以有一家叫 simple closure 的公司，它的整个业务过去有点像是帮那些失败的……
52:09
Yes.
对。
52:10
So there is a company simple closure whose whole business used to be sort of helping failed
所以有一家公司叫 simple closure，它过去的整个业务就是帮助那些失败的
52:13
startups wind down.
创业公司逐渐停摆。
52:15
But now they have these guys are like the undertakers of Silicon Valley.
但现在他们这些人就像是硅谷的殡葬师。
52:19
Yes.
对。
52:20
Exactly.
没错。
52:21
Their corporate logo is just like the grim reaper.
他们公司的标志简直就像死神。
52:23
Yes.
对。
52:24
Yes.
对。
52:25
If these guys show up at your office or you get a call from them, it's a very bad day for
如果这些人出现在你的办公室，或者你接到他们的电话，那对你来说是非常糟糕的一天。
52:28
your company.
你的公司。
52:29
So basically, they were sort of like, you know, helping do the orderly wind downs of
所以基本上，他们就是，你知道，帮忙做这些东西的有序清盘。
52:33
these things.
但后来他们发现，哦，这些死掉的 startup 的 data 其实有市场。
52:34
But then they realize, oh, there's actually a market for the data from these dead startups.
于是他们开始把 data 卖给 AI 公司。
52:39
And so they start selling it to AI companies.
截至今年四月，他们差不多做了近一百笔交易，每笔大概在一万到十万美元每家公司。
52:41
And as of April of this year, they had done almost a hundred deals ranging from roughly
还有，我就是想知道，当买家真的拿到这些...
52:47
$10,000 to $100,000 per company.
每家公司 1 万到 10 万美元。
52:50
And again, I just want to know like what happens when the buyer actually gets a hold of the
然后我还是想知道，当买家真的拿到那些数据之后会发生什么。
52:55
data?
data 呢？
52:56
Like, where does it go?
比如说，它会去哪里？
52:57
And does it violate my HIPAA race?
那它会不会违反我的 HIPAA 权利？
52:59
It does not violate your HIPAA race.
它不会违反你的 HIPAA 权利。
53:01
These are presumably not things that are covered by HIPAA.
这些大概都不在 HIPAA 的覆盖范围内。
53:04
But this is basically this new strategy.
但这基本上就是这种新策略。
53:07
So there was an era where all of the data collection and scraping that the AI companies
所以曾经有一个时代，AI 公司做的所有 data collection 和 scraping
53:13
did was sort of focused on getting the highest quality text images and video they could.
都算是集中在尽可能拿到最高质量的 text、images 和 video 上。
53:20
This was used for pre-training.
这用于pre-training。
53:21
For the first step in the model process, you thrown as much data as a can.
在模型流程的第一步，你把能塞进去的数据全都塞进去。
53:24
The model learns from it.
模型从中学习。
53:26
This is sort of how you saw the models improved for many years.
这就是多年来你看到模型进步的方式。
53:29
And this is where all these stories came from about scraping Reddit, feeding it into
也是那些关于抓取Reddit数据、喂给模型的故事的由来。
53:34
the models.
甚至还有我们的故事，Kevin？是的。
53:35
Or even our stories, Kevin?
或者甚至是我们的故事，Kevin？
53:36
Yes.
是的。
53:37
And we're sort of the first era of AI training.
我们算是处在 AI training 的第一个时代。
53:42
Now we are in this different era, which I would call the era of experience, which is basically
现在我们处于一个不同的时代，我会称之为体验的时代，它基本上就是
53:47
where these models now, the way that they are improved is through reinforcement learning.
这些模型现在的改进方式是通过 reinforcement learning。
53:52
Reinforcement learning is sort of this trial and error process where you go out, you sort
Reinforcement learning 是一种试错的过程，你出去，你有点
53:57
of do a little task or a test or play a game and you get a score or you get some indication
去做一个小任务、做个测试或玩个游戏，然后你会得到一个分数，或者得到一些指示
54:04
of whether you've succeeded or not.
关于你是否成功了。
54:06
And maybe you've broken into a hucking face.
而且也许你已经进入了 hacking 阶段。
54:08
Success.
成功。
54:09
Yes.
对。
54:10
That one was a success.
那一次成功了。
54:11
And then you sort of try it over and over again using slightly different strategies or
然后你会一遍遍地尝试，每次用稍微不同的策略，或者
54:15
slightly different techniques every time and you sort of get signals about what it works
稍微不同的技术，然后你会得到一些信号，知道什么有效
54:18
and what doesn't.
什么没用。
54:19
And that's how you improve at things like autonomous coding.
这就是你提升像autonomous coding这类能力的方法。
54:22
So what's happening with these data sets, including the data set of the dearly departed
那么这些data sets，包括那位已故的
54:27
spirit airlines, presumably, is that they are being turned into reinforcement learning
Spirit Airlines的data set，大概正在被用于reinforcement learning。
54:31
environments for AI agents to learn new tasks.
用于 AI agents 学习新任务的环境。
54:36
So basically, you use this data to rebuild whatever company you've acquired their data.
所以基本上，你用这些数据来重建任何你拿到数据的公司。
54:42
So you're basically rebuilding an airline or an insurance company or a startup as an
所以你基本上是在把一家航空公司、一家保险公司或者一家初创公司重建为 AI agents 的环境，本质上就像一座训练健身房，让这些 agents 出去尝试不同的任务，看看它们是成功还是失败。
54:47
environment for AI agents, as a training gym essentially for these agents to go out and
你这是在创造一个噩梦般的平行宇宙，在那里 spirit airlines 依然存在，而且还在接受航班预订。
54:53
try different tasks and see whether they succeed or fail.
对。
54:56
You're creating a nightmare parallel universe where spirit airlines still exists and it's
你在创造一个噩梦般的平行宇宙，spirit airlines 仍然存在，而且它还在
55:02
booking flights.
订机票。
55:04
Yes.
是的。
55:04
So you can kind of rebuild the company as a video game.
所以你可以把公司重新打造成一个视频游戏。
55:08
You can sort of mine tasks from these emails and team's messages.
你可以算是从这些邮件和团队消息里挖掘任务。
55:12
And then you can actually see how things played out.
然后你就能真正看到事情是怎么发展的。
55:15
So for example, if you have the transaction history of an airline, you can say what happened
所以举个例子，如果你有一家航空公司的交易历史，你就可以说发生了什么，
55:22
in 2015, when there was a big storm on the East Coast, how did the routing decisions
2015年，当东海岸有一场大风暴时，航线决策
55:29
get made and did that result in people getting to their destinations on time?
是怎么做出的，以及这是否让人们按时到达了目的地？
55:34
And if you can get spirit airlines to turn a profit in the sandbox, that's AGI.
那么如果你能让 Spirit Airlines 在 sandbox 里扭亏为盈，那就是 AGI。
55:40
Yes.
对。
55:41
So it's basically you have these kind of data points that come from these companies about
所以基本上就是，这些公司提供的 data points 讲的是人们怎么跟系统互动、系统之间怎么互相互动，还有客户怎么在这些大型系统里穿行。
55:47
how people interact with systems, about how systems interact with each other and about
哎呀，听你说这些我总算放心了，Kevin，因为看到那篇报道的时候，我就想，天哪，Google 要开航空公司了，而且你也知道他们，肯定不会只有一家航空公司，还会有一个 app，然后就像，好吧，你得选择，你坐的是 Google Airlines、Google Airways 还是 Google Air，它们全都一样，但又完全不一样，而且它们可能还会……
55:51
how customers navigate through these sort of giant systems.
客户是如何在这些庞大的系统中穿行的。
55:55
Well, I am so relieved to hear you say all of this, Kevin, because when I saw this story,
好吧，Kevin，听你这么说，我真是松了一口气，因为当我看到这个故事时，
56:00
I thought, oh, my God, Google is going to start an airline and you know with them,
我想，天哪，Google 要开航空公司了，而且你知道，和他们……
56:06
it wouldn't just be one airline, there would be an app, and it was like, okay, you have
不会只有一家航空公司，会有一个 App，然后就像，好吧，你得
56:09
to choose, are you flying Google Airlines, Google Airways, or Google Air, and they would
来选择，你是坐 Google Airlines、Google Airways，还是 Google Air，然后它们会
56:13
all be the same, but they would all be completely different, and also they would probably
全都一样，但又全都完全不同，而且它们可能还会
56:16
be different apps.
会是不同的apps。
56:17
Well, and also since they were trained on spirit airlines, they would also charge you for
嗯，而且既然他们是用Spirit Airlines的数据训练的，他们也会向你收取
56:20
peanuts, who charge you for, you know, slightly bigger seat, you know, charge you to use
花生米的费用，还会收你，你知道，稍微大一点的座位，你知道，收你使用
56:26
the bathroom, maybe, everything would be charges.
厕所的钱，也许，什么都会收费。
56:29
Yeah.
对。
56:30
So not, not eager to see that, that business.
所以不，不期待看到那种，那种生意。
56:31
So what's interesting about the dead company sort of data market is that you are, you
那么，关于这些死掉的公司的数据市场，有意思的是，你，你
56:37
are sort of able to turn these companies into sort of living, zombified, simulacra of
某种程度上能把它们变成某种活生生的、僵尸化的拟像
56:43
the original company, but you're also training these systems on companies that ultimately did
原来的公司，但你也用那些最终没有成功的公司的数据来训练这些系统。
56:49
not succeed.
所以我很好奇，这些 datasets 到底是不是真的在帮助这些 models 在这些 tasks 上提升，还是说它们以某种微妙的方式被不成功公司的数据 conditioned，从而在它们试图学习的 tasks 上变得更糟。
56:50
So I'm very curious to know like if these data sets actually are helping these models improve
你担心这些未来的 models 会有一种失败者心态。
56:56
at these tasks, or if there's some subtle way in which they are being conditioned on the
是啊。
57:01
data of unsuccessful companies, and thereby are becoming worse at the tasks that they're
用那些不成功公司的数据，从而在它们试图
57:05
trying to learn.
学习的任务上变得更差。
57:06
You're worried that these future models are going to have a loser mentality.
你担心这些未来的模型会有一种失败者心态。
57:09
Yeah.
对。
57:10
They don't have what it takes to cut it in the modern economy.
他们没有本事在现代经济里立足。
57:13
There's other interesting data story this week that came from 404 media, which was that
这周还有一个来自 404 media 的有趣数据故事，内容是
57:17
they slipped an air tag into a rare book that was part of a bulk book order on a marketplace
他们把一枚 AirTag 塞进了一本珍本书里，这本书是一批图书订单的一部分，订单是
57:24
site called Biblio.
在名为 Biblio 的市场网站上下的。
57:26
They finally determined that this book lands at an Amazon warehouse in Las Vegas, specifically
他们最终确定这本书落在了 Las Vegas 的一个 Amazon 仓库，具体是
57:32
an internal unit called VGT3, which has, according to 404 media, a door marked with the logo
一个叫 VGT3 的内部单位，据 404 media 说，这个单位有一扇门，门上标着一个
57:40
of a dinosaur eating a book.
恐龙吃书的标志。
57:42
That feels a little on the nose, even for this simulation, I have to say.
我不得不说，就算是在这个模拟里，这也太明显了。
57:46
So Casey, why are these books ending up at mysterious Amazon warehouses?
那么 Casey，为什么这些书最后会出现在亚马逊的神秘仓库里？
57:51
What are they doing with them?
他们拿这些书做什么？
57:52
It is a good question, and this ties into some of the lawsuits that have been filed
这是个好问题，这跟一些已经提起的诉讼有关，
57:58
against the big AI labs.
针对大型 AI 实验室的诉讼。
58:01
In particular, this big case against anthropic that you may remember, and the judge in
特别是你可能还记得的那个针对 Anthropic 的大案子，案中的法官
58:07
that case ruled that because anthropic had bought these millions of print books, scanned
裁定，因为 Anthropic 购买了这些数百万本纸质书，扫描了
58:14
them, and then discarded the paper, that this was fair use of the material because each
它们，然后丢弃了纸张，这属于 fair use，因为每份
58:21
digital copy had replaced a legally purchased print original.
数字副本都替代了一本合法购买的纸质原版。
58:26
So there was no multiplication of the number of copies.
所以副本数量并没有成倍增加。
58:30
It was just kind of a one-to-one shift in format, and so what I think the other labs have
这其实只是格式上的一对一转换，所以我觉得其他实验室从Kevin这件事中
58:35
taken away from this Kevin is you're not going to run into as many legal issues if you destroy
学到的是：如果你销毁
58:42
these books.
这些书，就不会遇到那么多法律问题。
58:43
Right.
对。
58:44
So it's not like the AI companies are kind of giddy about like destroying these relics
所以并不是说AI公司会为销毁这些文明的遗物而
58:47
of civilization.
感到兴奋。
58:48
Well, they might be.
嗯，它们可能确实会。
58:49
They might be.
他们可能是。
58:50
I'm sorry, all of Kevin, who is this book?
对不起，所有的Kevin，这本书是谁？
58:51
I'd be having a good day at the office, but it is the sort of fallout of this legal environment
我本来在办公室过得挺好的，但这就是他们所处的法律环境的后果，在这种环境里，法律上来讲，他们扫描完书之后把书销毁更安全。
58:56
that they're in, where it's just safer for them legally to destroy the books after they
我只能说，这整件事在我看来太蠢了，真的是一个我们尊重法律条文却不尊重法律精神的情况，对吧？
59:00
have finished scanning them.
就好像，是的，你确实对数据做了转换，但显然，你知道，真正的抱怨是……
59:01
And I just had to say, like, this whole thing seems so stupid to me, truly a case where
然后我不得不说，这整件事在我看来太蠢了，真是一个...的例子
59:05
we are honoring the letter of the law, but not the spirit, right?
我们算是遵守了法律条文，但没遵守法律精神，对吧？
59:08
It's like, yes, you literally transformed the data, but obviously, you know, the real complaint
就是说，对，你确实对数据做了转换，但显然，你懂的，真正的抱怨
59:13
here that the authors have, at least, you know, the ones who have sued is I didn't want
听到作者们至少，你知道，那些起诉的人，是我不想
59:17
to use my, you know, book this way.
用我的，你知道，书这种方式。
59:20
So I expect we're going to see a lot more angst over this as we continue to see more
所以我预计我们会看到更多的焦虑，随着我们继续看到更多
59:25
books destroyed.
书籍被毁掉。
59:26
Yeah.
是啊。
59:27
You know, back in my day, Kevin, we would only see books destroyed because the Republicans
你知道，在我那个年代，Kevin，我们只会看到书被毁掉，因为共和党人
59:31
had read a case vaccine, and I want to get back to that point.
读了一个案例疫苗，而且我想回到那一点。
59:35
All right, Casey, one more data story to talk about this week, which is sort of related
好的，Casey，这周还有一个 data 故事要谈，这有点相关。
59:40
to the first one that we discussed, both because it involves Google and because it involves
回到我们讨论的第一个，既因为涉及 Google，也因为涉及
59:44
these sort of high quality training environments for reinforcement learning that all these AI
这类高质量 reinforcement learning 训练环境，所有 AI 公司
59:50
companies are now racing to build.
现在都在竞相构建。
59:52
This one is about the 50 person startup mechanize.
这家是 50 人规模的初创公司 mechanize。
59:57
We have talked about mechanize on the show before we interviewed two of their co-founders.
我们之前在节目里聊过 mechanize，还采访过他们的两位联合创始人。
60:01
They are about a year old, and they specialize in companies, they're a little older than
他们大约成立一年了，并且他们专精于公司，他们其实比那
60:05
that.
要老一些。
60:06
Yes.
对。
60:07
Yes.
对。
60:08
They specialize in creating these RL environments for coding and other tasks.
他们专门为 coding 和其他任务创建这些 RL environments。
60:14
And they are reportedly in talks to be acquired by Google for over $1.5 billion.
据报道，他们正在和 Google 谈收购，价格超过 15 亿美元。
60:19
Not bad for a year's work.
干了一年，真不错。
60:22
Yes.
对。
60:23
So this is a big boom area inside the AI boom, basically.
所以这基本上就是 AI 热潮里面的一个大热领域。
60:27
If you want these sort of high quality tasks that you can put your AI agents into and
如果你想要这些高质量的任务，能把你的 AI agents 放进去，
60:33
have them sort of hill climb on them, get a little bit better every time.
让它们在这些任务上 hill climb，每次都能好一点。
60:37
They need to be good tasks, right?
它们得是好的任务，对吧？
60:39
They need to be thoughtfully created.
它们需要被精心设计。
60:41
They need to not have a bunch of sort of obvious flaws in them.
不能有一堆显而易见的毛病。
60:44
And they need to mirror the things that real people might be doing in their jobs.
而且要能反映真实工作中人们可能会做的事情。
60:49
So one way that you might create an RL environment is like create a fake version of Amazon.com, right?
所以创建RL环境的一个方法，就是做一个Amazon.com的仿冒版，对吧？
60:56
And everything about it is exactly identical to Amazon.com, except it's not called Amazon.com.
除了不叫Amazon.com之外，其他一切都完全一样。
61:01
And it's just for these AI agents to learn how to click around, put things in the cart, browse
就是专门让这些AI agents学习怎么点击浏览、往购物车里放东西、逛网站，
61:07
the site, destroy books, destroy books.
销毁书籍，销毁书籍。
61:11
And a warehouse in Las Vegas.
还有拉斯维加斯的一个仓库。
61:13
So this kind of sort of simulated environment is the kind of thing that mechanizes specializes
所以这种模拟环境正是 mechanizes 擅长构建的东西，大概也是 Google 有兴趣收购他们的原因。
61:21
in building and presumably why Google is interested in acquiring them.
所以，这个故事和我们节目开头聊的 hugging face 故事之间有一个挺有意思的联系——我觉得很多 RL environments 在设计、构建或安全性上都不太行，对吧？
61:25
So one interesting sort of connective tissue between this story and the hugging face story
部分问题，或者说可能正是为什么像 Google 这样的公司开始把这种 expertise 收归内部，而不是外包给 vendor 或 startup。
61:33
that we discussed at the top of the show is that I think a lot of these RL environments
我们节目开头讨论的是，我觉得很多这些RL环境
61:40
are not particularly well-designed or built or secured, right?
并不是特别设计良好、构建良好或者安全，对吧？
61:44
Part of the problem, and maybe the reason that companies like Google are starting to bring
部分问题在于，也许也是像Google这样的公司开始把
61:49
this expertise in-house rather than sort of contracted out to a vendor or a startup
这种专业能力引入内部，而不是外包给供应商或初创公司
61:55
is that they are finding issues with the way that these tests and these RL environments
就是他们发现这些测试和RL环境的构建方式有问题。
62:00
are constructed.
所以现在已经有几例有缺陷的security test了，都来自同一个供应商 Irregular，Meta和Anthropic都依赖过它。
62:01
So there's been a couple instances now of a flawed security test by this, by the same vendor
你可能一两个星期前看到过这件事。
62:09
irregular that both meta and anthropic relied on.
我的猜测，再加上我和实验室里一些人聊过，基本上就是他们已经到了这个地步：这些测试必须非常好、非常安全，所以不得不自己内部构建，用极高质量的数据。
62:12
You may have seen this story a week or two ago.
你可能一两个星期前看到过这个故事。
62:15
My guess, and the conversations that I've had with some of the people at the labs is basically
我的猜测，以及我和一些实验室的人聊过之后，基本上就是
62:19
they have gotten to the point where these tests need to be so good and so secure that they
他们已经到了这种程度，这些测试必须做得足够好、足够安全，以至于他们不得不用极其高质量的数据在内部自行构建。
62:25
have to be building them in-house using extremely high-quality data.
是啊。
62:29
Yeah, irregular put out a report about some of the incidents that you just mentioned.
对，irregular 发了一份报告，关于你刚才提到的那些事件。
62:35
And it got criticism from the security community saying you're not offering us enough detail
然后它遭到了安全社区的批评，说你们没给我们足够的细节来理解到底出了什么问题。
62:40
to understand what went wrong.
所以我怀疑，如果 irregular 不更加开诚布公，那些 labs 就会觉得他们别无选择，只能把这一切都收归内部。
62:42
And so I suspect if irregular is not more forthcoming than the labs are going to feel
对。
62:46
like they have no choice but to bring this all in-house.
这真是让我匪夷所思，现在这些 AI 公司基本上就是在构建 the Sims，只不过是在能想象的最宏大的行星尺度上，他们正在组装 data。
62:48
Yeah.
这对我来说简直难以置信，这些AI公司现在基本上就是在打造 the Sims，只不过是在能想象的最宏大的行星规模上整合数据。
62:49
It is just like mind-boggling to me that these AI companies now are sort of essentially
如果你有兴趣获取 hard fork 播客的数据存储，hard fork
62:55
building the Sims, but on the grandest planetary scale imaginable, they are assembling data
当你第一次下载 New York Times 应用时，你在应用里的第一个月
63:04
from the corpses of failed startups.
从那些失败初创公司的残骸中。
63:07
They are turning them into simulations and video games and then they are running their
他们把它们变成 simulations 和电子游戏，然后让他们的
63:11
AI agents through them to try to make them superhuman at everything.
AI agents 通过它们运行，试图让这些 AI agents 在一切事情上变得超人类。
63:15
And it is just like if you made that the plot of a science fiction novel 10 years ago,
这就好比如果你在 10 年前把这个当作科幻小说的情节，
63:20
people would have like criticized it for being a little over the top.
人们大概会批评它有点太过头了。
63:22
It is pretty wild.
这相当疯狂。
63:23
Now, let me ask you this, Kevin, I'm sure you've already thought about this, but hard fork
现在，让我问你这个，Kevin，我相信你已经想过这个了，但 hard fork
63:27
is preparing to wind down.
正准备收尾。
63:30
Have you thought about how much money we'd be in bold to get for the data?
你有没有想过咱们这些数据能卖出多少钱？
63:35
I would be open to seeing bids, now obviously, you know, it's not the best training data.
我愿意看看报价，不过显然，你懂的，这并不是最好的 training data。
63:42
There would be a lot of bad jokes, a lot of questionable interviews, but if spirit
肯定会有很多烂笑话，很多有问题的采访，但如果 spirit airlines 能拿到一千万美元，没错，有人就能请我们吃顿午饭。
63:50
airlines can get $10 million, exactly, somebody could buy us lunch.
我们没破产。
63:55
We didn't go bankrupt.
看看我们。
63:56
Look at us.
他们说过这行不通。
63:57
They said it would never work.
如果你有兴趣收购 hard fork 播客的 data stores，hard fork。
63:59
If you're interested in acquiring the data stores of the hard fork podcast, hard fork
把你破产航空公司的训练数据发给我们。
64:02
nytimes.com.
nytimes.com。
64:38
The New York Times app unlocked?
The New York Times app 解锁了？
64:41
Everyone knows the times is behind a paywall.
谁都知道 The Times 有 paywall。
64:44
Only subscribers have access to all the reporting.
只有订阅用户才能访问全部报道。
64:48
But what if you could explore the times for a month, for free, without putting in a credit
但如果你能免费探索 The Times 一个月，还不用填信用卡呢？
64:52
card?
现在你可以了。
64:53
Now you can.
当你第一次下载 The New York Times app，你在 app 里的第一个月
64:55
When you download the New York Times app for the first time, your first month in the app
当你第一次下载 New York Times app 时，你在 app 里的第一个月
64:59
is free.
是免费的。
65:00
A month to go behind the paywall.
有一个月的时间可以访问付费墙后的内容。
65:02
To see what time subscribers get every single day.
看看Times订阅用户每天都能获得什么。
65:05
All the investigations, the reviews, the recipes, the deeply reported fact-based journalism.
所有的调查、评论、食谱，以及深度报道的基于事实的新闻。
65:11
If you don't already subscribe to the New York Times, download the Times app today and
如果你还没有订阅New York Times，今天就下载Times应用，
65:15
get free access for 30 days.
就能获得30天的免费访问权限。
65:22
Hard fork is produced by Whitney Jones and Rachel Cohn.
Hard fork由Whitney Jones和Rachel Cohn制作。
65:25
We're edited by Veer and Povitch, we're fact-checked by Kate and Love.
我们由Veer和Povitch编辑，由Kate和Love进行事实核查。
65:29
Today's show is engineered by Katie McMurrin, original music by Marion Luzano, Rowan
今天的节目由 Katie McMurrin 负责工程制作，原创音乐来自 Marion Luzano、Rowan Nemisto、Alyssa Moxley 和 Dan Powell。
65:34
Nemisto, Alyssa Moxley, and Dan Powell.
视频制作由 Sawyer Roké、Jake Nichol 和 Chris Schott 负责。
65:37
Video production by Sawyer Roké, Jake Nichol, and Chris Schott.
你可以在 YouTube 上观看完整剧集，访问 youtube.com/hard fork。
65:41
You can watch this full episode on YouTube at youtube.com slash hard fork.
特别感谢 Paula Schumann、Clewing Tam、Brooke Mentors 和 Dahlia Hadad。
65:45
Special thanks to Paula Schumann, Clewing Tam, Brooke Mentors, and Dahlia Hadad.
你可以像往常一样给我们发邮件，地址是 hard fork@nytimes.com。
65:49
You can email us as always at hard fork at nytimes.com.
把你的破产航空公司训练数据发给我们。
65:53
Send us your bankrupt airline training data.
把你们破产航空公司的 training data 发给我们。