H-1B签证改革：打击滥用与市场化手段 [00:02:25]

The topic of the week: H-1B visas are being overhauled.

本周话题：H-1B签证 (H-1B Visa: 允许美国雇主临时雇佣外国高技能人才的非移民工作签证) 正在进行全面改革。

The Trump administration announced a new $100,000 fee for all future H-1B applications.

特朗普政府宣布，未来所有H-1B申请将征收一次性10万美元的费用。

This is a huge jump; the current fee is nothing—it's like $2,000 to $5,000 that you pay to the government.

这是一次巨大的跃升 [00:02:36]，目前的政府费用几乎微不足道，大约在2000到5000美元之间。

This hits on a lot of the Trump campaign promises: tougher on immigration, looking out for US workers.

这契合了特朗普竞选时的多项承诺 [00:03:00]：对移民采取更强硬立场，并关注美国本土工人。

We've talked before here about the abuse in the H-1B system.

我们此前在这里讨论过H-1B系统中的滥用问题。

I think it's a good idea to have this $100,000 fee.

我认为收取这10万美元的费用是个好主意 [00:03:40]。

Right now, there are something like five times as many H-1B applications as there are slots.

目前，H-1B的申请数量大约是名额数量的五倍 [00:03:47]。

They grant about 85,000 H-1Bs a year, and many more apply for it, which results in a lottery.

他们每年发放约8.5万份H-1B签证 [00:03:55]，但申请人数更多，因此采用了随机抽签的方式。

Over the past decade, roughly half the H-1Bs go to these IT consulting firms, and the average salary is like $65,000 a year.

纵观过去十年，大约有一半的H-1B签证流向了这些IT咨询公司 [00:04:09]，而平均年薪仅为6.5万美元左右。

This kind of puts the lie to this idea that H-1Bs are for high-skilled engineers and AI researchers.

这揭穿了H-1B签证是为高技能工程师和AI研究人员准备的说法 [00:04:15]。

A huge chunk of them end up going to low-end IT jobs.

事实上，很大一部分名额最终被用于低端IT工作 [00:04:34]。

By putting this $100,000 fee on it, you encourage the applications to go to the actual higher-skilled, higher-paid jobs where there's actually a shortage of Americans.

通过征收这10万美元的费用 [00:04:44]，可以鼓励申请流向那些真正缺乏美国人才的高技能、高薪职位。

It also encourages US companies to try to fill those jobs with Americans first.

这也激励了美国公司优先尝试用美国人来填补这些职位 [00:04:52]。

When I was in IT in the early 90s, the abuse was happening all the time, and it was 契约奴役 (Indentured Servitude: 一种劳工制度，工人通过签订契约在特定时间内为雇主工作以偿还债务或费用)—it was disgraceful.

早在90年代初我在IT行业时 [00:05:27]，这种滥用就一直存在，而且是可耻的契约奴役 (Indentured Servitude: 一种劳工制度，工人通过签订契约在特定时间内为雇主工作以偿还债务或费用)。

The IT people would hire typically Indians and say that these guys are going to work for half as much and twice as long, and they can't say no.

那些IT公司通常会雇佣印度人 [00:05:34]，并声称这些人会以一半的薪水工作两倍的时间，而且他们不能拒绝。

They can't say no when we put them on weekend coverage or if we want to do a buildout and they have to work 10 days in a row because we can kick them out of the country, and they have 30 days to find a new job.

当我们安排他们周末值班 [00:05:46]，或者需要连续工作10天进行部署时，他们不能拒绝，因为我们可以把他们踢出这个国家，而他们只有30天时间找新工作。

It's a giant scam on the bottom half of these.

对于低端的H-1B申请者来说 [00:05:57]，这是一个巨大的骗局。

I additionally think they should do an auction for one-third of these.

我另外认为 [00:06:25]，应该对三分之一的H-1B名额进行拍卖。

Let all these big tech companies that are truly trying to get in very unique PhDs from Oxford in AI man, let them just put out how many they want to buy and at what price, do a reverse auction, and fill one-third of them.

应该让那些真正想引进牛津大学AI领域独一无二博士人才的大型科技公司 [00:06:28]，公布他们想购买的数量和愿意支付的价格，进行逆向拍卖，填补三分之一的名额。

I came to the United States initially on a TN visa, then switched to an H-1B, and then got my green card and citizenship.

我最初是通过TN签证来到美国 [00:06:56]，随后转为H-1B，并在2010年代初获得了绿卡和公民身份。

Elon, Sundar Pichai, and Satya Nadella all came in on an H-1B.

埃隆·马斯克、桑达尔·皮查伊和萨蒂亚·纳德拉都是通过H-1B进入美国的 [00:07:13]。

That being said, I think Sacks is right that people have found an end run and have been abusing this H-1B system.

尽管如此 [00:07:18]，我认为萨克斯说得对，人们找到了规避的途径，并一直在滥用H-1B系统。

The H-1B program is supposed to be 85,000 visas a year, but the data shows it's been upwards of 10 times that number in many years, including the last several.

H-1B项目本应每年发放8.5万份签证 [00:07:49]，但数据显示，包括过去几年在内的许多年份里，实际数量高达这个数字的10倍。

When you see this, you can start to see why a lot of people are saying that there is wage suppression and that it's taking away from American jobs.

看到这一点 [00:08:16]，就可以理解为什么很多人说存在工资压制，以及它抢走了美国人的工作。

Secondly, there was a myth that these H-1Bs were these extremely highly skilled people.

其次 [00:08:48]，有一种迷思认为这些H-1B人才都是极度高技能的人才。

The average salary is slightly under $120,000.

但平均工资略低于12万美元 [00:08:57]。

If you told me that these were the best-in-class PhDs in all of these industries where companies are raising billions of dollars, you know that this salary would not pass the smell test.

如果你告诉我这些人是在那些筹集了数十亿美元的行业里最顶尖的博士 [00:09:14]，你就知道这个薪水是经不起推敲的。

The third thing is that a large plurality of these visas don't actually go to American companies that are looking to hire talent to make this American business do better.

第三点是 [00:10:03]，很大一部分签证实际上并未流向那些寻求人才以发展美国业务的美国公司。

These are foreign companies that are arbitraging labor and bringing people in.

这些是进行劳动力套利 [00:10:21] 并将人员引进的外国公司。

This is a very important reset.

这是一次非常重要的重置 [01:03:00]。

It is really important to remember that when you are in the United States for a master's or PhD, you already get an automatic visa—it's called 选择性实习培训 (Optional Practical Training/OPT: 允许在美国完成学业的国际学生，在与其专业领域相关的行业获得临时工作授权)。

重要的是要记住 [00:11:08]，当你来美国攻读硕士或博士学位时，你已经自动获得了签证，它被称为选择性实习培训 (Optional Practical Training/OPT: 允许在美国完成学业的国际学生，在与其专业领域相关的行业获得临时工作授权)。

This gives us a very good amount of time to figure out how exceptional they are, and then, quite honestly, I would gladly pay the $100,000 to get these guys on an H-1B program.

这给了我们足够的时间来判断他们有多优秀 [01:03:00]，老实说，我将很乐意支付10万美元让他们加入H-1B项目。

These changes are very good measures to course correct and get us towards what it was meant to be, which is to help American companies excel, get the best of the brightest.

这些改变是非常好的纠正措施 [00:11:59]，能够使项目回归其初衷，即帮助美国公司脱颖而出，吸引最优秀的人才。

The history of this is that it was something that was started after World War II to get really specialized people, like Polish and German geniuses building out rockets.

追溯历史 [00:12:06]，该项目始于二战后，目的是吸引真正专业的人才，比如波兰和德国的天才来建造火箭。

I think there should be two separate programs for what we could call highly skilled workers.

我认为应该为我们所说的高技能人才设立两个独立的计划 [00:13:58]。

What you were referring to after the end of World War II: there was a secret US operation called 回形针行动 (Operation Paperclip: 二战后美国一项秘密计划，旨在招募前德国科学家、工程师和技术人员到美国工作) where we tried to recruit German scientists and engineers.

你提到的二战后的计划 [00:14:04]，是美国一项名为回形针行动 (Operation Paperclip: 二战后美国一项秘密计划，旨在招募前德国科学家、工程师和技术人员到美国工作) 的秘密行动，我们试图招募德国科学家和工程师。

One could make the case that a similar sort of scenario should exist today—a second Operation Paperclip, perhaps a continuing process—where we identify some of the top scientists and top domains and go after those scientists proactively with government action.

有人可能会说 [00:14:53]，今天应该存在类似的场景——第二次“回形针行动”，或许是一个持续的过程——由政府主导，与私营产业合作，积极识别和争取那些顶尖领域的最优秀科学家。

The majority of papers today across nearly every scientific domain are being published out of China.

如今 [01:03:00]，几乎所有科学领域的多数论文都是在中国发表的。

There's a real case to be made that perhaps those scientists would be better off, and America would be better off, if they were pioneering their research here rather than there.

确实有理由认为 [01:03:00]，如果这些科学家能在美国而不是在中国进行研究和创新，对他们个人和美国都会更好。

The alternative would be to force a higher salary range such that you as a company are now basically being forced to pay a higher salary, which means you have to justify that this person is worth it to bring them in from ex-US.

另一种选择是强制要求提高薪资范围 [01:03:00]，这样公司就必须支付更高的工资，这意味着你必须证明这个人值得从美国境外引进。

We can always create a different class of visa for them.

我们随时可以为他们创建不同类别的签证 [00:20:46]。

We have the ability to do these things called 国家利益豁免 (National Interest Waiver/NIW: 绿卡申请中的一类，允许符合美国国家利益的高级人才在无雇主担保的情况下申请豁免劳工证)。

我们有能力进行被称为国家利益豁免 (National Interest Waiver/NIW: 允许符合美国国家利益的高级人才在无雇主担保的情况下申请豁免劳工证) 的操作 [00:20:46]。

If we can clean the decks on the H-1B stuff, it'll give people a lot more incentive to support the national interest waiver concept.

如果我们能够清理H-1B方面的乱象 [00:20:54]，人们将更有动力去支持国家利益豁免这一概念。

The term for the "Paperclip mode" is recruitment, then you have dissidents, which is compassion for true dissidents, and then there's family.

“回形针模式”的术语是“人才招募” [00:23:59]，然后是“持不同政见者”，这是对真正持不同政见者的同情，接下来是“家庭”。

Then you have this big thing in the middle, which is everybody in the world wants to live here—that's immigration.

接着是中间的一大类 [00:24:13]，即全世界的人都想住在这里，这就是“移民”。

We need to put it in three buckets and have thoughtful discussions about each.

我们需要将其归入三个类别 [00:24:17]，并针对每一个类别进行深思熟虑的讨论。

自闭症、泰诺争议与环境因素 [00:25:26]

Major news in science this week: let's talk about autism and the press conference that happened with Bobby Kennedy and President Trump.

本周的重大科学新闻 [00:25:26]：让我们谈谈自闭症，以及鲍比·肯尼迪和特朗普总统召开的新闻发布会。

Autism has increased dramatically over the years; we went from 1 in 10,000 in 1970 to 1 in 32 in 2022.

多年来 [00:25:45]，自闭症的发病率急剧增加；从1970年的1/10000上升到2022年的1/32。

I think autism, just like Alzheimer's, there may be several underlying conditions that lead to what we would call the 表型 (Phenotype: 生物体可观察到的特征或特性，由其基因型和环境相互作用决定) of autism.

我认为自闭症就像阿尔茨海默病一样 [00:26:17]，可能有几种潜在的疾病导致我们所说的自闭症表型 (Phenotype: 生物体可观察到的特征或特性，由其基因型和环境相互作用决定)。

One of the drivers that came up is that they've identified a receptor that absorbs folate, a type of vitamin B, and that folate receptor may be attacked by the immune system.

出现的一个驱动因素是 [00:26:41]，他们发现了一个吸收叶酸（一种B族维生素）的受体，而这个叶酸受体可能会被免疫系统攻击。

As a result, you can't really uptake vitamin B, and when those cells are dysfunctional, you end up having what looks like what we call autism.

结果是 [00:27:01]，你无法正常吸收维生素B，当这些细胞功能失调时，最终就会表现出我们所说的自闭症。

One of the things they announced is they're going to work on getting the label updated for leucovorin, which will resolve the folate receptor issue for many people.

他们宣布的其中一件事是 [00:27:12]，他们将致力于更新亚叶酸（leucovorin）的标签，这将为许多人解决叶酸受体的问题。

The other thing they brought up is a paper by Andrea Bacarelli, Dean of the Harvard T.H. Chan School of Public Health.

他们提出的另一件事 [00:27:26] 是哈佛陈曾熙公共卫生学院院长Andrea Bacarelli的一篇论文。

This paper analyzed several studies and showed that 27 out of 46 studies had a slightly positive association, meaning it was having some contributory effect to both ADHD and 自闭症谱系障碍 (Autism Spectrum Disorder/ASD: 一种神经发育障碍，特征为社交互动和沟通困难，以及重复刻板行为) when women would take 对乙酰氨基酚 (Acetaminophen: 一种常用的止痛和退烧药物，例如泰诺/Tylenol) while pregnant.

这篇论文分析了几项研究 [00:27:35]，显示46项研究中有27项存在轻微正相关，意味着孕妇在怀孕期间服用对乙酰氨基酚 (Acetaminophen: 一种常用的止痛和退烧药物，例如泰诺/Tylenol) 对多动症（ADHD）和自闭症谱系障碍 (Autism Spectrum Disorder/ASD: 一种神经发育障碍，特征为社交互动和沟通困难，以及重复刻板行为) 有一定的促成作用。

One of the underlyings might be this autoimmune condition associated with the folate receptor.

其中一个潜在原因 [00:28:43] 可能是这种与叶酸受体相关的自身免疫状况。

What causes autoimmunity is a whole 'nother conversation, and we can get into the vaccine stuff if you guys want to.

导致自身免疫的原因是另一个话题 [00:28:49]，如果你们想谈，我们可以讨论疫苗相关的内容。

It's pretty clear that in our modern world, there's a cumulative effect of environmental exposures that children are getting in the last couple of decades.

很明显 [00:29:04]，在过去的几十年里，我们现代世界的儿童所遭受的环境暴露有着累积效应。

Whether it's microplastics, chemicals in the food, environmental exposure in the air, or other things we're putting in our body, every one of these things might increase your chance of autism by a small percentage.

无论是微塑料、食物中的化学物质 [00:29:14]、空气中的环境暴露，还是我们摄入的其他物质，每一样都可能使你患自闭症的几率增加一个小百分比。

When you add up all the things in our environment, there may be a cumulative effect that results in different underlying conditions in our body that may result in what looks like things that we call autism.

当你把我们环境中的所有因素加起来 [00:29:42]，可能存在一个累积效应，导致我们身体内出现不同的潜在状况，这些状况可能表现为我们称之为自闭症的症状。

One of the controversies is that the paper by Bacarelli has been challenged because in 2023, the judge threw out his testimony as unreliable in a lawsuit against the maker of Tylenol, because he was being paid $150,000 as an expert witness.

其中一个争议是 [00:31:53]，Bacarelli的论文受到了质疑，因为在2023年针对泰诺制造商的诉讼中，法官认为他作为专家证人收取15万美元报酬的证词不可靠，并将其驳回。

This association study indicates an increased risk, but that doesn't mean that if you take Tylenol, your child is going to have autism; there is a statistical chance that there is a slight increase.

这项关联研究表明风险有所增加 [00:32:41]，但这并不意味着你服用泰诺你的孩子就一定会患自闭症；这只是一种统计学上的概率略有增加。

We know for sure that there is this potential autoimmune issue that you as the mother can express as an antibody.

我们确切地知道 [00:33:03]，存在这种潜在的自身免疫问题，母亲可以通过抗体表达出来。

There's an antibody test for it, and the child can also be tested for it.

有针对这种抗体的测试 [00:33:13]，孩子也可以接受测试。

At a minimum, we're now at a point where we can create a very thorough, well-funded study to get to the core of this issue.

至少 [00:33:20]，我们现在可以进行一项资金充足的、非常彻底的研究，以探究这个问题的核心。

I don't think you're adding to the conversation when you say that there's no determinism because on the off chance that there is, I would say that we don't know yet.

我不认为你说“没有决定性”是在为对话添砖加瓦 [00:34:26]，因为万一存在决定性，我想说的是我们尚不清楚。

This 叶酸受体自身抗体 (Folate Receptor Auto-antibody: 一种针对叶酸受体的抗体，可能导致叶酸吸收障碍，与某些神经系统疾病相关) test has been available since 2012.

这种叶酸受体自身抗体 (Folate Receptor Auto-antibody: 一种针对叶酸受体的抗体，可能导致叶酸吸收障碍，与某些神经系统疾病相关) 测试自2012年以来就已经可用了 [00:35:01]。

As scientists, we have to constantly interrogate, and this cannot be just a political point.

作为科学家 [00:42:06]，我们必须不断进行审视和质疑，这不能仅仅成为一个政治观点。

吉米·坎摩尔、政治暴力与仇恨言论立法 [00:43:46]

We've got a lot of news in the censorship space: Jimmy Kimmel is back on the air.

审查领域有很多新闻 [00:43:46]：吉米·坎摩尔（Jimmy Kimmel）已经复播了。

Disney explained why they suspended Kimmel last week: to avoid further inflaming a tense situation at an emotional moment for our country.

迪士尼解释了上周暂停坎摩尔节目的原因 [00:44:00]：是为了避免在我们国家情绪化的紧张时刻进一步煽动局势。

Kimmel's statement was heartfelt and deft; he said it was never his intention to make light of the murder of a young man.

坎摩尔的声明真挚而得体 [00:44:46]；他说自己从未有意轻视一位年轻人的遇害事件。

He didn't apologize, he just offered an explanation that it wasn't his intent.

他没有道歉 [00:46:59]，他只是解释说这不是他的本意。

Nobody was accusing him of making light of the murder; what he did, and what people were upset about, is that he lied and said that the shooter was MAGA.

没有人指责他轻视谋杀 [00:47:06]；他所做的，以及让人们感到不安的是，他撒谎称枪手是MAGA（让美国再次伟大）支持者。

I still think that his statement, let's call it an apology, was constructive and positive because at least he is showing empathy towards the other side.

我仍然认为他的声明 [00:47:24]，姑且称之为道歉，是建设性和积极的，因为至少他表现出了对另一方的同情。

In the current overheated political environment, expressing empathy for the other side is a positive statement, and I think he definitely brought the temperature down.

在当前这种过度狂热的政治环境中 [00:47:33]，对另一方表达同情是一种积极的表态，我认为他确实降低了温度。

He did not fess up to what he really did wrong here, which was to claim that the shooter was MAGA; that was the thing that was deeply offensive.

他没有坦白自己真正的错误 [00:48:35]，那就是声称枪手是MAGA支持者；这才是最令人反感的事情。

The left really does need to look in the mirror and rid itself of this poisonous ideology that political violence can be used to solve problems.

左翼确实需要自我反省 [00:49:06]，并去除这种认为政治暴力可以解决问题的有害意识形态。

This guy might have been crazy in the sense that he was willing to use murder to achieve his objectives, but it doesn't mean he wasn't animated by an ideology that lots of people believe.

这个人可能疯狂到愿意使用谋杀来实现其目标 [00:50:14]，但这并不意味着他没有受到很多人认同的意识形态的鼓动。

The proof of this was the celebratory reaction to the assassination of Charlie Kirk on TikTok, Blue Sky, and Reddit.

证据就是对查理·柯克（Charlie Kirk）遇刺事件在TikTok、Blue Sky和Reddit上的庆祝反应 [00:50:21]。

Thousands, maybe hundreds of thousands of people were celebrating his death and buying into this idea of political violence as the solution to their problem.

成千上万 [00:50:39]，甚至可能是数十万人在庆祝他的死亡，并认同政治暴力是解决他们问题的方案。

We showed you plenty of data last week showing that there were three times as many people on the left who are celebrating or endorsing political violence.

我们上周向你们展示了大量数据 [00:51:13]，显示左翼人士中庆祝或支持政治暴力的人数是右翼的三倍。

The difference here is that you saw thousands, maybe even upwards of a hundred thousand people on the left on social media rejoice and celebrate his assassination or downplay it, on the grounds that somehow he deserved it.

这里的区别在于 [00:58:18]，你在社交媒体上看到了数千，甚至多达十万的左翼人士为他的遇刺而欢呼庆祝或轻描淡写，理由是某种程度上他罪有应得。

There is this kind of hate speech bill that passed out of the Assembly and the Senate in California that's now on Gavin Newsom's desk to sign.

加州参众两院通过了一项关于仇恨言论的法案 [01:14:18]，目前已提交给加文·纽森（Gavin Newsom）签署。

It basically would fine social networks that allow content to show up that the state of California deems to be 仇恨言论 (Hate Speech Law: 立法中未被宪法定义或允许豁免的言论类别，通常指攻击或贬低特定群体的言论).

该法案实质上将对允许出现加州认定为仇恨言论 (Hate Speech Law: 立法中未被宪法定义或允许豁免的言论类别，通常指攻击或贬低特定群体的言论) 的社交网络处以罚款 [01:14:31]。

This could become a very slippery slope very fast, as social media companies, afraid of fines, block all that content.

这可能很快成为一个非常危险的滑坡 [01:14:37]，因为社交媒体公司担心巨额罚款，会封锁所有相关内容。

I think the bill you're referring to is SB 771, and it is an EU-style suppression of "hate speech" on social networks.

我认为你指的是SB 771法案 [01:15:12]，这是欧盟式在社交网络上压制所谓“仇恨言论”的尝试。

The problem is that there is no constitutional exception for hate speech under the First Amendment, and there is no definition of hate speech—it's just whatever the people in power say it is.

问题在于美国宪法第一修正案中没有仇恨言论的宪法例外 [01:15:25]，而且仇恨言论没有明确的定义——它完全取决于当权者如何界定。

Offensive content suddenly can get wrapped up and be called hate speech, and then the government starts to tell us all what we are and aren't allowed to see or say.

冒犯性内容可能突然被归类为仇恨言论 [01:16:22]，然后政府就开始告诉我们什么可以看、什么可以说了。

AI研究突破与能源效率 [00:59:28]

Sacks and I are going to wrap it up with a quick conversation on AI.

萨克斯和我将以一段关于AI的简短对话来结束本期节目 [00:59:28]。

There were two papers that were published this week, each of which on their own was pretty important.

本周发表了两篇论文 [00:59:32]，每一篇都相当重要。

The first is the MIT paper called "Teaching LLMs to Plan," which created an instruction tuning framework that teaches LLM to do 符号规划 (Symbolic Planning: 一种人工智能方法，涉及使用符号表示（如逻辑语句）来制定解决问题的步骤序列).

第一篇是麻省理工学院的论文 [01:00:00]，名为《教导大型语言模型（LLM）进行规划》，它创建了一个指令调优框架，用于教导LLM进行符号规划 (Symbolic Planning: 一种人工智能方法，涉及使用符号表示（如逻辑语句）来制定解决问题的步骤序列)。

This means that the LLM can think step-by-step, or "chain of thought," in a smarter way by making them generate explicit state-action-state chains.

这意味着LLM能够通过生成明确的“状态-行动-状态”链 [01:00:09]，以更智能的方式进行循序渐进或“思维链”式的思考。

They trained that model by giving them feedback with an external plan validator.

他们通过一个外部计划验证器向模型提供反馈来进行训练 [01:00:27]。

They were able to achieve planning accuracy of up to 94% on some standardized benchmarks, a 66% absolute improvement over baseline models.

他们 [01:00:45] 在一些标准化基准测试中实现了高达94%的规划准确率，比基线模型绝对提高了66%。

This sort of a system can be used to train LLMs to do better reasoning and better chain of thought in such a dramatic way that LLMs will look like they are starting to reason.

这种系统可用于以惊人的方式训练LLM [01:01:08]，使其具备更好的推理和思维链能力，LLM看起来将像是开始真正进行推理。

The symbolic framework they are talking about is PDDL, or 规划领域定义语言 (Planning Domain Definition Language/PDDL: 一种用于描述人工智能规划问题的标准语言), an attempt to standardize AI planning languages that has been around for a long time.

他们所说的符号框架是PDDL [01:01:59]，即规划领域定义语言 (Planning Domain Definition Language/PDDL: 一种用于描述人工智能规划问题的标准语言)，这是一种旨在标准化AI规划语言的尝试，已存在很长时间。

The other paper is from a team in Germany, published in the journal Nature Computational Science.

另一篇论文来自一个德国团队 [01:03:15]，发表在《自然计算科学》杂志上。

These folks were able to reduce the physical memory size that's needed to run the attention window on a GPU.

他们成功减少了在GPU上运行注意力窗口所需的物理内存大小 [01:03:28]。

As a result, the energy and the total token cost to run 推理 (Inference: 指人工智能模型使用训练好的参数对新数据进行预测或生成输出的过程) went down significantly.

结果是 [01:03:51]，运行推理 (Inference: 指人工智能模型使用训练好的参数对新数据进行预测或生成输出的过程) 所需的能源和总的Token成本显着下降。

This mechanism, if it scales, can have a pretty dramatic effect on the energy consumption needed to run AI.

这种机制如果能够规模化 [01:04:34]，将对运行AI所需的能源消耗产生相当大的影响。

Because you need far less memory, you can actually move a lot of AI inference to the edge of the network.

由于所需内存大大减少 [01:04:47]，你实际上可以将大量AI推理转移到网络的边缘。

This means you could put a very high-powered LLM model that could be run in a robot or on your phone that historically you'd need to run in a data center.

这意味着你可以将一个功能强大的LLM模型部署在机器人或你的手机上运行 [01:04:56]，而历史上你需要在一个数据中心才能运行它们。

This architecture could be one of these big architectural breakthroughs that are coming in AI that could ultimately lead to many orders of magnitude reduction in the energy cost needed to run inference.

这种架构可能是AI领域即将到来的重大架构突破之一 [01:05:12]，最终可能使运行推理所需的能源成本降低多个数量级。

I think the writing was on the wall that models are going to get smaller and smaller and more efficient to the point where they can run on the edge on local devices.

我认为 [01:06:08]，模型将变得越来越小、越来越高效，直到可以在本地设备上运行的趋势早已显现。

I don't pay a lot of attention to papers; I do pay a lot of attention to product launches.

我不太关注学术论文 [01:07:07]，我更关注产品发布。

I think we're going to need a lot more power, a lot more electricity, and we haven't even gotten to the robot revolution yet.

我认为我们将需要更多的电力和能源 [01:07:46]，而且我们甚至还没有迎来机器人革命。

If this thing's even close to being correct, then you could run the most sophisticated LLMs in a robot without it needing to be run out of a data center going forward.

如果这个研究接近正确 [01:08:02]，那么将来你就可以在机器人中运行最复杂的LLM，而无需依赖数据中心。S