← 返回任务列表

Why Companies Want AI They Can Own

259 段 · 1 位说话人 · 原片 26:34
M1
M10:00

一个尚未发布的open-weight AI模型引起了不小轰动。

A yet-to-be-released open-weight AI model is getting a lot of buzz.

人们说,这个模型可能把open-source AI的王冠带回美国。

This could be the model, people say, that brings the open-source AI crown back to US shores.

不过,我感兴趣的不是模型本身,而是美国围绕open-weight模型不断变化的讨论。

What's interesting to me, though, is less the model itself and more the evolving discourse around open-weight models in the US.

现在这越来越不只是一个话题,而是三个:AI安全话题、国家安全话题和企业战略话题。

Increasingly, this is not just one conversation, but 3: an AI safety conversation, a national security conversation, and an enterprise strategy conversation.

那么,如果这些讨论可能指向不同方向,最后哪个会胜出?

Now, if those conversations point in potentially different directions, which will win out?

它们将如何调和?

How will they be reconciled?

今天我们来探讨一下目前开源AI在美国的现状,以及它未来的发展方向。

Today we explore where open-source AI is in the US right now and where it might head next.
M1
M11:09

hyperscalers们逐渐明白,他们不能再把数据中心周边社区的态度当作次要事项。

Slowly but surely, the hyperscalers are getting the message that they cannot treat community attitudes around data centers as a secondary priority.

最新的例子来自Amazon,他们周五承诺未来五年投入超过10亿美元,用于数据中心周边的社区项目。

The latest example comes from Amazon, who on Friday committed to spending more than $1 billion over the next 5 years on community projects surrounding their data centers.

此外,公司还表示,他们已停止使用保密协议来隐瞒与地方官员达成的协议。

In addition, the company has said that they've stopped using nondisclosure agreements to keep their deals with local officials under wraps.

这些承诺来自AWS首席执行官Matt Garman一篇3000字的文章。他在做出这些承诺的同时,还恳请公众把数据中心视为现代生活的关键基础设施。

The commitments came as part of a 3,000-word essay from AWS CEO Matt Garman, who, in addition to making those commitments, implored the public to think about data centers as critical infrastructure for modern life.

他称数据中心的建设是一场我们的国家输不起的比赛,并将其比作州际公路系统的建设。

He called the data center buildout the race our nation can't afford to lose and compared it to the construction of the interstate highway system.

Garman写道,任何变化都会引发重要问题,但也会伴随错误信息和彻头彻尾的谎言。

Garman wrote, with any change, there will be important questions raised, but there will also be misinformation and outright lies.

在社交媒体和全天候新闻的时代,误解会比以往任何时候都更快地扎根。

And in the age of social media and 24/7 news, myths take hold faster than ever before.

事实上,这种建设在地缘政治上如此重要,以至于有广泛报道称,多个国家故意在美国散布关于数据中心的虚假信息,诱骗我们放慢速度。

In fact, this buildout is so important geopolitically that there are widespread reports of various countries intentionally seeding misinformation in the U.S. about data centers to trick us into slowing down.

Garman指出,全国有100项数据中心建设暂停措施正在审议中。他接着说,如果这些措施通过,美国可能就是在给自己写一张输掉比赛的入场券,后果将影响几代人。

Noting 100 data center moratoriums being considered across the country, Garman continued, if these measures are enacted, the U.S. could be writing its own losing ticket to this race, and the consequences would last generations.

作为一个国家,我们不能让自己落到那种境地。

As a country, we can't afford to find ourselves in that position.

新的社区承诺包括了那些正在迅速成为标准的所有承诺:防止当地能源价格上涨、创造本地就业机会,以及确保社区有更多机会讨论数据中心项目。

The new community pledge included all of the commitments that are quickly becoming standard: preventing increases to local energy rates, creating local jobs, and ensuring communities have more opportunities to discuss data center projects.

这一公告的发布方式在媒体上引起了一些反弹。

The approach to the announcement saw some pushback in the press.

The Verge几乎没有提及新的社区承诺,而是聚焦于它所称的可怕的博客文章,该文章警告社区不要阻止数据中心。

The Verge barely touched on the new community pledges, focusing instead on what it called the scary blog warning communities not to block data centers.

部分问题在于,Amazon把他们的新承诺埋在了很长一段辟谣内容之下,那段内容在驳斥关于数据中心的误解。

And part of the issue was that Amazon buried their new pledge beneath a long section debunking the myths around data centers.

Tom's Hardware写道,Garman文章的第一部分实际上回应了这些项目附近社区居民提出的许多担忧,并将这些担忧称为虚假和误导,尽管有各种报道对公用事业涨价和噪音污染提出了关切。

Tom's Hardware wrote, the first part of Garman's post actually addressed many of the concerns raised by community members near these developments, calling them false and misleading, despite various reports that have raised concerns about utility hikes and noise pollution.

从我个人的角度看,虽然我很高兴看到这类承诺开始变得常态化,但我确实认为这篇帖子相当不接地气、也不太明智,尤其是在试图打动那些数据中心将落地的社区方面。

From my perspective, while I am very glad to see these sort of commitments starting to become normalized, I tend to agree that this post was fairly tone deaf and ill-advised to the extent that it was trying to actually appeal to the communities where these data centers are going to happen.

很难想象有人会因为数据中心开发商说他们的真实担忧是假的就改变看法。

It's pretty difficult to imagine anyone will change their opinion because a data center developer said that their genuinely held concerns were false.

不管这是对是错,这些公司必须意识到,他们打交道的对象是非常不同的群体。

And whether that's right or wrong, these companies have to realize that they have very different constituencies that they're dealing with.

社区里的普通人远远不会考虑什么“国家输不起的比赛”。

The individual in a community has vastly less consideration for the race our nation can't afford to lose.

他们在乎的是自己的生活,以及身边人的生活,就这。

What they care about is their own lives and the lives of people near them, period.

把本来是积极的承诺包裹在一层公关和辟谣的外衣里,只会削弱这个承诺,正如The Verge写的那篇文章所证明的那样。

Wrapping what is an otherwise positive pledge in a blanket of PR and myth-busting just undermines the pledge, as witnessed by the article that The Verge was able to write.

就目前而言,数据中心建设者们采取的行动是积极的。

For now, the moves that the data center builders are making are positive.

但这种说辞还需要很多改进。

The pitch still needs a lot of work.

说到这个宣传需要改进,Sam Altman 给所有不喜欢 AI 的人送上了一大袋弹药。

And speaking of the pitch needing some work, Sam Altman has handed a huge new bag of ammunition to everyone who dislikes AI.

在接受 Politico 采访时,Altman 说,引用他的话,OpenAI 和 Anthropic 在 AI 安全上的观点有很大的分歧。

In an interview with Politico, Altman said that there was a, quote, lot of daylight between OpenAI and Anthropic's views on AI safety.

当被问到具体细节时,Altman 回答说,我们认为世界应该接受一些坏事发生,以换取这项技术的好处和人们拥有自主权。

Asked for specifics, Altman responded, we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency.

现在,如果你坐在那里想,这肯定只是 Altman 试图做出的更全面评估和更细致论证的一部分,你当然是对的。

Now, if you are sitting there thinking to yourself, well, that has to be just part of a larger assessment, and a much more nuanced argument that Altman is trying to make, you would of course be right.

很明显,他并不是事先准备好那句话,想要在采访中特意说出来以强调某个重点。

It's pretty clear he didn't come into this interview with that line prepared, wanting to drop it for some big emphasis.

但总有一天,那些安排自己参加这些采访的人会意识到,不管他们喜不喜欢,我们生活在一种金句文化中。

But at some point, the folks who are scheduling themselves on all of these interviews are going to realize that whether they like it or not, we live in a soundbite culture.

如果你说,我们认为世界应该接受一些坏事发生以换取这项技术的好处,那接下来几天,所有人只会报道这一句。

And if you say, we believe the world should accept some bad things happening for the benefits of this technology, that's gonna be the only thing that anyone publishes for the next couple of days.

那么,关于 Altman 实际上想深入讨论的内容,就是围绕当前的安全提案,以及 OpenAI 与 Anthropic 在哪些方面一致,哪些方面不一致。

Now, as to what Altman was actually trying to get into, the discussion was around current safety proposals and where OpenAI agreed with Anthropic and where they didn't.

例如,在一致方面,Altman 说双方都支持第三方安全审计员等措施。

For example, in terms of agreement, Altman said that both support things like third-party safety auditors.

然而,Altman 试图提出的一个重大区别是关于监管俘获的风险。

However, one of the big distinctions Altman tried to present was around the risk of regulatory capture.

Altman 将 Anthropic 的观点描述为:这项技术会变得如此强大,如此危险,以至于旧金山的一家实验室应该掌控它,确保不出任何坏事。

Altman characterized the Anthropic view as, this technology is going to get so powerful and it's so dangerous that a single lab in San Francisco should have it and make sure nothing bad happens.

并且要设法分配它的好处。

And kind of figure out how to dole out its benefits.

然而,Altman 本人称这是一种完全不可接受的权衡。

Altman, for his part, though, called this a completely unacceptable trade-off.

他继续说,我不会接受这样的交易,即我们确保没有重大黑客攻击,没有技术滥用,没有诈骗,没有所有其他可能发生的坏事,因为我认为人们做的善事将比坏事多出几个数量级。

He continued, I wouldn't take a trade of saying, we'll make sure there's no major hacks, there's no misuse of this technology, there's zero scams, there's zero all the other bad things that will happen, because I think people will do tremendously orders of magnitude more good stuff than bad stuff.

换句话说,他触及的是 AI 领域经典的安全与自由的两难问题,并且提出了一个观点——他正确地指出,在主流媒体中这个观点没有得到充分阐述——即严格限制 AI 的使用或将其限制在某个特定群体中,本身也带来一系列风险,这些风险至少与 AI 安全讨论中提到的任何问题一样真实。

In other words, he is getting at the AI version of the classic conundrum of safety versus freedom, and is making the case which he is accurate has not been made enough in the popular press, that tightly controlling the use of AI or limiting it to some very specific group of people comes with its own set of risks that are certainly no less real than anything being discussed in the AI safety discourse.

总的来说,OpenAI 确实似乎在 AI 安全问题上试图开辟越来越不同的领域。

Overall, it definitely seems like OpenAI is trying to carve out increasingly different territory when it comes to AI safety.

但不幸的是,至少在这一轮讨论中,没有人能看到除了“我们应该接受一些坏事”这句话之外的东西。

But unfortunately, at least for this round in the discourse, no one was seeing beyond the line that we should accept some bad things.

然而,如果坏事真的发生,财政部长 Scott Bessent 已经强调,当它们发生时,那些促成这些事的前沿实验室要为自己的行为负责。

To the extent that bad things do happen, however, Treasury Secretary Scott Bessent has reinforced that when they do, the frontier labs who enabled them are responsible for their own actions.

上周在白宫举行的 AI 峰会最终达成了前沿实验室之间的协议,该协议规范并重申了第三方审计师和董事会监督的概念。

Last week's AI summit at the White House ended in an accord between the frontier labs, which normalized and reaffirmed the concept of third-party auditors and board-level oversight.

在上周六接受 Axios 采访时,Bessent 说,我们想要安全的加速。

In an interview with Axios on Saturday, Bessent said, we want safe acceleration.

实验室里的人必须承担责任,我同意这一点。

It's the people at the labs who have to accept responsibility, and I agree with that.

我认为实验室也已经接受了这种想法。

I think the labs have come around to that way of thinking too.

Bessent 还对 AI 领域的一系列其他热点问题发表了看法,这代表了财政部长的一些新思考。

Bessent weighed in on a range of other hot-button issues in AI, representing some updated thinking for the Treasury Secretary.

关于 AI 风险总体,Bessent 批评了某些方面的言论,他评论说,我们必须要为各种情况做好准备,但另一方面,AI 社区中一些人没有解决方案的危言耸听,那不是领导力。

On AI risk generally, Bessent scolded the communication from some quarters, commenting, I believe that we have to be prepared for every occasion, but on the other side, this alarmism without solutions by some of the AI community, that's not leadership.

记得在 Mythos 推出的早期阶段,Bessent 似乎完全相信该模型是危险的,不适合公开发布,这表明他的观点至少有一些轻微的改变。

Remember that during the early stages of the Mythos rollout, Bessent seemed fully convinced the model was dangerous and unsuitable for general release, suggesting at least some slight change in perspective.

Bessent 还驳斥了 AI 泡沫的说法,指出像微软这样的行业领导者正从他们的基础设施支出中获得可观的回报。

Bessant also dismissed the notion of an AI bubble, pointing out that industry leaders like Microsoft are seeing solid returns from their infrastructure spend.

总的来说,Bessent 似乎仍然有顾虑,但在表达越来越细致的 AI 安全观点,超越了简单的放缓要求。

Overall, Bessant appears to still have concerns but is expressing increasingly nuanced views of AI safety that go beyond some simplistic slowdown.

现在,说到上周的白宫 AI 活动,在特朗普总统宣布将不再使用“人工智能”这个术语后,Elon Musk 通过发帖表示支持总统的改名,他说:“不再叫 AI,叫 SI,更好。”

Now, speaking of last week's White House AI events, after President Trump declared that the term will no longer be artificial intelligence, Elon Musk showed his support for the president's name change by posting, no more AI, SI, it's better.

一位 X 用户呼吁他将 Space XAI 改名为 Space XSI,Musk 回应说,是的,我们会做出这个改变。

After an X user called on him to rename Space XAI to Space XSI, Musk responded, yes, we will make that change.

不过,朋友们,目前我没有任何打算将 AI Daily Brief 改名为 SI Daily Brief。

At the moment, however, friends, I do not have any intentions of renaming the AI Daily Brief to the SI Daily Brief.

最后今天,有个有趣得多的话题。

Lastly today, a much more fun one.

Meta 开源了他们的代码,让用户自己打造 Muse 设备。

Meta has open-sourced their code to let users build their own Muse gadgets.

当 Mark Zuckerberg 在 Meta Connect 上展示 Muse 吊坠设备时,它看起来像是硬件部门用现成组件拼出来的东西。

When Mark Zuckerberg unveiled the Muse charm device at Meta Connect, it seemed like something the hardware division had cooked up from off-the-shelf components.

Zuckerberg 展示了一个能工作的原型,但说他们还在敲定最终的物料清单,准备在12月发布。

Zuckerberg showed off a working prototype, but said that they're still nailing down the final build sheet ahead of a December release.

很快,人们就开始用便宜的通用硬件自己拼装版本了。

Immediately, people started hacking together their own version from cheap generic hardware.

而 Meta 对此的回应是:尽管去做。

And Meta's response on this one was go right ahead.

Nat Friedman 宣布了针对 ESP32 设备的 Muse 固件开源构建,以及 Linux SDK。

Nat Friedman announced an open-source build of Muse firmware for ESP32 devices and the Linux SDK.

硬件极客现在可以轻松在 Raspberry Pi 或 ESP32 开发板上构建完全可定制的 Muse 设备了。

Hardware hackers can now easily build a fully customizable Muse gadget on top of a Raspberry Pi or ESP32 dev board.

Meta 自己也拼了一个新设备,叫 Muse Home Link。

Meta has also hacked together their own new gadget with the Muse Home Link.

这个设备是一个小型 USB 供电的 Wi-Fi 连接器,让 Muse 能接入你的智能家居系统。

The device is a small USB-powered Wi-Fi link that lets Muse tap into your smart home setup.

Meta 正在生产5000个这种设备,等它们准备好发货时,我们会免费送给 Muse 订阅用户。

Meta is making a batch of 5,000 of these devices, and we'll be giving them away to Muse subscribers once they're ready to ship.

结果掀起了巨大浪潮,大家把 Muse 移植到了大量各种设备上。

The result was a huge wave of people porting Muse onto a ton of random devices.

Robert Soriano 翻出了他的旧 PSP,把它变成了一个 Muse 设备,还支持麦克风。

Robert Soriano dusted off his old PSP and turned it into a Muse device, complete with microphone support.

Wes Bos 让它跑在了一个会议胸牌上,把固件移植到了 MicroPython。

Wes Bos got it running on a conference badge, porting the firmware across to MicroPython.

Riley Brown 直接在 Mod Retro Game Boy 上安装了 Muse,只是让 Codex 帮忙移植了一下。而 Meta 工程师 Jessick Min 走了更简单的路,把 Muse 装到了一个50美元的 ESP32 设备上。

Riley Brown installed Muse on his mod-retro Game Boy by simply asking Codex to port it over. While Meta engineer Jessick Min went the easier route, loading Muse onto a $50 ESP32 device.

他目前用它当扫描仪来盘点自己的红酒收藏,但指出可能性是无限的。

He's currently using it as a scanner to inventory his wine collection, but notes that the possibilities are endless.

基本上,任何有 Wi-Fi 连接的设备现在都可以当作 Muse 设备了。

Basically any device with a Wi-Fi connection can now act as a Muse gadget.

不过,也不是夸大其词——考虑到我们在 AI 设备上见过多少次出师不利,或许这种让大家自己动手做东西的方法,才是探索智能设备应有形态的更好策略。

Now, not to overstate this, but given how many false starts we've seen when it comes to AI devices, maybe this approach of just letting people build things is a better strategy for figuring out what the form factor for agentic gadgets should be.

从历史上看,推出专用设备比如智能音箱,一直是个高风险的做法。

Historically, launching a dedicated device like a smart speaker has been a risky play.

这些设备利润率低,而且如果功能不完美,很容易在市场上失败。

The devices are low margin and prone to failing in the market if they don't function perfectly.

Meta 反其道而行之,直接面向技术社区,给他们开源固件,让他们用在手头任何硬件上。

Meta is instead going straight to the tech community and giving them open-source firmware to use on whatever hardware they happen to have lying around.

虽然这一系列设备可能只适合那些有足够技术能力自己动手的爱好者,但如果那个所谓的智能设备的正确形态真的来自这些实验之一,我一点也不会惊讶。

And while this set of devices might only be for those hobbyists who have enough technical capability to do it themselves, it wouldn't at all surprise me if the quote-unquote right form factor for an agentic device actually comes out of one of these experiments.

而不是只在实验室里闭门造车。

Instead of just brainstorming inside one of the labs.

如果你想看更多例子,可以去 X 上搜索 Muse Gadgets。

If you wanna see more examples, go to an X search for Muse Gadgets.

不过现在,新闻部分就到这里。

For now, however, that's gonna do it for the headlines.

接下来是主要环节。

Next up, the main episode.

KPMG 和德克萨斯大学奥斯汀分校的一项新研究发现,当人和 AI 一起工作时,相似的技能并不能保证相似的结果。

A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don't guarantee similar outcomes.

研究人员研究了500多名早期职业人士,发现表现最好的人通过指导、评估和优化 AI 的输出,持续放大了 AI 的价值。

Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs.

这些顶尖表现者被称为 AI 放大器,他们的定义不是仅仅靠知道什么,而是靠如何与 AI 合作。

These top performers, called AI amplifiers, weren't defined by what they knew alone, but by how they worked with AI.

想了解更多关于 AI 放大器与其他人的区别,请访问 kpmg.com/us/aiamplifiers。

Learn more about what separates AI amplifiers from everyone else at kpmg.com/us/aiamplifiers.

这里有个残酷的事实。

Here's a harsh truth.

你的公司可能正在花数千或数百万美元在AI工具上,但这些工具被严重低估使用。

Your company is probably spending thousands or millions of dollars on AI tools that are being massively underutilized.

一半的公司有AI工具,但只有12%真正用于创造商业价值。

Half of companies have AI tools, but only 12% use them for business value.

大部分员工仍然只是用AI来总结会议纪要。

Most employees are still using AI to summarize meeting notes.
M1
M112:48

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

今天我们聊聊美国开放权重AI的话题。

Today we are talking about open weight AI in the United States.

这个讨论在过去一年里已经发生了很大变化。

This is a conversation that's evolved quite a bit in the last year.

某种程度上,它的第一个重大时刻是2025年一月的DeepSeek时刻,当时中国实验室DeepSeek在一个免费应用里推出了一款推理模型。

Its first big moment in some ways was the January 2025 DeepSeek moment, where Chinese lab DeepSeek dropped a reasoning model in a free app.

这成了很多人第一次接触推理模型,跟当时市面上其他模型的体验很不一样。

Which ended up being most people's first experience with the reasoning model, which was very different than their experience using the existing crop of models that were available at that time.

那个时刻引起了市场的注意,导致NVIDIA市值一天内蒸发了5800亿美元,当然,担忧在于,如果大家纷纷转向更便宜的中国模型,那么支撑整个经济的那些基础设施成本可能就没那么合理了。

That moment caught the market's attention to the tune of $580 billion wiped off NVIDIA's market cap in a single day, with the concern, of course, being that if cheaper-to-produce Chinese models were going to become what everyone shifted to, then maybe these infrastructure costs, which were propping up the entire economy, didn't really make as much sense.

随着时间的推移,这个说法发生了变化,尽管它从未完全消失。

Over time, that narrative shifted, although it never fully went away.

到了2026年,围绕开放权重模型的讨论又发生了变化。

And in 2026, the discussion around open weight models has changed yet again.

今年,关于开放模型的讨论更多地集中在它们如何可能解决关键问题,不仅是对愿意探索未知领域的初创公司和开发者,也包括那些担心成本和数据主权的企业。

This year, the discourse around open models is much more about how they potentially solve key issues, not just for startups and developers who are willing to play in uncharted waters, but for enterprises who are worried about costs and data sovereignty.

过去几个月的一个大问题是,美国是否会在这一块市场参与竞争,还是干脆全部让给中国?

And one of the big questions for the last few months has been whether this was a section of the market that the US was going to compete in, or was it just going to cede it all to China?

然后努力在闭源前沿模型和最先进技术上保持领先。

And try to stay ahead on the closed-source frontier models and the state-of-the-art.

Axios的一份新报告引起了很大关注,甚至是超出比例的关注,我觉得这显示了大家对这一领域的兴趣。

A new report from Axios is getting a lot of buzz— a disproportionate amount, in fact— that I think shows the interest in this particular area.

周日,Axios发布了一篇文章,标题是《独家:强大开放模型将震撼AI竞赛》。

On Sunday, Axios dropped a piece called Scoop: Powerful Open Model Is Set to Shake Up AI Race.

传闻中的模型来自Reflection AI,这家公司自2024年初成立以来一直有些神秘。

The rumored model is coming from Reflection AI, who themselves have been a bit of an enigma since their founding in early 2024.

这家公司筹集了数十亿美元,在没有产品的情况下就达到了250亿美元的估值。

The company raised several billion dollars and reached a $25 billion valuation without a product.

他们五月份赢得了五角大楼的合同,并与SpaceX签署了价值63亿美元的算力协议。

They won a Pentagon contract in May and signed a $6.3 billion compute deal with SpaceX.

Reflection AI确实把自己定位为美国对日益流行的中国开放权重模型的制衡力量,但没有现成可用的模型,很多人认为这只是一个融资话术。

Reflection AI did position themselves as a US-based counterweight to the growing popularity of Chinese open weight models, but without a currently available model, many assumed this was just a fundraising pitch.

然而周日,Axios报道称Reflection AI正准备本月晚些时候发布他们的第一款模型,传闻说这个模型相当不错。

On Sunday, however, Axios reported that Reflection AI is preparing to release their first model later this month, with rumors that it's pretty good.

消息人士称该模型将与中国竞争对手有竞争力,并且将,引述,

Sources said that the model will be competitive with Chinese rivals and will, quote,

拥有足够强大的智能,可以帮助企业构建自己的专有低成本AI系统。

boast powerful intelligence capable enough to help companies build their own proprietary low-cost AI systems.

而这当然引起了我的兴趣,也应该引起那些在考虑企业AI部署的人的兴趣。

And this is of course where it piques my interest and should be of interest to those of you who are thinking about enterprise AI deployments.

根据报道,Reflection的目标是建立一个AI工厂——一个旨在帮助企业基于Reflection的模型创建自己的AI系统的系统。

According to the report, Reflection's goal is to build an AI factory— a system designed to help enterprises create their own AI systems based on Reflection's models.

Reflection已经在试点这种方法,最近宣布与韩国新世界集团建立主权AI工厂合作伙伴关系。消息人士称,Reflection上周在华盛顿举行了简报会,讨论新模型和AI工厂这一概念。

Reflection is already piloting this approach, recently announcing a sovereign AI factory partnership with Shinsegae Group in South Korea. And sources said that Reflection held briefings in Washington last week to discuss the new models and this concept of AI factories.

他们确实已经为这次发布预定了大量的算力。

They've certainly been securing a ton of compute in anticipation of this launch.

除了与SpaceX的数十亿美元交易外,Reflection还与Nebius有另外十亿美元的合同。

In addition to the multi-billion-dollar SpaceX deal, Reflection has another billion dollars in contracts with Nebius.

那么,这些都很好,都非常酷。

Now, all of that's great, all of that's very cool.

看看Reflection会发布什么,会很有意思。

It'll be interesting to see what Reflection puts out.

但比单个模型及其工厂策略更重要的是它所代表的趋势。

But what matters much more than the individual model and their approach to factories is the trend that it represents.

实际上,消息人士告诉Axios,还有几家美国实验室准备在本月发布开源模型。

In fact, sources told Axios that several other U.S. labs are preparing to launch open models this month.

我们听到了传言说NVIDIA正在训练Nemotron 4,而且我们可能很快会看到Thinking Machines Lab的新版本。

We've heard rumors that NVIDIA is training Nemotron 4, and we could be due for a new release from Thinking Machines Lab.

Andrew Curran总结写道:美国开源的复兴即将开始?

Summing it up, Andrew Curran wrote, American OSS renaissance about to begin?

Howard Lutnick一直以来都希望看到这样的局面,据说他一度甚至考虑直接通过政府拨款来推动这件事。

Howard Lutnick has been hoping for something like this for some time, and at one point was reportedly even considering direct government funding to get the ball rolling.

美国政府希望美国开源模型能在全球范围内与中国模型竞争。

The U.S. government wants American open models to compete with Chinese models globally.

Andrew Curran所指的讨论是今年夏天发生的那场争论,当时有一度看起来美国政府可能会把开源权重模型纳入他们的测试体系。

The discussion that Andrew Curran is referring to is the one that happened over the summer when it seemed for a moment like the U.S. government could be trying to include open weight models in their testing regime.

很多人认为,如果这个举措真的实施,实际上会扼杀掉这个行业的那一部分。

Many thought that that move, if it actually happened, would effectively snuff out that segment of the industry.

当然,这些呼吁是在Kimi K3发布后不久出现的,媒体很快将其归为开源模型。

The calls, of course, came shortly after the release of Kimi K3, which the media was very quick to jump on as open source models.

根据《连线》当时对白宫辩论的报道,商务部长Howard Lutnick是支持竞争而非禁止的关键人物。

According to Wired's coverage of that White House debate at the time, Commerce Secretary Howard Lutnick was a key defender of competition over bans.

他们在七月曾写道:Lutnick一直在考虑如何激励美国顶尖实验室打造自己的开源权重模型以制衡中国,并且在最近几周与多家AI实验室的负责人进行了交流。

Back in July, they wrote, Lutnick has contemplated ways to create incentives for top U.S. labs to create their own open-weight models to counterbalance China and has spoken with leaders at a number of AI labs in recent weeks.

Lutnick似乎在监管上采取了一种中间立场。

Lutnick appears to be straddling a middle ground of regulation.

他对Anthropic实施了出口管制以使其就范,但比其他官员更加宽松。

He imposed export controls on Anthropic to bring them to heel, but has been more freewheeling than others.

一个长期的战略问题以及不同观点之间的分歧在于:美国保持领导地位的正确做法是切断中国获取能制造先进模型的输入资源,还是试图掌控从闭源模型到开源模型的整个技术栈,让整个世界从硬件到软件都完全运行在美国的AI基础设施之上。

One of the long-term geostrategic questions and divides between different people who have different takes on this has been whether the right approach to U.S. leadership is to try to cut off China's access to the inputs by which they can make advanced models, or to try to own the entire stack from closed models to open models so that the world runs end-to-end on U.S. AI infrastructure, both from a hardware and software perspective.

关于这一点从未有过真正的战略一致性,但随着讨论的推进,权力走廊里这种反复争论一直在进行。

At no point has there ever really been any strategic coherence around that, but that is the constant back and forth that's happening in the halls of power as the conversation evolves.

不过有一家公司一直是美国开源的倡导者,它要么低调要么不那么低调,取决于你有多关注——那就是NVIDIA。

One company though that has quietly, or not so quietly depending on how well you're paying attention, been a champion of U.S. open source is NVIDIA.

这家公司现已训练了多代大语言模型,其中Nemotron 3.5得到了大量使用,而Nemotron 4有望接近前沿水平。

The company has now trained multiple generations of LLMs, with Nemotron 3.5 seeing significant use and Nemotron 4 promising to be relatively close to the frontier.

除了大语言模型,NVIDIA还在机器人和自动驾驶等垂直领域训练了模型。

Alongside LLMs, NVIDIA has also trained models in verticals like robotics and self-driving cars.

也许并不意外,NVIDIA似乎认为他们的GPU是核心产品,而这一核心产品将从自由可用的AI模型更广泛的使用中大大受益。

Perhaps unsurprisingly, NVIDIA seems to hold the view that their GPUs are the core product, that core product stands to benefit dramatically from more use of freely available AI models.

他们的外部投资也表明了对开源模型的坚定承诺。

Their outside investments also speak to a strong commitment to open models.

NVIDIA是Reflection AI以及其他几家开源实验室的主要支持者。

Nvidia is a key backer of Reflection AI, as well as several other open-source labs.

当然,最近他们还对Hugging Face进行了巨额投资,上个月花了129亿美元收购了该平台。

And of course, more recently, they made that huge investment into Hugging Face, spending $12.9 billion to acquire the platform last month.

在今年七月的NVIDIA开发者大会上,当开源权重模型监管问题被正式摆上台面时,NVIDIA CEO黄仁勋动情地呼吁保持这些模型的可用性。

At Nvidia's developer conference in July, when open-weight model regulation was firmly on the table, Nvidia CEO Jensen Huang made an impassioned plea to keep the models available.

他说:研究人员需要开源。

He said, researchers need open source.

开发者需要开源。

Developers need open source.

全世界的公司都需要开源。

Companies around the world need open source.

开源模型真的真的非常重要。

Open source models are really, really important.

我们在开源贡献方面领先。

We lead in open source contribution.

我们在排行榜上有23个模型。

We have 23 models on leaderboards.

我们有所有这些不同的领域,从语言模型到物理AI模型再到生物学模型。

We have all these different domains from language models to physical AI models to biology models.

每一个这样的模型都有庞大的团队。

Each one of these models has enormous teams.

我们专注于这个领域,原因是科学家们需要它。

We are dedicated to this, and the reason for that is that scientists need it.

研究人员需要它,创业公司需要它,企业也需要它。

Researchers need it, startups need it, and companies need it.

问题是,中国的模型够不够好?

The question is, are the Chinese models good enough?

当然,很多创业公司和小型企业已经将他们的工作负载转移到了开源模型上。

Certainly a ton of startups and smaller companies have already shifted their workloads over to open models.

X平台上的Signal写道,很多人都没有意识到有多少新东西是建立在中国开源权重模型之上的。

Signal on X writes, tons of people do not realize how much new stuff is being built on top of Chinese open-weight models.

你甚至不需要透露这一点,因为它运行在美国的基础设施上。

You don't even have to disclose it because it's running on US-based infrastructure.

当产品设计师Sam Solomon跟进说,我一直听到这个说法,但老实说,这实际上看起来并不比使用Luna便宜,Signal指出,你无法对Luna进行微调。

When product designer Sam Solomon followed up and said, I keep hearing this, but honestly it doesn't actually seem cheaper than using Luna, Signal pointed out, you cannot fine-tune Luna.

而且,微调越来越成为关键。

And increasingly, that fine-tuning is the story.

今年,关于开源模型的讨论已经从两个方面转向了:一方面是大型基础设施投资的市场影响,另一方面是与中国的总体地缘政治竞争,现在还包括了一个强有力的维度,即关于开源模型如何可能解决企业AI问题的讨论。

This year, the conversation around open models has shifted from being about either A, the market implications of big infrastructure investments, or B, the general geopolitical competition with China, to instead also including this strong dimension of discussion around how open models potentially solve enterprise AI problems.

投资人Joe McCann写道,这正是我投资Reflection AI的原因。

Investor Joe McCann wrote, This is precisely why I invested in Reflection AI.

开源是企业与政府最终信任AI的方式。

Open source is how enterprises and the government will ultimately trust AI.

关于企业对AI信任的讨论越来越热烈。

The discourse about enterprise trust of AI is getting louder.

早在七月份,在与CNBC的对话中,Palantir的CEO Alex Karp讨论了客户真正想要什么,他认为Frontier Labs的真正业务是什么,以及因此开源模型的重要性。

Back in July, in a conversation with CNBC, Palantir CEO Alex Karp discussed what customers actually want, what he considers the real business of Frontier Labs, and consequently the importance of open models.

Karp说,技术客户想要的是对他们计算资源、模型、数据栈和他们的优势(alpha)的控制权。

Said Karp, what the technical customers want is control over their compute, their models, their data stack, and their alpha.

他们想知道他们拥有生产资料,而这些生产资料没有被转移给其他人。

They want to know that they own the means of production, and it's not being transferred to someone else.

谁拥有数据?

Who owns the data?

提示词(prompts)安全吗?

Are the prompts secure?

这些信息是否被转移给了你?

Is this being transferred to you?

如果它这么有价值,我能为你赚十亿美元,那我难道不会说,我为你赚十亿美元,但我要30%吗?

If it was so valuable and I can make you a billion dollars, wouldn't I say, I'll make you a billion dollars and I want 30%?

既然它这么有价值,为什么他们还要按token收费呢?

Why are they charging for tokens if it's so valuable?

现在,Karp有点另类,但他不是唯一提出这个观点的人。

Now, Karp is a bit of a renegade, but he's not the only one making this argument.

越来越多地,微软也在向市场传递这样的主张。

Increasingly, this is the pitch that Microsoft is bringing to the market as well.

微软CEO Satya Nadella几个月前写了一篇博客,叫做《反向信息悖论》。

Microsoft CEO Satya Nadella wrote a blog post a couple of months ago called The Reverse Information Paradox.

他在文中说,在AI时代,买家为了使用他们购买的东西,有可能泄露知识。

In it, he said, in the AI age, the buyer risks giving away knowledge just in order to use what they bought.

你基本上为智能付了两次钱,一次是用钱,另一次是用更有价值的东西——为了让那个智能有用,你必须透露的专有知识。

You essentially pay for intelligence twice, once with money and again with something even more valuable, the proprietary knowledge you must reveal to make that intelligence useful.

你希望模型表现越好,就必须给它喂更多的知识。

The better you want the model to perform, the more of that knowledge you have to feed it.

随着时间的推移,信息不对称变得越来越严重。

Over time, the information asymmetry becomes increasingly skewed.

当你使用你购买的东西时,卖家越来越了解你,而你却几乎不了解卖家在反过来学习什么。

The seller learns more and more about you as you use what you purchased, while you learn very little about what the seller is learning in return.

在同一篇文章里,Satya 后来又说,企业需要一个真正的信任边界,这样他们的人力资本和代币资本才能不断累积。

In the same post, Satya later argued, enterprises need a real trust boundary for their human capital and token capital to compound.

在这个边界里,组织的数据、轨迹、评估、调整过的权重和记忆会一起积累并改善。

It's where an organization's data, traces, evals, adapted weights, and memory accumulate and improve together.

并且这是一个硬边界,未经同意,任何东西都不能穿过,包括智能的副产品。

And it is a hard boundary across which nothing crosses, not even the intelligence exhaust, without consent.

企业会要求拥有使用模型输出来微调或训练自己模型的权利。

Enterprises will demand the rights to use model outputs to fine-tune and/or train their own models.

我认为这是每个公司都有权让模型符合自身企业问责义务。

I think of this as every firm's right to align models to their enterprise accountability obligations.

有意思的是,微软当然不是在卖开源模型。

Now, what's interesting about this is that Microsoft is, of course, not selling open models.

但他们卖的是自己的模型,作为基础,让你可以在现有的可信合作伙伴那里做这种训练。

What they are selling, however, is their own models as a base to do this sort of training with an existing trusted partner.

微软 AI 首席执行官 Mustafa Suleyman 写道,这一切都是微软 Frontier Tuning 的基础。

Wrote Microsoft AI CEO Mustafa Suleyman, all of this is the foundation for Microsoft Frontier Tuning.

它让你可以定制我们的模型,创建只有你控制的、公司专属的智能体。

It lets you customize our models to create custom company-specific agents that only you control.

你可以把我们的模型变成你的模型,你的数据,你的智能体,你的护城河。

You can make our model your model, your data, your agents, your moat.

虽然这个宣传主要是围绕主权,但显然在效率提升的话题上也有好处。

And while the pitch was largely around sovereignty, there were also clearly benefits on this growing efficiency conversation as well.

比如 Suleyman 说,当我们为 McKinsey 的任务微调模型时,MAI,也就是微软的模型,取得了最高的胜率,在质量上超过了当时的顶尖模型 GPT-5.5,同时成本低了十倍。

For example, Suleyman said, when we tuned our models for McKinsey's tasks, MAI, which is Microsoft's models, delivered the highest win rate, outperforming GPT-5.5, at the time the state of the art, on quality while being 10x lower on cost.

这个趋势已经出现了一段时间,而且越来越明显。

Now this trend has been coming for a while and it's getting more and more serious.

无论你看哪里,都有更多迹象表明行为正在朝这个方向转变。

And everywhere you look, there are more indications that behavior is shifting this way as well.

比如几周前,Vercel 首席执行官 Guillermo Roche 写道,今天可能创下了 Vercel AI Gateway 上开源模型代币使用量的纪录。

A couple of weeks ago, for example, Vercel CEO Guillermo Roche wrote, looks like today may be a record day for token volume percentage of open models on Vercel AI Gateway.

开源模型使用了 78.4% 的代币,闭源模型只用了 21.6%。

Open models used 78.4% of tokens while closed used 21.6%.

而且在我们 AIDB 和 Superintelligent 的讨论中,关于开源模型的话题,去年这个时候可能连议程都上不了,现在很多企业都在认真对待了。

And certainly in the discussions that we're having both at AIDB and at Superintelligent, a conversation about open models, which might have not even made the agenda last year at this time, is now something that many enterprises are taking much more seriously.

不过,美国开源 AI 复兴的一个大问题是,这类 AI 的监管环境最终会是什么样。

One big question of any potential U.S. open-weight AI renaissance, however, is what the regulatory environment for this type of AI might end up being.

在这方面,特朗普政府任命了新的 AI 主管,并宣布成立一个新的 AI 工作组。

And on that front, the Trump administration has named their new AI czar and announced a new AI task force alongside him.

《华尔街日报》报道说,现任国家情报总监 Jay Clayton 将被任命为新的 AI 主管。

The Wall Street Journal reports that current Director of National Intelligence Jay Clayton will be empowered as the new AI czar.

Clayton 在两届特朗普政府中都是高级官员。

Clayton has been a senior official across both Trump administrations.

他在特朗普第一届政府中担任 SEC 主席,在本届政府中,他先担任纽约南区联邦检察官,后被任命为国家情报总监。

He served as SEC chair during Trump 1, and during the current administration, he began as a U.S. attorney in the Southern District of New York before being appointed Director of National Intelligence.

除了 Clayton,政府还将设立一个跨部门的 AI 工作组,成员包括研究工程副部长 Emile Michael、人事管理办公室主任 Scott Cooper 和联邦贸易委员会主席 Andrew Ferguson。

Alongside Clayton, the administration will now have a cross-department AI task force as well, with that group including Emile Michael, Undersecretary of War for Research and Engineering, Scott Cooper, the Director of the Office of Personnel Management, and Andrew Ferguson, the Chairman of the Federal Trade Commission.

联邦贸易委员会似乎越来越成为 AI 行业的重要监管机构,不过工作组中还有几位负责政府内 AI 实施的官员。

The FTC seems to be becoming an increasingly important regulator for the AI industry, although the task force also includes several officials responsible for AI implementation across the government.

除了这个核心小组,《华尔街日报》报道说,布什时代国家安全顾问 Condoleezza Rice、副总统 J.D. Vance、财政部长 Scott Bessent 以及前 AI 主管 David Sacks 也将加入工作组。

Alongside this core group, the Wall Street Journal reported that Bush-era National Security Advisor Condoleezza Rice, Vice President J.D.

万斯、财政部长斯科特·贝森特和前AI主管戴维·萨克斯也将被纳入该工作组。

Vance, Treasury Secretary Scott Bessent, and former AI czar David Sacks will also be included on the task force.

总统在 Truth Social 上确认了工作组,并写道:超级智能力量的任务是协调联邦政府的努力,确保美国在超级智能领域继续领先世界。

Confirming the task force in a Truth Social post, the president wrote, the Superintelligence Force is tasked with coordinating the effort of the federal government to ensure that America continues to lead the world in superintelligence.

这个超级智能力量将协调联邦政府与消费者、公共利益团体、宗教组织、关键基础设施提供商以及超级智能公司的互动。

This superintelligence force will coordinate the federal government's engagement with consumers, public interest groups, religious organizations, critical infrastructure providers, and superintelligence companies.

目前非常清楚的是,Clayton 把自己在人工智能或超级智能方面的角色首先定位为国家安全角色。

Now, what's super clear so far is that Clayton views his role around artificial and/or superintelligence as first and foremost a national security role.

Clayton 说,不是第一的风险很高。

Clayton said, the risk of not being first is high.

不作为先行者会增加已知和未知的风险,尤其是来自对手的风险。

Not being first increases the identified and unidentified risks, particularly from our adversaries.

成为先行者能让我们更好地代表美国人民应对这些风险。

Being first will better enable us to address those risks on behalf of the American people.

甚至在这次任命之前,Clayton就在接受CNBC采访时反对美国公司因同样的国家安全原因暂停AI模型开发。

Even before this appointment, Clayton argued in an interview with CNBC against U.S. companies pausing development of AI models for these same national security sorts of reasons.

Steve Bannon对此不满,指责Clayton基本上与中国走得太近。

Steve Bannon doesn't like it, accusing Clayton basically of being much too close to China.

在他的播客中,Bannon说:'Clayton是个很好的律师,但我不明白一个帮阿里巴巴完成IPO、为中共军方信息网络服务的总法律顾问,怎么会当上国家情报总监。'

On his podcast, Bannon said, Clayton's a great lawyer, but I don't know how an Alibaba general counsel who got them through an IPO, a military info network for the CCP, ends up Director of National Intelligence.

在批评光谱的另一端,Tornado Cash开发者Roman Storm根据他与Clayton的个人经历,对此事对开源可能意味着什么持相当负面的看法。

On the other end of the critique spectrum, Tornado Cash developer Roman Storm is pretty negative on what this likely means for open source in his estimation based on his personal experience with Clayton.

他写道:'看起来我们并没有走向一个支持开源AI的未来。'

He wrote, it doesn't look like we're headed towards a future that supports open source AI.

Jay Clayton曾担任纽约南区联邦检察官办公室的主管,那个办公室起诉了我。

Jay Clayton headed the U.S. Attorney's Office for the Southern District of New York, which prosecuted me.

我的直觉是,我们现在关于开放权重模型的两个非常不同的讨论——或者老实说可能是三个——即开放模型对企业的影响、对AI安全的影响、以及对国家安全的影响,鉴于我们即将进入新的监管阶段,这些讨论将不可避免地以某种形式更加接近。

My instinct is that the 2 very different conversations we're having around OpenWeights models— or maybe honestly even 3— Open models in terms of their implications for the enterprise, open models in terms of their implications for AI safety, and open models for their implications in terms of national security are inevitably going to get a bit closer in one form or another, given the new regulatory phase we're getting into.

不过,你可以说我是乐观主义者,但我认为有理由对美国开源模型的未来感到兴奋,哪怕仅仅是因为我们往往非常积极地跟随市场机会信号。

Still, call me an optimist, but I think there are reasons to be excited about the future of American open models, if for no other reason than we tend to pretty aggressively follow market opportunity signals.

在一个企业都在应对成本、数据隐私和数据主权问题的世界里,开放模型变得有趣得多。

And in a world where enterprises are looking to deal with costs and issues of data privacy and data sovereignty, Open models get a lot more interesting.

这绝对是一个我们会持续关注的讨论。

Certainly a conversation that we will continue to watch.

不过,今天《AI Daily Brief》的内容就到这里。

For now though, that is going to do it for today's AI Daily Brief.

一如既往,感谢您的收听或观看。

Appreciate you listening or watching as always.

下次再见,祝平安。

Until next time, peace.
已剔除 2 处广告(点击展开查看)
M1
M10:38广告 · 已剔除

The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

All right, friends, quick announcements before we dive in.

First of all, thank you to today's sponsors, KPMG, Section, Robots and Pencils, and Blitzy.

To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts.

And if you want to learn more about sponsoring the show, send us a note at [email protected].

M1
M110:59广告 · 已剔除

If you're the one responsible for AI adoption at your company, you need Section.

Section is a platform that helps you manage AI transformation across your entire organization.

It coaches employees on real use cases, tracks who's using AI for business impact, and shows you exactly where AI is and isn't creating value.

The result?

You go from rolling out tools to driving measurable AI value.

Your employees move from meeting summaries to solving actual business problems.

And you can prove the ROI.

Stop guessing if your AI investment is working.

Check out Section at sectionai.com.

That's S-E-C-T-I-O-N-A-I dot com.

The best teams don't have a single star carrying everyone else.

They know their own strengths and each other's weaknesses and play to both.

That's the team Robots and Pencils has built on purpose.

Nobody there is grinding through busywork to pad a headcount number.

People come for the hard problems, and they stay because everyone around them is leveling up at the same time.

In a market full of companies that are just trying to hire fast, that's worth a look.

Check out robotsandpencils.com/careers.

Every AI coding tool on the market does the same thing first: it starts writing code.

Blitzee does the opposite.

Before writing a single line, Blitzee spends days reverse engineering your entire codebase.

Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade.

The result is a dynamic knowledge graph that understands your software the way a principal engineer would after 30 years in the building.

Other tools guess at context with grep searches and markdown files.

Blitzy never guesses.

It builds true understanding first, then delivers over 80% of entire software epics autonomously.

Validated, end-to-end tested, production-grade pull requests.

That's why Fortune 500 engineering teams trust Blitzy with the codebases that matter most.

See for yourself at blitzy.com.

That's blitzy.com.