为何新一代AI“判断模型”可能带来重大商业影响

不是每天都能遇到一个新模型来玩，更不是每天都能遇到一种全新的模型构建方法，而且这种方法对我们如何使用模型有着相当不同的影响。

It's not every day that we get a new model to play around with, and it's certainly not every day that we get an entirely new approach to model building with some fairly different implications for how we even use it.

不过，今天我们讨论的是一类新模型，你可以称之为AI判断模型。

Today, though, we are talking about a new class of models which you might refer to as AI Judgment Models.

这些判断模型，比如我们今天要讨论的来自TypeSafe的Jev，不是生成大段文本，而是针对特定问题给出概率。

Rather than producing long strings of text, these judgment models, like the one we're discussing today, Jev, from TypeSafe, produce probabilities around specific questions.

这个客户是不是生气了？

Is this customer angry?

这封邮件里有没有新的依赖关系？

Is there a new dependency in this email?

我们需不需要因为这个改变运营计划？

Do we need to change the operational plan because of this?

今天我们来探讨这些模型背后的理念，它们是如何以不同的方式训练的，它们如何能更快、更便宜地做出这些判断，以及最重要的是，它们在你的整体模型栈中会扮演什么角色。

Today we're exploring the idea behind these models, how they're trained differently, how they can produce these judgments much more quickly and much less expensively, and most importantly, where they're going to fit in your overall model stack.

《AI每日简报》是一个每日更新的播客和视频，内容涵盖人工智能领域最重要的新闻和讨论。

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

好的，朋友们，在我们开始之前，先快速说几个通知。

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

首先，感谢今天的赞助商：KPMG、Blitzy、Section 和 Hyperagent。

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

想要收听无广告版本，请访问 patreon.com/aidailybrief，或者你也可以在 Apple Podcasts 上订阅。

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

订阅每月只需 3 美元，即可享受无广告体验。

Subscriptions are just $3 a month for ad-free.

如果你想了解如何赞助这个节目，请发送邮件至 sponsors@aidailybrief.ai，或者直接访问 aidailybrief.ai，那里有详细信息。

And if you want to learn about sponsoring the show, send us a note at sponsors@aidailybrief.ai, or just go to aidailybrief.ai, where you can learn all about it.

AI 安全相关的讨论继续在科技行业和主流社会中传播。

The AI safety discourse continues to trickle out through the tech industry as well as mainstream society.

不过，除非发生什么惊天动地的大事，否则我们会把这个话题放到头条新闻里，而不是作为主要节目内容。

But for now, unless something absolutely seismic happens, we're going to move it into the headlines and away from the main episode.

考虑到这一点，马克·扎克伯格在周末保持沉默后，终于就 AI 放缓这一想法发表了自己的看法。

With that in mind, after staying quiet over the weekend, Mark Zuckerberg has made his thoughts known on this idea of an AI slowdown.

周二，扎克伯格在X上发帖说：“每个实验室都有责任和动力，以安全训练模型所需的速度前进，并且有能力采取自己的行动来确保这一点。”

On Tuesday, Zuckerberg wrote in a post on X, "Every lab has the responsibility and incentive to move at the pace required to train its models safely and the ability to take its own actions to ensure that happens."

基本上，他的观点是，节奏是各个实验室自己的责任，而不是集体行动，这个观点基于两个核心想法。

Basically, his view is that pacing is the responsibility of individual labs rather than a collective action, and that view hinges on two core ideas.

第一，他说：“人们不想使用与自己不一致、不按自己要求行事的代理，所以实验室有很强的自然动力让模型更加对齐。”

First, that, quote, "People don't want to use agents that are misaligned with them and that don't do what they ask, so labs have a strong natural incentive to make their models more aligned."

第二，“如果模型造成伤害，实验室会面临重大责任，所以他们也有很强的动力来防止这种情况。”

And two, "Labs face significant liability if their models cause harm, so they have a strong incentive to prevent this as well."

为了强调这一点，扎克伯格说Meta已经将Muse的发布推迟了几个月，以进行安全工作。

Emphasizing the point, Zuckerberg said that Meta had delayed the release of Muse by several months to work on safety.

他接着说：“我们没有要求其他人都这样做才行动。”

He continued, "We didn't call for everyone else to do this before we would.

“我们只是把它作为日常工作的一部分，因为这对人们和我们来说显然是正确的事情。”

We just did it as part of our day-to-day work because it was clearly the right thing for people and for us."

基本上，扎克伯格是在说，现有的个人激励和后果已经足够迫使AI实验室进行对齐工作，并正确把握前沿节奏，而不需要外部政府强制放缓。

Essentially, Zuckerberg is saying that the individual incentives and consequences that are already in place are enough to force AI labs to work on alignment and to pace the frontier correctly, rather than needing some exogenous government-enforced slowdown.

不过，扎克伯格确实支持独立评估员和顾问的想法，认为这是最佳实践，而不是监管要求。

Now, Zuckerberg did support the idea of independent evaluators and advisors as a matter of best practice rather than regulation.

他声称Meta已经聘请了外部评估员，不是因为被要求这么做，而是因为这有助于产出更好的成果。

He claimed that Meta has already engaged outside evaluators, not because it's required of them, but because it helps produce better work.

最后，他总结道：“把绝大部分算力用于服务人类，而不是竞相追求递归自我改进，是确保我们安全开发这项技术的最佳方式之一。”

Finally, he concluded, "Committing the significant majority of compute towards serving people rather than racing towards recursive self-improvement is one of the best ways to ensure that we develop this technology safely.

Meta已经做出了这一承诺，其他实验室也可以这样做。

Meta has made this commitment, and other labs can do this as well.

我相信，为所有人创造积极未来的关键在于保持权力的适当平衡。

I believe the key to building a positive future for everyone is maintaining the right balance of power.

这是我们力所能及的。”

This is within our power to do."

这话说得挺委婉，但基本上整段话就是在说：“拜托，各位。”

It's carefully worded, but basically, this whole thing says, "Come on, guys.

咱们别搞那些虚头巴脑的表演了。”

Let's please stop with the theatrics."

说实话，很多人都觉得这话让人耳目一新。

And a lot of people, frankly, found this a breath of fresh air.

YouTuber约瑟夫·卡尔森写道：“等一下。

YouTuber Joseph Carlson wrote, "Hold up a minute.

你是说公司可以放慢脚步，确保安全，而不必要求所有竞争对手都放慢脚步？”

You are telling me companies can slow down, make sure things are safe without telling all their competitors to slow down?"

我觉得总的来说，大家的反应大致分成两类。

And I would say that broadly speaking, reactions fell into one of two categories.

第一类就像刚才那个，还有马修·伯曼的这个：“我喜欢这个。

The first was like that one and like this from Matthew Berman, "Love this.

模型安全和对齐是一个特性，而且在经济上是有激励的。”

Model safety and alignment is a feature and economically incentivized."

或者比尔·阿克曼，他干脆称这是AI开发的正确做法。

Or Bill Ackman, who simply called this the proper approach to AI development.

但另一方面，也有人认为，不管这个论点本身有没有道理，扎克伯格根本没有足够的信誉或立场来提出这个论点。

But then on the other side were those arguing effectively that Zuckerberg just does not have the trust or standing to make this argument, regardless of the merits of the argument itself.

《Hard Fork》的凯文·鲁斯写道：“无论你是民主党人还是共和党人，是有效利他主义者还是有效加速主义者，我想我们都同意，最适合保护我们免受强大新技术危害的人，就是马克·扎克伯格。”

Hard Fork's Kevin Roose wrote, "Whether you're a Democrat or a Republican, an EA or an EACC, I think we can all agree that the person best suited to protect us against the harms of powerful new technology is Mark Zuckerberg."

与此同时，在《奇怪的同床异梦日报》那边，伯尼·桑德斯和史蒂夫·班农联手呼吁制定以人为本的AI法规，这成了本政治周期最奇怪的联盟。

Now, meanwhile, over in Strange Bedfellows Daily, Bernie Sanders and Steve Bannon have joined forces to call for human-centric AI regulations in the strangest alliance of this political cycle.

这两位高度意识形态化的领导人，周二在华盛顿未来生命研究所的“亲人类大会”上，同台发表了讲话。

The two highly ideological leaders spoke from the same stage on Tuesday at the Future of Life Institute's Pro-Human Assembly in Washington.

桑德斯对人群说：“如果这项技术将从根本上改变每个男人、女人和儿童的生活，那么这个国家的人民必须对AI做出决定，而不是仅仅由少数寡头来决定。”

Sanders told the crowd, "If the lives of every man, woman, and child are going to be fundamentally changed by this technology, then the people of this country must make decisions about AI and not just a handful of oligarchs."

班农的言论也非常相似，他说：“美国公民将不再向寡头们俯首称臣。”

Bannon had very similar remarks, stating, "The American citizens are not going to be supplicants to the oligarchs anymore.

我们不能这样。

We can't do it.

这是历史的转折点。

This is a hinge in history.

我们必须正确处理这件事。

We have to handle this correctly.

要正确处理这个问题，第一，我们永远不能相信寡头说的话。

To handle it correctly, number one, we can never trust what an oligarch says."

现在我要指出，虽然一开始看起来很奇怪，这两个意识形态截然对立的人站在同一个舞台上，但他们所代表的更广泛的运动其实有不少共同的历史背景和交集。

Now, I will note that while this seems strange at first, these two highly ideologically opposed people sharing the same stage, the broader movements they connect to have a fair bit of shared context and history.

2016年总统大选，伯尼以微弱劣势输给希拉里，未能成为民主党候选人，而班农显然策划了特朗普的第一届政府，这两件事都源于对后金融危机时代金融格局的民粹主义愤怒，以及对那些参与制造这场经济危机的机构和机构领导缺乏问责的不满。

The 2016 presidential election, in which Bernie was narrowly beaten by Hillary to miss out on becoming the Democrat nominee, and in which Bannon obviously architected the first Trump administration, both had their roots in populist anger at the post-GFC financial landscape and the lack of accountability for the institutions and institutional leaders who were involved in creating that particular economic crisis.

显然，左派和右派对这一特定背景的反应不同，但我认为，这两个人最终找到共同点，可能并不像你想象的那么令人惊讶。

Obviously, the left and the right's reaction to that particular context were different, but it is, I would contend, perhaps less surprising than you think, that at some point, these two would find common ground.

需要说明的是，我并不是仅仅因为这个特定问题本身的价值以及它打乱现有政治联盟的力量，就否定他们找到的共同点。

And to be clear, I am not dismissing the common ground that they find just on the merits of this particular issue itself and the power of this particular issue to scramble existing political alliances.

我只是想指出，他们的故事实际上比乍看起来更加交织在一起。

I'm just making the point that their stories are actually more intertwined than you might think at first glance.

无论如何，这次活动似乎更多是关于技术所代表的阶级斗争，而不是本周占据头条的AI存在风险。

In any case, the event seemed somewhat less about the existential risks of AI that have filled the headlines this week and more about the class struggle that technology has come to represent.

电视屏幕上播放了一段对虚构AI CEO的恶搞采访，描述这个人热爱人类，但对人类本身并不狂热。

TV screens played a parody interview from a fictional AI CEO described as someone who loves people but isn't crazy about humans.

整个活动下来，感觉AI本身并不是需要解决的风险，真正的问题在于寡头们不受约束的权力。

Throughout the event, it seemed that AI itself wasn't the risk that needs addressing, but rather the unchecked power of oligarchs.

你可能还记得，八月份的时候，记者Jasmine Sun走访全国，跟那些反对数据中心的人聊过，发现大家最常抱怨的不是电力、用水或者噪音污染，而是缺乏控制感，觉得未来是被强加给他们的社区，他们几乎没有发言权。

Now, you might remember that back in August, journalist Jasmine Sun toured the country speaking to real people who were working in opposition to data centers and found that the most common complaint was not about electricity, water use, or noise pollution, but a lack of control and the sense that the future was being forced upon their communities with little input.

这次活动也是一样。

The same was true about this event.

多位演讲者都没有把AI的风险说成是网络安全、生物恐怖主义或者其他什么X风险。

Across multiple speakers, the risk of AI was not framed around cybersecurity, bioterrorism, or other X risks.

他们说的是，在决定未来形态这件事上，人们缺乏自主权。

It was about a lack of agency in determining the shape of the future.

由此引出了他们最关心的问题，其实就是担心失去对AI的控制。

From this followed their main concern, which was simply the idea of losing control of AI.

正如德克萨斯州民主党人Greg Casar，也就是伯尼·桑德斯超级智能法案的发起人所说："答案很简单。

As Texas Democrat Greg Casar, who is sponsoring Bernie Sanders' super intelligence bill said, "The answer is simple.

我们要禁止那些强大到人类无法控制的AI系统。"

We ban AI systems that are too powerful for humans to control."

尽管最近这些言论听起来很激烈，但我觉得我们已经进入了一个新阶段，这个阶段主要是关于协商公民和政府在其中都有利益的关系，而且既然现在大家基本上都处于这个阶段，接下来的阶段可能会包含更多的具体细节，甚至可以说是更多的微妙之处。

And as strange as it might seem, despite all the intensity of this rhetoric recently, I still contend that we've moved into a new phase that's all about negotiating the relationship where citizens and governments have a stake in this, and where given that that is now pretty much where everyone is, the next phase is likely to include a lot more specificity and, dare I say, nuance.

比如，格伦·贝克在他的节目里有一段很长的独白，讲他既签署了《支持人类AI宣言》，又在数据中心的具体问题上和他们意见不同。

Glenn Beck, for example, had a long monologue on his show about how he can both have signed the Pro-Human AI declaration alongside them, but also disagree with them on specifics of data centers.

说实话，虽然听起来很疯狂，但我觉得政治讨论越是让人困惑、让人摸不着头脑，这反而可能说明我们真的在取得进展。

And honestly, as crazy as it sounds, I think that the more that the political discourse kind of frags your brain for how confusing and all over the place it is, that might just be a sign that we're actually making progress.

这个话题也出现在了Salesforce的年度Dreamforce大会上，他们发布了新的AI模型，但社交媒体上的很多讨论都集中在山姆·奥特曼、黄仁勋和达里奥·阿莫迪的出席上，他们各自阐述了自己的安全观点。

The conversation also found its way into Salesforce's annual Dreamforce event, where the company unveiled a new AI model, but where a lot of the chatter on social media focused on appearances from Sam Altman, Jensen Huang, and Dario Amodei, who each iterated their own safety views.

达里奥说，他只是想提出一套行业可以围绕其组织的标准，而奥特曼则表示，他对我们公司和行业安全做到这一点的能力非常有信心。

Dario argued that he was just trying to put forward a set of standards that the industry could organize around, while Altman said that he was very confident in our company's ability and our industry's ability to do this safely.

与此同时，黄仁勋重申了扎克伯格式的观点，说：“尽可能快地跑，但如果你在任何时候觉得公司失控或者产品不安全，就停下来，确保你做对了。”

Jensen Huang, meanwhile, reiterated the Zuckerberg-ian view, saying, "Run as fast as you can, but if you feel at any given point in time the company's out of control or the product's not going to be safe, take a pause and make sure you get it right."

至于Salesforce的首席执行官马克·贝尼奥夫，他认为每家公司都有责任维护道德标准。

And as for Salesforce CEO Mark Benioff, he believes that every company has a responsibility to uphold ethical standards.

不过，为了暂时缓解一下安全辩论的紧张气氛，Salesforce还发布了两项重要的实际AI公告。

Still, to give the safety debate a bit of a rest, Salesforce also had two big practical AI announcements.

首先，他们发布了很长时间以来的第一个自研模型。

First, they're releasing their first in-house model in quite some time.

这个模型叫Qoa，是基于英伟达的Nemotron微调而来的，专门用于CRM中的销售管理。

Called Qoa, the model is a fine tune of NVIDIA's Nemotron and is designed to handle sales management within the CRM.

这一发布进一步凸显了开源在企业领域的作用，因为它让这种窄垂直模型成为可能。

The announcement reinforces the role that open source has to play in the enterprise by enabling this sort of narrow vertical model.

其次，Salesforce推出了一个名为AI Force的新计划。

Secondly, Salesforce unveiled a new initiative called AI Force.

这将是Salesforce连接器的总称，这些连接器允许第三方代理访问Salesforce的数据。

This will be the umbrella term for Salesforce's connectors that will allow third-party agents to access Salesforce data.

这次发布再次表明Salesforce致力于以平台无关的方式向无头软件转型，让任何代理都能成为界面。

The release reinforces Salesforce's commitment to moving towards headless software in a platform-agnostic way, allowing any agent to become the interface.

显然，这些都是重要的大话题，而且还会继续讨论下去，但今天的头条新闻就到这里。

Now, obviously, these are big conversations that are important and are going to continue, but for now, that is where we will close the headlines.

接下来，是本期主节目。

Next up, the main episode.

毕马威和德克萨斯大学奥斯汀分校的一项新研究发现，当人们与人工智能合作时，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多名早期职业专业人士，发现表现最好的人通过引导、评估和完善人工智能的输出，持续放大了人工智能的价值。

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“人工智能放大器”，他们的定义不仅仅在于他们知道什么，而在于他们如何与人工智能合作。

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

想了解更多关于人工智能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.

这就是为什么大多数传统系统现代化项目会失败的原因。

Here's why most legacy modernization projects fail.

做这项工作的人工智能无法大规模理解代码库。

The AI doing the work can't understand code bases at scale.

它只看到一小部分上下文，检查语法，却错过了分布在全球应用生态系统中的多年决策。

It sees a small slice of context, examines syntax, and misses years of decisions distributed across the global application ecosystem.

Blitzy解决这个问题的方式，就像它解决其他一切问题一样。

Blitzy solves this the way it solves everything.

在迁移开始之前，Blitzy的智能代理会基于你的代码，把整个遗留系统逆向工程成一个持久的知识图谱。

Grounded in your code before any migration begins, Blitzy's agents reverse engineer the entire legacy system into a persistent knowledge graph.

每一个依赖、每一个约束、每一条曾经只存在于某个工程师脑子里的隐性知识。

Every dependency, every constraint, every piece of tribal knowledge that used to live in one engineer's head.

基于这些理解，Blitzy能自主执行语言迁移、框架升级，以及从单体架构到微服务的改造，并且全程端到端验证。

From that understanding, Blitzy autonomously executes language migrations, framework upgrades, and monolith to microservices transformations, all validated end to end.

Blitzy的一个客户用编码代理，在16周内完成了价值1000万美元的单体保险系统的现代化改造，而传统基线是137周。

One Blitzy customer modernized a $10 million monolithic insurance stack in 16 weeks against a 137-week baseline with coding agents.

这相当于压缩了9倍的时间。

That's 9x compression.

在加速你的路线图的同时，消除技术债务。

Retire technical debt while accelerating your roadmap.

想了解更多，请访问blitzy.com。

See how at blitzy.com.

网址是B-L-I-T-Z-Y点com。

That's B-L-I-T-Z-Y dot com.

这里有个残酷的事实。

Here's a harsh truth.

你的公司可能花了成千上万甚至数百万美元在AI工具上，但这些工具被严重浪费了。

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%的公司用它们来创造商业价值。

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

大多数员工还在用AI来总结会议纪要。

Most employees are still using AI to summarize meeting notes.

如果你负责公司AI的落地，你需要Section。

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

Section是一个平台，帮你管理整个组织的AI转型。

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

它指导员工实际应用场景，追踪谁在用AI创造商业影响，并让你清楚看到AI在哪里创造价值，哪里没有。

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.

结果就是，你从推出工具转变为推动可衡量的AI价值。

The result, you go from rolling out tools to driving measurable AI value.

你的员工从整理会议纪要转向解决实际业务问题，你也能证明投资回报率。

Your employees move from meeting summaries to solving actual business problems, and you can prove the ROI.

别再猜测你的AI投资是否有效了。

Stop guessing if your AI investment is working.

去sectionai.com看看Section吧。

Check out Section at sectionai.com.

就是S-E-C-T-I-O-N-A-I点com。

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

本期《AI每日简报》由Hyperagent赞助，你可以通过它管理团队协作的智能体集群。

This episode of the AI Daily Brief is brought to you by Hyperagent, where you run fleets of agents your team can manage together.

别再依赖本地智能体和等待你笔记本电脑提示的聊天工作流了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent在云端部署常驻智能体，在你团队已经使用的工具中完成实际工作。

Hyperagent deploys always-on agents in the cloud, doing real work across the tools your team already uses.

营销智能体把竞争对手的动态转化为落地页。

Marketing agents turn competitor moves into landing pages.

销售代理会丰富潜在客户信息、起草邮件，并更新客户关系管理系统。

Sales agents enrich leads, draft emails, and updates the CRM.

运营代理负责处理文书工作，并跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个代理都能访问共享上下文，并遵循你关于范围和审批的规定。

Every agent has access to shared context and follows your rules about scope and approvals.

是时候让代理们像队友一样工作了。

It's time you add agents that feel like teammates.

去Hyperagent雇佣你的代理吧。

Hire yours at Hyperagent.

在hyperagent.com/aidailybrief注册，即可获得100美元的信用额度。

Get $100 in credits at hyperagent.com/aidailybrief.

欢迎回到AI每日简报。

Welcome back to the AI Daily Brief.

在今天的正片中，我们要探讨一个非常不同且相当罕见的东西，简单来说，就是一种完全不同的AI方法，而不仅仅是另一个大语言模型。

On today's main episode, we are looking at something really different and quite rare, which is, in short, a totally different approach to AI that is not just another LLM.

现在，距离ChatGPT发布已经将近四年，而这个节目也播出了三年半，我们和整个行业关注的大部分内容，在某种程度上都与大语言模型有关。

Now, in the nearly four years since ChatGPT was released and the three and a half that the show has been around, the vast majority of the things that we and the rest of the industry have focused on have been in some ways related to large language models.

然而，正如一些人指出的那样，尽管这种观点往往没有引起太多关注，更像是对着虚空呐喊，但大语言模型实际上并不是人工智能的全部。

And yet, as some have pointed out, although often in a way that didn't get much traction and was more or less screaming into the void, LLMs are not, in fact, the totality of artificial intelligence.

昨天，Diogo Almeida在X上发帖说：“在共同发明ChatGPT之后，我一直在问自己，‘为什么超人类的聊天模型没有带来AGI？’”

Yesterday, Diogo Almeida posted on X, "After co-inventing ChatGPT, I kept asking myself, 'Why have superhuman chat models not led to AGI?'

过去两年，我一直在秘密开发一种新的模型训练方法，RLCD，即用于校准决策的强化学习，以及我们今天发布的一种新型前沿AI模型，Jev。

I've spent the last two years in stealth building a new way to train models, RLCD, or Reinforcement Learning for Calibrated Decisions, and a new type of frontier AI model that we are releasing today, Jev.

Jev的速度快20到200倍，成本低40到400倍，输出令牌免费，是面向决策优化的前沿可组合智能。

Jev is 20 to 200 times faster, 40 to 400 times cheaper with output tokens free, frontier composable intelligence optimized for decisions.

据我所知，这是通往基于AI的经济革命的最短路径。”

As far as I can tell, the shortest path to AI-based economic revolution."

现在，看到20到200倍更快、40到400倍更便宜这样的数字，人们完全有理由持怀疑态度，至少可以说。

Now, one could absolutely be forgiven for seeing numbers like 20 to 200 times faster and 40 to 400 times cheaper and being a bit skeptical of the claims, to say the least.

如果这仅仅是另一个大语言模型，那么这种怀疑是完全合理的。

And were this just another LLM, that skepticism would be entirely warranted.

不过，对于Jev和这个新策略，我们面对的是完全不同的东西。

However, with Jev and this new strategy, we're dealing with something that is quite different.

Jev背后的公司叫Typesafe，在发布公告的博客文章中，他们更详细地介绍了自己方法的独特之处。

The company behind Jev is called Typesafe, and in the announcement blog post, they talk a little bit more about what makes their approach different.

他们写道：“我们构建了一个全新的技术栈，完全专注于自动化，采用了新的模型架构、并行采样器以实现最大效率，以及我们称之为‘校准决策强化学习’的训练方法。”

They write, "We built a new stack entirely focused on automation with a new model architecture, parallel sampler for maximum efficiency, and training method we call Reinforcement Learning for Calibrated Decisions.

现有的LLM优化的是人类偏好，也就是人类评分者更喜欢的写作和聊天回复，而新的System 1模型（第一个就是Jev）优化的是校准决策，即带有认知上诚实的概率的答案。”

Whereas existing LLMs optimize for human preference, i.e. write-ups and chat responses that human raters prefer, the new System 1 models, the first of which is Jev, optimize for calibrated decisions, or answers with epistemically honest probabilities."

那这到底是什么意思呢？

So what does that actually mean?

我们来看看Every的Mike Taylor是怎么描述的。

Well, let's look at how Mike Taylor from Every describes it.

他写道：“可以把它想象成一个智能的if/then语句，在自动化工作流程时决定接下来会发生什么。”

He writes, "Think of it as a smart if/then statement that determines what happens next when you're automating a workflow.

假设你在构建一个优先处理客户服务请求的软件，你写代码问模型：‘这个客户听起来生气吗？’

Say you're building software that prioritizes customer service requests, and you write code that asks the model, 'Does this customer sound angry?'

Jev可能会回答0.9，这意味着根据模型在训练中学到的内容，答案有大约90%的可能性是肯定的。

Jev might answer 0.9, which means there's an estimated 90% probability that the answer is yes based on what the model learned in training.

你还可以定义一些类别，比如恼火、烦躁、被冒犯、愤怒和暴怒，然后了解到客户有60%的可能性被归类为愤怒，而只有10%的可能性是暴怒。

You could also provide categories you define, like annoyed, irritated, offended, furious, and enraged, and learn that the customer was 60% likely to be classified as furious with only a 10% probability of being enraged."

接着解释为什么这很重要，以及它与大型语言模型有何不同，Mike继续说：“如果答案是0.9，很可能客户很生气，软件可能会自动将客户的问题升级给经理处理。

Going on to explain why this matters and how it differs than LLMs, Mike continues, "With an answer of 0.9, very likely to be angry, the software might automatically proceed to escalate the customer concern to a manager.

或者如果答案是0.1，不太可能生气，那么这个请求可能会被降低优先级。

Or if it answers 0.1, not likely to be angry, that request might be deprioritized.

然而，聊天机器人被训练成用华丽的语言回应，比如：‘您说得完全正确，这位客户听起来确实非常生气。

However, chatbots are trained to respond with flowery text like, 'You're absolutely right, this customer does sound very angry.

您需要我以友好、支持的语气起草一封电子邮件回复吗？’

Would you like me to compose a draft email response in a friendly, supportive tone?'

这种文本回应会导致你正在构建的程序崩溃，因为它期望的是0到1之间的数字，而不是一篇长篇大论。”

This text response would cause the program you're building to crash because it was expecting a number between zero and one, not an essay."

Teal的同事Michael Li说：“这开启了一类决策方式，既不适合笨拙、无智能的代码，也不适合缓慢且昂贵的LLM。”

Teal fellow Michael Li says, "This allows a class of decision-making that was neither suited to dumb, unintelligent code nor to slow, expensive LLMs."

正如Chubby总结的那样，“Jev是一个专为决策而非文本生成而构建的AI模型。”

As Chubby sums up, "Jev is an AI model built for decisions rather than text generation."

而且需要指出的是，这里的权衡在于这个模型不生成文本。

And it is important to note that the trade-off here is that this model does not generate text.

换句话说，它并不是LLM的通用替代品。

It is not, in other words, a replacement for LLMs in general.

它替代的是LLM所承担的某类工作，在这些工作中，LLM就像方钉被硬塞进圆孔一样，用得很别扭。

It's a replacement for a certain category of work that LLMs do, where they've been very square peg mashed into a round hole to do it.

Chubby继续说：“其理念是将快速、廉价的AI决策嵌入到软件中。”

"The idea," continues Chubby, "is to embed fast, cheap AI decisions into software."

那么这个模型到底是用来做什么的呢？

So what is this model actually built for?

想想看，办公室工作中有多少是阅读内容然后决定下一步该做什么。

Well, think of how much of office work consists of reading something and deciding what should happen next.

这条消息需要回复吗？

Does this message need a response?

应该由哪个部门来处理？

Which department should handle it?

这份文件回答这个问题了吗？

Does this document answer the question?

这个客户是在描述一个漏洞，还是在要求一个新功能？

Is this customer describing a bug or asking for a feature?

这份草稿是否提出了其来源不支持的主张？

Does this draft make a claim its source doesn't support?

这种情况是常规到可以自动化处理，还是需要有人来审核？

Is the situation routine enough to automate, or should someone review it?

这些都是关于意义的判断，而且往往很难用固定的规则来表达。

These are judgments about meaning, and they're often difficult to express as fixed rules.

因此，Jev 的设计目的是获取相关信息，并以概率、类别或分数的形式回答定义明确的问题。

Jev then is designed to take the relevant information and answer narrowly defined questions with probabilities, categories, or scores.

周围的软件再利用这些答案来进行路由、排序、标记或继续处理。

The surrounding software then uses those answers to route, rank, flag, or proceed.

在Typesafe的文档里，他们明确建议把复杂的决策拆成小问题，然后在代码里把结果组合起来。

In the Typesafe documentation, they explicitly recommend breaking complex decisions into small questions and then combining their results in code.

那这在日常业务中会体现在哪里呢？

So where would this show up in normal business?

一个明显的领域就是客户支持，模型能做的简单判断比如：客户是不是很沮丧，之前的回复是不是没解决问题？

One obvious area is customer support, where the small judgment the model could make would be something like, is the customer frustrated, and have previous replies failed to address it?

有了这些简单判断，软件就可以接着把工单转给相应部门、提高优先级，或者请求人工审核。

With those small judgments in hand, the software could then next route the ticket, raise its priority, or request human review.

在销售领域，模型可以判断：这是不是购买咨询？

In the sales domain, the model could judge, is this a buying inquiry?

这个潜在客户适合这个产品吗？

Does the prospect fit the product?

他们是不是在要求安排会议？

Are they requesting a meeting?

然后软件就可以根据这些判断来分类进来的线索，并安排后续跟进。

The software could then take those judgments to sort inbound leads and assign follow-up.

在市场营销和编辑领域，模型可能会判断文案是否符合特定的风格规则，或者报价是否清晰。

In marketing and editorial, the model might judge whether the copy meets specific style rules or whether the offer is clear.

周围的软件可以在发布前标记需要修改的段落。

The software that surrounds it could then flag passages for revision before publication.

重要的是，很多人关注的不仅仅是他们会在哪里用它来替代大语言模型，而是如何将它与大语言模型结合使用。

And importantly, where a lot of people went was not just understanding where they would use this instead of LLMs, but how they might use it alongside LLMs.

YC创始人Nathan Fleury写道：“我设想很多工作流程会是这样的：大语言模型提出选项，Jev做决定，代码执行。”

YC founder Nathan Fleury wrote, "I'd imagine a lot of workflows that look like LLM proposes options, Jev decides, code executes.

举个具体的例子，想象一下客户写道：“这是我第三次联系你们了。”

To put a clear example on this, imagine that a customer writes, 'This is the third time I've contacted you.

“我们还是无法导出报告，而且我们的续约就在下周。”

We still can't export our reports, and our renewal is next week.'

像Jev这样的支持工作流可以同时提出几个问题。

A support workflow supported by something like Jev could ask several questions together.

第一，客户是否在描述一个产品问题？

One, is the customer describing a product problem?

第二，这条消息是否表明客户多次联系但问题一直没解决？

Two, does the message indicate repeated unsuccessful support?

第三，是否有重要的商业截止日期临近？

Three, is a commercially significant deadline approaching?

第四，哪个团队最适合处理这个问题？

Four, which team is best equipped to help?

然后，围绕它的软件可以把这些信号和实际的账户信息结合起来，比如续约日期，然后升级工单。

The software that surrounds it could then combine those signals with actual account information, such as the renewal date, and escalate the ticket.

接下来，你仍然可以用生成式模型来起草回复，但Jev可以再次根据严格的标准来检查草稿。

From there, you would still have a generative model draft the reply, but Jev, once again, could check the draft against narrow criteria."

回复是否承认了客户多次联系？

Does the response acknowledge the repeated contacts?

是否解决了导出问题？

Does it address the export problem?

是否承诺了基于现有信息无法支持的内容？

Does it promise something unsupported by the information provided?

而且大家对这个新方法感到兴奋的部分原因是，便宜的判断让频繁检查变得更实际。

And part of what people are excited about opening up with this new approach is that cheap judgment makes frequent checking more practical.

如果一次检查会带来明显的延迟或成本，团队可能只在选定的案例上或任务结束时运行它。

If a check adds a noticeable delay or expense, a team may run it only on selected cases or at the end of a task.

如果检查变得足够快和便宜，它就可以在每一个进来的请求上运行，在每次草稿修改后，在代理采取重要步骤之前，跨多个候选文档运行。

If it becomes sufficiently fast and inexpensive, it could run on every incoming request, after each draft revision, across many candidate documents before an agent takes a consequential step.

回到迈克·泰勒在《Every》上的文章，他把所有27篇文章的文本给了Jev，同时附上10篇故意模仿AI风格的反驳文章，然后对所有文章同时提出同样的21个问题，以检查是否有AI痕迹。

Back in Mike Taylor's article from Every, he gave Jev the text from all 27 of his articles, alongside 10 deliberately AI-styled counterpoints, then asked the same 21 questions concurrently across all articles to check for AI tells.

基本上，他用Jev做的检查是：这篇文章是否做了某些特定的事情，让人们觉得它是AI写的？

Basically, the check that he was doing with Jev was, does this essay do specific things that indicate to people that it is AI-composed?

比如，文本是否在没有添加证据的情况下重复一个观点？

Things like, does the text repeat an idea without adding evidence?

它是否强行制造对称的“双方”论证？

Does it force a symmetrical both sides argument?

它是否过度解释一个简单的观点？

Does it overexplain a straightforward point?

不到0.7秒，迈克说，Jev，引号，反引号，"读完了全部37份文档，并为每份回答了21个问题，返回了777条判断，估计成本只有四分之一美分。"

In less than 0.7 seconds, Mike said, Jev quote, unquote, "read all 37 documents and answered all 21 questions for each, returning 777 judgments for an estimated quarter of a cent."

正如他指出的，这速度快、成本低，足以检查你公司里每个人写过的所有内容，而且瞬间就能拿到结果。

As he points out, that's fast and cheap enough to AI check everything everyone at your company has ever written and get the results back in an instant.

迈克做的类比是，这就像是知识工作的代码检查工具。

A comparison that Mike makes is a code linter for knowledge work.

他写道："在软件开发中，代码检查工具是一种分析你的工作并几乎立即标记语法错误、捕捉漏洞、发现不良模式并强制风格一致性的工具。

He writes, "In software development, a code linter is a tool that analyzes your work and almost instantly flags syntax errors, catches bugs, spots bad patterns, and enforces stylistic consistency.

TypeSafe的模型在将模糊任务转化为清晰、结构化的答案方面速度极快，因此它可以充当知识工作的代码检查工具。

TypeSafe's model is so fast at turning fuzzy tasks into clear, structured answers that it could act as a kind of code linter for knowledge work.

给Codex或Claude接入Jev和一系列问题，它就能快速检查自己的工作，看是否出现了你告诉它要避免的问题。"

Give Codex or Claude access to Jev and a list of questions, and it can quickly check its own work for problems you've told it to avoid."

Perangrat说："大多数软件最终都是一棵巨大的'如果这样，就那样做'的树。

Perangrat says, "Most software is ultimately a giant tree of if this, do that.

如果这样，就路由到这里。

If this, route here.

如果是这种情况，就升级处理。

If this, escalate.

如果是这种情况，就拒绝。

If this, reject.

如果是这种情况，就找人工处理。

If this, ask a human."

Jev 基本上是在问，如果那些条件判断能理解混乱的人类语境会怎样？

Jev is basically asking, what if those if statements could understand messy human context?

这个框架比另一个 AI 模型有意思多了。

That's a much more interesting framing than another AI model.

我觉得这在欺诈和风险控制、客服路由、内容审核、公关和质检自动化、销售线索评分、合规、工作流编排以及智能体路由方面会非常有用。

I can see this being very useful for fraud and risk, support routing, moderation, PR and QA automation, lead scoring, compliance, workflow orchestration, and agent routing.

他说：“显然，这是早期技术，但这个类别本身很有意义。”

"Early tech, obviously," he says, "but the category itself makes a lot of sense."

有趣的是，Matt Stockton 指出，从某种程度上说，公司采用这种技术相当于一种后 LLM 的 AI 技术，让前 LLM 时代的机器学习技术变得更容易上手。

Now interestingly, Matt Stockton points out that in some ways, companies adopting this amounts to a post-LLM AI technology, making pre-LLM machine learning techniques a little bit more accessible.

正如他所说，商业中的很多很多问题都是分类或回归问题。

As he writes, "Lots and lots of problems in business are classification or regression problems.

很多很多公司并不知道他们遇到的问题可以通过经典的机器学习方法来解决。

Lots and lots of companies don't know that the types of problems they have are solvable by classic machine learning methods.

他们通常用人和流程来解决这些问题，而不是用技术。

They often solve them with people and process instead of technology.

随着大语言模型的兴起和流行，越来越多的公司开始想，也许我们可以用人工智能来做这件事，于是用大语言模型来解决分类和回归问题。

With the emergence and popularity of LLMs, more companies are thinking, 'Maybe we can use AI for that,' and are solving classification and regression problems with LLMs.

这在某些方面是好的，因为公司可能自动化了一些手工工作，但在某些方面也是不好的，因为大语言模型往往不是解决问题的合适工具，可能不如经典的机器学习方法效果好。

This is good in some ways because companies are potentially automating some manual work, but also bad in some ways because it's often the wrong tool for the job and possibly not as good as classic ML methods for what they are trying to do."

但他指出，经典方法需要你标记数据、训练模型，然后把模型部署到某个地方。

But he points out, the classical techniques require you to label your data, train a model, and host that model somewhere.

它们不像调用大语言模型API那样容易使用。

They aren't as easy to use compared to calling an LLM API.

而且这需要你和你的组织了解这些技术，并且有能力投入其中。

And it requires you and your org to be aware of those techniques and capable of investing in them.

不过这个比喻先不展开太多，他基本上是说，可以把Jev看作是一种用户体验，用LLM式的交互模式来操作经典的机器学习技术。

Without going too far on the analogy, he basically says one way to look at Jev is as a UX for using LLM style user interaction patterns for classical ML techniques.

那么，一个有趣的问题是，这到底是给个人用的，还是给团队构建系统用的。

Now, one interesting question that comes up is whether this is for individuals or teams building systems.

简单来说，虽然两者都适用，但它也进一步突出了我们最近一直在讨论的多人在线AI主题。

And the short answer is that while it is absolutely both, it also puts a fine point on the multiplayer AI themes that we've been talking about recently.

当然，个人可以用这种能力来打造自己的工具，比如按照自己的优先级整理收件箱，按照编辑标准检查写作草稿，根据研究兴趣给保存的文章排序，或者在会议记录中标记承诺事项，诸如此类，这些都能帮你更好地完成工作。

Certainly, individuals could use this sort of capability to have personal tools that sort an inbox against their own priorities, check drafts of their writing against an editorial rubric, rank saved articles against research interests, or flag commitments in meeting transcripts, things like that, that are going to personally help you do your work better.

但我认为，这种技术真正大放异彩的地方，是在团队协作中那些关于谁需要知道、谁应该行动、谁需要批准的小判断。

But I think that where this sort of technique is going to really shine is in the domain of teamwork that happens through small judgments about who needs to know, who should act, and whose approval is required.

当智能体只服务一个人时，只要学会那个人的偏好，就能做得相当不错。

When an agent is serving a single person, it gets pretty far simply by learning that person's preferences.

然而，一个在团队中运作的智能体，需要理解人们工作之间的关系。

An agent operating across a team, however, needs to understand the relationships between people's work.

这就带来了一系列不同的问题。

That creates a different set of questions.

这是谁负责的？

Who owns this?

这会影响谁的工作？

Whose work does this affect?

有人在等这个决定吗？

Is someone waiting on this decision?

这个承诺会不会给另一个团队带来义务？

Does this promise create an obligation for another team?

现在的负责人能决定，还是需要更广泛的同意？

Can the current owner decide, or does this need broader agreement?

从某种程度上说，交接成了工作中最关键的环节。

In some ways, the interesting unit of work becomes the handoff.

想象一下，销售人员对客户说：“在您续约之前，我们应该能支持那个集成。”

Consider a salesperson telling a customer, "We should be able to support that integration before your renewal."

对于销售人员的个人代理来说，接下来的步骤可能很简单。

For the salesperson's personal agent, the next steps might be straightforward.

更新客户记录，起草一封跟进邮件，并设置提醒。

Update the account record, draft a follow-up, and create a reminder.

但在组织内部，这句话牵涉到多个责任。

But inside the organization, that sentence implicates several responsibilities.

对销售团队来说，这句话意味着他们可能通过支持那个集成来确保续约。

For the sales folks, that sentence means for them a potential way to secure the renewal, i.e. supporting that integration.

对工程团队来说，这句话意味着一个可能涉及不确定工作的交付承诺。

For engineering, that sentence means a possible delivery commitment involving uncertain work.

对产品团队来说，这意味着路线图优先级可能发生变化，而对客户成功团队来说，这是一个他们可能需要管理的期望。

For product, it means a potential change to roadmap priorities, and for customer success, it's an expectation they may have to manage.

一个多智能体系统需要识别这句话可能是一个跨团队的承诺，而一个判断模型可以评估具体问题。

A multiplayer agent would need to recognize the sentence as a possible cross-team commitment, and a judgment model could assess specific questions.

这条消息是否暗示了交付承诺？

Does this message imply a delivery promise?

这个承诺是否涉及说话者权限之外的工作？

Does the promise concern work outside the speaker's authority?

这和提供的路线图有冲突吗？

Does it conflict with the supplied roadmap?

有没有证据表明负责的团队已经同意了？

Is there evidence that the responsible team agreed?

更大的系统就能生成一个提议的承诺，确定必要的负责人，并请求缺失的决策。

The larger system could then create a proposed commitment, identify the necessary owners, and request the missing decisions.

再强调一下成本和收益，Jev以及这类判断模型的主要成本，如果这成为一个类别的话，就是它本质上是一个不完整的类别。

Now again, to reinforce costs and benefits, the big cost of Jev and this type of judgment model in general, to the extent that this becomes a category, is that it is an incomplete category by definition.

它无法完成我们目前让生成式AI做的所有工作。

It cannot do all the work that we currently have generative AIs do.

判断模型必须成为更复杂的模型架构的一部分，也就是我们过去几个月一直在讨论的那种模型栈。

Judgment Models are going to have to be part of a more complex model architecture, the type of model stack that we've been discussing for the last several months.

当然，好处是它能以极低的成本做到这一点。

The benefit, of course, though, is that it can do this extraordinarily inexpensively.

而且因为这种判断智能可以如此廉价地应用，这意味着如果做得好，它可以极其深入地集成到我们都在构建的自动化系统中。

And because this sort of judgment intelligence can be applied so cheaply, it means that done well, it can be integrated incredibly deeply into the automated systems we're all building.

这显然只是这个全新概念的第一天，而且这个概念非常重要，以至于这个非常能干的团队已经花了两年时间在它上面。

This is obviously just the first day of a very new concept, and a concept which is significant enough that this very competent team has spent two years working on.

所以很明显，我们需要看看它在实践中会如何发展，但当你深入研究时，它确实给人一种既重要又显而易见的感觉，就像那种一旦存在，我们将来会惊讶于我们竟然这么久都没有它的东西。

So obviously, we're going to need to see how it all plays out in practice, but it does have the feel when you dig in of something both important and obvious, the type of thing that once it exists, we will be surprised in the future that we didn't have it for so long.

当然，我会持续关注这个，并且会继续寻找更多人们使用它的例子，以及公司在构建这些新型系统时遇到的挑战。

Certainly, I'm going to be keeping an eye on this, and I will continue to look out for more examples of how people are using it, as well as where companies are running into challenges as they try to build these new types of systems.

不过现在，这些都是非常酷的东西，值得去看看，今天的AI每日简报就到这里。

For now, though, very cool stuff to go check out, and that's going to do it for today's AI Daily Brief.

感谢你的收听或观看。

Appreciate you listening or watching.

和往常一样，下次再见，祝一切安好。

As always, until next time, peace.