← 返回任务列表

How People Are Actually Using Jev

269 段 · 1 位说话人 · 原片 25:42
M1
M10:00

Jev 是我们很长时间以来遇到的最火的模型之一,这不是因为它只是另一个像 GPT-6、Opus 或 Fable 那样的 LLM。

Jev is one of the buzziest models we've had in a long time, and that's because it's not just another LLM like a GPT-6 or an Opus or Fable model.

它在本质上完全不同。

It is something fundamentally different.

但正因为不同,最初它到底最适合用来做什么,并不那么清晰。

But because it's different, it's not necessarily clear at the beginning exactly what it's going to be best used for.

不过现在过了一个半星期,人们已经开始发现并分享各种使用场景,充分利用了 Jev 独特的地方。

With the benefit of a week and a half under our belts now though, people are discovering and sharing a slew of different use cases that take advantage of what makes Jev unique.

今天我们打算聊聊其中最好的那些,以及它们对你来说可能有什么用。

And today we're going to get into the best of them and where they might be relevant for you.

AI Daily Brief 是一个每日更新的播客和视频,内容是关于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.

好了朋友们,开始之前先说几件小事。

Alright friends, quick announcements before we dive in.
M1
M10:59

还有一个快速通知。

One more quick announcement.

我们马上要举办下一场免费的网络研讨会了。

We have our next free webinar coming up shortly.

主题是如何构建你自己的个人 AI 基准测试,这样当新模型出来时,你可以测试一下它对你来说有多好,以及它在你自己的 AI 栈中适合放在哪里。

It's all about how you can build your own personal AI benchmark so that when a new model comes out, you can test it and see how good it is for you and where it will fit into your AI stack.

这次还是由 Nufar Gaspar 主持。

It's led once again by Nufar Gaspar.

完全免费,你可以在 aidailybrief.ai 上找到所有信息。

It will be free, and you can get all the info you need at aidailybrief.ai.

过去几周,AI 圈最火的新东西之一,就是一个叫 Jev 的新模型。

Over the last couple of weeks, one of the buzziest new things to come up in the AI world has been a new model called Jev.

Jev 有意思的地方在于,它不只是另一个 LLM,拿来和 GPT-6 或 Opus 5.5 做同样的事,而是工作方式略有不同,并且正如我们将看到的,是互补的。

Now, what makes Jev interesting is that it is not just another LLM that you would use for the same thing as GPT-6 or Opus 5.5, but actually works in a slightly differently and, as we will see, complementary way.

自发布以来的大约十天里,不仅仅是热度高——我在 X 上看到了几百条帖子,每条都有几百甚至几千个点赞和转发。

In the 10 days or so since launch, not only has there been a ton of buzz— I'm talking hundreds of different posts on X, which each themselves have hundreds or even thousands of likes and shares.

而且这股关注也转化成了巨大的财务机会,《The Information》报道说该公司正在谈判,以至少 100 亿美元的估值融资高达 10 亿美元。

But that attention has also translated into significant financial opportunity with The Information reporting that the company is in talks to raise as much as $1 billion at a $10 billion or higher valuation.

相比于之前 2 亿美元估值下的 4000 万美元种子轮,这算是一个不小的飞跃。

That is a decent jump from its $40 million seed that was completed at a $200 million valuation.

但现在人们已经有机会实际上手 Jev,我想回过头来聊聊,除了社交媒体上那些酷炫的视觉和游戏演示之外,大家实际上是怎么用这玩意的?

But now that we've had a chance for people to actually get their hands on Jev itself, I wanted to go back through and talk about how people are actually using this thing outside of the buzzy visual and game demos that have been all over social media?

基本上,Jev 是那种不是游戏设计师也不是开发者的普通人,也应该关注甚至考虑纳入他们更大的 AI 栈里的东西吗?

Basically, is Jev something that the average person who's not a game designer or not a developer should be paying attention to and even thinking about as part of their larger AI stack?

现在,回顾一下 Jev 是什么,我之前称它为判断模型。

Now, to recap what Jev is, previously I called it a judgment model.

Jev 做不到的是,以你通常认为的聊天机器人那种传统方式写作。

What Jev can't do is write in the traditional way that you think about chatbots writing.

相反,构建 Jev 的 Typesafe 团队称之为系统 1 模型,这个说法来自 Daniel Kahneman 的《思考,快与慢》。

Instead, the team at Typesafe who built Jev calls it a System 1 model after a concept from Daniel Kahneman's Thinking Fast and Slow.

系统 1 思考是快速、本能的模式匹配,也就是快速判断,而 Jev 正是为此而生。

System 1 thinking is fast, instinctive pattern matching, i.e., snap judgments, and that's what Jev is built for.

一个判断某个任务是否适合 Jev 的好方法是,你是否需要反复阅读某样东西,做出一个小判断,然后执行一个可预测的下一步。

A good way to think about a test for whether something is a good job for Jev is where you repeatedly read something, make a small judgment, then take a predictable next step.

比如,一个文件落在你的下载文件夹里。

So take, for example, a file landing in your downloads folder.

现有软件可能有一个普通规则,把它归类为 PDF。

An ordinary rule that existing software might have would be to classify it as a PDF.

但 Jev 能更进一步,识别出它是不是一张发票?

But what Jev can do is go a step further, identifying, is it an invoice?

对应哪个项目?

Which projects is it for?

以及是否需要有人看它?

And does someone need to see it?

一旦做出判断,就可以按部就班地进入系统中的下一步。

Once the judgment is made, it can be predictably moved on to the next step in a system.

你可以问Jev三种核心类型的问题。

There are 3 core types of questions that you can ask Jev.

第一种是选一个,或者Jev所谓的“choice”。

The first is pick one, or what Jev calls choice.

它会回答哪个选项更合适。

It answers which of these fits.

你可以列出最多255个选项,Jev能找到最匹配的那个。

You list up to 255 options and Jev can find the best fit.

举个例子,哪个小组应该处理这个工单?

So an example of this might be, which team should handle this ticket?

是计费、技术支持、销售还是其他?

Is it billing, tech support, sales, or other?

Jev模型会返回你选中的那个,同时还会给出每个选项的概率。

The Jev model is going to give you back the pick plus a probability for every other option.

第二种可以问Jev的问题,他们叫“score”,也就是按等级打分。

The second type of question that you can ask Jev is what they call a score, which is basically rating on a scale.

你把每个等级用文字描述出来,等级从2到10个,然后Jev能判断某个东西落在哪个等级上。

You describe each level in words from 2 to 10 levels, and Jev can answer where something falls.

比如,在从平静到非常愤怒的范围内,这个客户有多不满?

So for example, on the scale from calm to very angry, how frustrated is this customer?

Jev会返回一个在这个范围内的位置,还会显示模型有多确定。

You're going to get back from Jev a position on that scale as well as an indication of how sure the model is.

最后一种问题是是非题,Jev管它叫“null”,是Bernoulli的缩写,它只是回答:这是真的吗?

The last type of question is a yes or no question, what Jev calls a null, which is short for Bernoulli, and it simply answers, is this true?

顺着这个客服的例子,这个人是不是明确要求退款?

So to stay on this customer service example, is this person explicitly asking for a refund?

你得到的是一个0到1之间的概率,表示这句话是真的程度。

What you're going to get back is a probability from 0 to 1 that the statement is true.

X上的Stefan做了一个非常酷又简单的可视化,他在屏幕底部放了一堆表情符号,然后在文本框里输入:你可以输入分类器,比如“能穿戴的东西”或者“冬天能穿戴的东西”。

Hey Stefan on X did a really cool, very simple visualization where he put a pile of emojis at the bottom of a screen and then into a text box he wrote, You can type in classifiers like things you can wear or things you can wear in winter.

匹配的表情会浮到顶部,如果分类变了,它们就会掉下去。

The emojis that match rise to the top, or fall off if the classification has changed.

所以在后台,每个表情都在回答同一个是非题:这个表情符合你输入的内容吗?

And so what's happening in the background is that every emoji is getting the same yes or no question: does this match what you typed?

重要的是,没有人给表情符号做过标注或分类标签——这在传统机器学习里是必需的。

And importantly, nobody labeled or tagged the emoji, as would happen in traditional machine learning.

换句话说,对于“冬天”并没有一条明确的规则。

There's no rule, in other words, for winter.

每个表情在同一时刻都在回答同一个问题:这个匹配吗?

Every emoji gets the same question at the same moment: does this match?

如果你在看这个演示,你可能会意识到:因为Jev能做的事有限且具体,所以它能做得特别快。

And part of what you might be recognizing if you're watching this demo is that because what Jev does is limited and specific, it can do it incredibly fast.

Typesafe发布这个模型的时候说,和类似的LLM流程相比,它速度快20到200倍,成本低40到400倍。

When Typesafe launched the model, they said it can work 20 to 200 times faster and 40 to 400 times cheaper than comparable processes with LLMs.

现在,把这一堆表情符号想象成你的潜在客户、你的工单、你的文件或者其他一大堆没有区分的大量信息。

Now instead of this pile of emojis, imagine that the pile is your leads, your tickets, your documents, or some other big undifferentiated mass of information.

重要的是,你可以同时针对同一个项目问多个问题。

Importantly, you can ask many questions about the same item all at once.

因为问题可以并行运行,问第9个或第10个问题不会花更多时间,只是多消耗一些token。

Because questions run in parallel, asking the 9th or 10th question doesn't cost more time, it just costs more tokens.

比如你有一封客户邮件。

Say for example you have a single customer email.

你可能想同时问三个是非题。

You might want to ask 3 yes or no questions at once.

客户是否生气?

Is the customer angry?

他们是否收到了延迟回复?

Did they get a delayed response?

他们遇到的是不是发错的商品?

Are they dealing with an incorrect item?

这是三个独立的问题,每个问题都会有自己的概率。

These are 3 separate questions that are each going to have their own probability.

在他们的测试中,Typesafe 发现,单次调用处理 13 个问题,比一次一个要便宜 12.2 倍,快 10 倍,而且得到的答案一模一样。

In their testing, Typesafe found that 13 questions handled in a single call was 12.2 times cheaper and 10 times faster than one at a time, and got to the identical answers.

Matthew Berman 对 724 条在投广告做了一次分析,每条广告被问了 12 个问题。

Matthew Berman did an analysis of 724 live ads with each ad being asked 12 questions.

有关于 hook 原型的,有关于格式的,有关于 offer 的,有关于 CTA 意图的,有关于 awareness 的,等等等等。

One about the hook archetype, one about the format, one about the offer, one about the CTA intent, one about the awareness, et cetera, et cetera.

他发现,就那一条广告,单独一次请求处理这 12 个问题,只用了 173 毫秒就完成了。

He found that that single request for the 12 questions for that ad took just 173 milliseconds to complete.

而这才是这个模型真正的厉害之处。

And this is really where the power of the model comes in.

它不仅擅长判断,而且它足够快、足够便宜,可以对每件事都进行提问和判断。

It's not just that it's good at judgment, it's that it's fast enough and cheap enough to ask and make judgments about everything.

一百万个输入 token 只需要 4.2 美分,输出 token 完全免费。

A million input tokens cost just 4.2 cents, and output tokens are free.

而且每次调用的耗时在 70 到 500 毫秒之间,并且就像我们看到的那样,可以把一堆不同的问题合并到一次调用里。

And each call is gonna take between 70 and 500 milliseconds, and as we've seen, can combine a bunch of different questions in a single call.

现在重要的是,Jev 做了一些取舍,这使它擅长做它做的事情,但这也意味着它不会做某些事情。

Now importantly, Jev makes trade-offs that allow it to be good at the things that it does, but that also means that it doesn't do certain things.

它不会写代码,不会起草合同,也不会做出那种涉及各种无法量化和不明确因素的细微决策。

It's not gonna write code, it's not gonna draft contracts, and it's not gonna make nuanced decisions that involve a variety of factors that aren't quantifiable and clear.

它会被用于大规模的小型判断。

It's going to be used for small judgments at volume.

哪个分类?

Which bucket?

有多紧急?

How urgent?

相关吗?

Is it relevant?

安全吗?

Is it safe?

那么接下来我们来谈谈,除了那些酷炫的演示之外,人们实际把 Jev 用起来的 6 种不同方式。

So now let's talk about 6 different ways that people are actually putting Jev to work outside of just cool demos.

第一类是分析你已经拥有的东西。

A first category is analyzing what you already have.

第二类是按语义搜索。

A second category is searching by meaning.

第三类是对进来的内容进行分类分流。

A third category is triaging what comes in.

第四类是根据你的规则检查工作。

A fourth category is checking work against your rules.

第五类是加速 AI agents。

A fifth category is speeding up AI agents.

第六类是即时响应。

And a 6th category is responding instantly.

那我们进入第一类,分析你已经拥有的东西,也就是“这一堆里是什么”。

So let's move to category 1, analyzing what you already have, AKA what's in this pile.

这基本上就是,你拿一个你已经有的集合,对每一个项目问 Jev 同样的问题,然后统计答案。

This is basically where you're gonna take a collection you already have, ask Jev the same questions about every item, and then count the answers.

这种分析,以前用手工做根本划不来,但现在只需要几秒钟,成本只要几分钱。

This is analysis that previously could not be justified doing by hand, but is now gonna take seconds and cost cents.

其中一个例子就是我们刚提到的 Matthew Berman,他分析拆解了来自 37 个品牌的 724 条在投广告,识别出每一个 hook、每一个格式、每一个 offer、每一个 CTA。

One example was that from Matthew Berman that we just talked about, where he analyzed and broke down 724 live ads from 37 brands to identify every hook, every format, every offer, every CTA.

这样做的目的是,拿到这些数据后,再跟真正跑出效果的广告做对比,给你提供关于广告效果极其深入、复杂且精细的信息,这在以前是极为困难的。

The idea is to be able to then take this data and compare it to what actually performed, giving you incredibly deep and complex fine-grained information about advertising performance that would've been extremely difficult before.

Berman 只用了 40 秒和 9 美分的 token,就拿到了那 724 条在投广告的完整拆解。

It took 40 seconds and 9 cents worth of tokens for Berman to get the full breakdown of those 724 live ads.

两天后,Berman 更进了一步,这次他不再只是泛泛地问那些问题,而是让 Jev 以 30 种不同的买家身份去浏览 723 条广告。

2 Days later, Berman went further, and instead of just asking those questions in general, this time he had Jev scroll 723 ads, as 30 different buyer personalities.

这些人物设定是像健身房老板、牙科诊所经理、带幼儿的妈妈,或者对AI好奇的工程师这样的典型角色。

The personalities were archetypes like gym owner, dental office manager, toddler mom, or AI-curious engineer.

Jev的任务是,对于每个广告里的每个买家,判断这个人会停下来看还是继续往下滑。

Jev's job was to ask, for every buyer in every ad, would this person stop scrolling or keep going?

花0.22美元就获得了21690次停或滑的决定。

It cost $0.22 to get 21,690 stop-or-scroll decisions.

当然,这不算真正的数据。

Now, of course, this is not data.

这不是30个真实的买家。

This wasn't 30 actual buyers.

只是AI设想出来的买家典型角色。

It was just buyer archetypes as considered by an AI.

换句话说,最好把它看作一个假设生成器,但我们很难不想到,市场营销会转向把它作为关键环节,在花钱做真实测试之前,用它来压力测试广告或落地页的角度。

In other words, it's better to view this as a hypothesis generator, but it's hard not to think that marketing will shift to using this as a key part of its process to pressure test ad or landing page angles before paying for real tests.

还有很多人把Jev用在整批存档上——不管是邮件存档、过去的X帖子、文档文件夹,还是研究项目的数据库。

Lots of other folks are running Jev over whole archives— whether that's email archives, past X posts, folders of documents, or databases of research projects.

在每种情况下,Jev做的事情都是问同样的几个问题,但针对存档中的每一个条目。

And in each case, what Jev is doing is asking the same few questions, but about every single item in the archive.

比如Ian Nuttall,让Jev对将近3300条他们以前的X帖子各问了8个问题——主题、钩子和语气——然后把这些数据与互动指标做对比,从而更清楚地了解什么内容真正有效。

Ian Nuttall, for example, had Jev ask 8 questions each— topic, hook, and tone— about nearly 3,300 of their previous X posts to be able to then compare that data with engagement metrics to get a better sense of what actually worked.

顺便说一句,3300条帖子各问8个问题,成本大约是13美分。

And by the way, those 8 questions each across 3,300 posts cost about 13 cents.

那么,我们怎么尝试类似的事情呢?

So how might we try something like this?

基本上,你手头任何一大堆积数据——一年的客户邮件、CRM记录、会议记录文件夹——只要你能写几个简单的问题,并对所有查询运行它们,那就是Jev任务可能发挥作用的地方。

Basically any big pile of data you have— a year of customer emails, CRM notes, a folder of meeting transcripts— Anything where you can write a few simple questions and run them against all of those queries, that's a potentially useful place for a JEV job.

第二类是语义搜索。

Category 2 is searching by meaning.

换句话说,哪些内容符合我的意思?

In other words, which of these match what I mean?

这个想法是可以用大白话描述你要找的东西,然后让Jev检查每一个候选。

The idea is to be able to describe what you're looking for in plain words and let JEV check every candidate.

厉害之处在于,即使关键词不匹配,它也能找到你想表达的意思。

The power is that it can find what you mean even when the words don't match.

这方面有很多例子,基本上人们把它当作一种新的网站自然语言搜索方式。

There were a bunch of examples of this, with people basically using it as a new approach to natural language search on websites.

a16z的Justine Moore举了个例子:扫描成千上万个Zillow房源,根据你通常无法筛选的条件对房产进行分类,比如建筑风格、翻新状况或是否靠近高速公路。

Justine Moore from a16z gave the example of scanning thousands of Zillow listings and classifying properties by things you can't normally filter for, such as architectural style, renovation status, or proximity to freeways.

有一个例子特别吸引我,Burhan用一段90分钟的视频,能根据他用自然语言描述的主题来剪辑。

In one example that piqued my interest, Burhan took a 90-minute video and was able to clip it based on themes that they described in natural language.

比如说,在那段90分钟的视频中,他可以查找并剪辑出他们预测AI何时会自动化AI研究的内容,而且在不到2秒内找到了好几处,花费不到2美分。

So for example, across that 90-minute video, they were able to look for and clip, quote, their predictions for when AI will automate AI research, And they were able to find a bunch of examples of that in under 2 seconds for under 2 cents.

另一个例子是SEO审计,Borgia用Jev读取了他们网站的全部586个页面,重建了网站的内部链接地图,耗时45.1秒,花费21美分。

Another example of this is an SEO audit where Borgia deployed Jev to read all 586 pages of their website to rebuild the site's internal link map, which it was able to do in 45.1 seconds for 21 cents.

Claude Opus 5只读了21个页面,花了1.43美元。

Claude Opus 5 only got through 21 of those pages and spent $1.43.

Borgia解释说,内部链接是完美的Jev任务。

Borgia explains, internal linking is the perfect Jev job.

它不需要写作,而是8790次是或否的判断。

It's not writing, it is 8,790 yes or no calls.

这个页面有真正的理由链接到那个页面吗?而且那个页面的锚文本是否已经存在于正文中?

Does this page have a real reason to link to that one, and is their anchor text already sitting in the copy?

这是一个分类问题,而我们之前一直花前沿模型的价钱,一页一页地做这件事。

That is a classification problem, and we've been paying frontier prices to do it one page at a time.

语义搜索的最后一个子类别是根据你关心的内容来过滤阅读。

A final subcategory of this searching by meaning is filtering what you read by what you care about.

比如Robin Bilgil构建了一个实时的AI垃圾内容检测器。

So for example, Robin Bilgil built a real-time AI slop detector.

当他们浏览X时,垃圾内容检测器会给出一个0到1的置信度分数,判断一条帖子是AI垃圾内容的可能性有多大,并屏蔽掉达到一定置信度的帖子。

As they scroll X, the slop detector gives a confidence score on that 0 to 1 scale about how much it thinks a post is AI slop or not, blocking out the ones that reach a certain confidence threshold.

不过,这种过滤功能其实可以用在很多不同的方面。

This sort of filtering though can be applied in a bunch of different ways.

Elvis Sun 在 X 上举了个例子,他让 Jev 浏览 384 条早间新闻,然后识别出其中 15 个不同品牌应该关注哪些新闻。

Elvis Sun on X gave an example of letting Jev browse 384 morning news stories and identifying which of them 15 different brands should be paying attention to.

这个任务花了 24.9 秒,费用是 19 美分。作为对比,在同样的时间里,Opus 5 只处理了 4 篇文章,剩下 380 篇没读。

That was completed in 24.9 seconds for 19 cents, and by way of comparison, in that same time period, Opus 5 got through 4 of those articles, leaving 380 unread.

Jev 的第三类使用场景是信息筛选分类。

Jev use case category three is triaging what comes in.

换句话说,就是搞清楚这是什么,以及该把它放到哪里去。

In other words, what is this and where does it go?

人们正在尝试的问题包括:这封邮件有多重要?

The kinds of questions that people are experimenting with are things like: How important is this email?

这个下载的文件是发票吗?

Is this downloaded file an invoice?

这个链接是不是恶意的?

Is this link malicious?

这个线索应该分给销售团队、让用户自助处理,还是继续培养?

Should this lead go to sales, self-serve, or nurture?

有一个特别有趣的实验来自 Jonathan Unikowski。

One really interesting experiment came from Jonathan Unikowski.

他写道:每个邮件应用都按时间倒序显示收件箱。但如果改成按重要性实时排序呢?

He wrote, "Every email app shows your inbox in reverse chronological order. What if it was live prioritized by importance instead?"

这个场景下,Jev 的任务是给每封邮件评出重要性。在测试中,它用 453 毫秒评了 100 封邮件,花费大约十分之一美分,而且 Jonathan 说 Jev 的评分跟他自己的完全一致。

Jev's job in this case was to rate every email's importance, and in the test, it was able to rate 100 emails in 453 milliseconds for about a tenth of a cent, with Jonathan reporting that Jev's rating matched his own on every single one.

这个功能我现在就想要,不是明天。

This is a use case that I want right now, not tomorrow.

实际上,我昨天就想用了,那些因为我还没回复而急得撞墙的人也一样。

In fact, I want it yesterday, and so do all of the people who are sitting there beating their heads against the wall because I haven't responded to them yet.

还有一些人在尝试对每一条信息快速提出一个问题的模式。

Some other folks are experimenting with a pattern of asking one quick question per item.

Beyond Code 的 CTO Marcel Pocchio 写道:我做了一个 macOS 应用,它可以监控我的下载文件夹,并配合一套可自定义的规则。

Marcel Pocchio, the CTO of Beyond Code, writes, I built a macOS app that monitors my downloads folder along with a customizable set of rules.

下载的文件是发票吗?

Is the downloaded file an invoice?

如果是,就把它移到专门文件夹,并改好文件名。

Move it to a special folder with the correct filename.

整个过程没有调用其他大语言模型,只用 Jev。

No other LLM calls involved, just Jev.

DevEd 用 Jev 做实时聊天审核,在消息出现时就过滤掉脏话和负面评论。

DevEd used Jev for live chat moderation, removing swearing and negative comments as they arrive.

短网址服务 Dubb 的 Steven Tay 把之前抓到的 1 万个恶意域名喂给 Jev,让它用这些标记免费短网址服务中的不良链接。

Steven Tay of the Dubb link shortener fed Jev 10,000 malicious domains they'd caught before to flag bad links on their free shortener service.

他写道:这个问题我们从第一天就在努力解决。

He wrote, this has been something that we've been wrestling with since day one.

用了 Jev,两个小时就搞定了。

With Jev, we solved it in 2 hours.

事实上,这种分类筛选在很多业务流程中都非常核心。

In fact, this sort of triage is so integral to so much of business.

在一篇关于营销人员 10 个 Jev 用例的文章中,HumanX 把工作流决策列为其中之一,并说:大多数工作流最终都会遇到同一个问题:下一步该做什么?

In a post about 10 Jev use cases for a marketer, HumanX identified workflow decisions as one, saying, most workflows eventually hit the same question: what should happen next?

那可能就是使用 Jev 的最佳场景。

That's probably the best place to use Jev.

Box 举了个例子,他们已经试过用 Jev 做事件分类,评估客户影响和严重程度,以及判断该升级到哪个团队。

Box gave an example, having already experimented with incident triage, where Jev is used to judge customer impact and severity, as well as things like which routes it can be escalated to.

Odoo CRM 有一个提议中的模块,对每个新线索问三个问题:优先级、购买意愿,以及是不是垃圾信息。

Odoo CRM has a proposed module that asks 3 things about every new lead: their priority, their buying readiness, and whether it's spam or not.

如果你只能选一个方向去探索的话——尤其是如果你在一家多人接触同样线索、客户或联系人的公司——那么我认为这种分类和路由功能,基本上从整合进系统的那一刻起,就会变得至关重要。

If you had to look at just one area to explore, especially if you were in a company with multiple people touching the same leads or customers or contacts, this sort of triage and routing is, I think, where Jev is going to become absolutely integral, basically from the moment that it gets integrated into the systems.
M1
M117:05

Jev 的第四类用例是:对照规则检查工作,也就是说,检查它是否符合规定。

A 4th category of JEV use cases is checking work against rules, i.e., does this meet the bar?

一般的语言模型会问‘请审查这个’,但你可以把它变成具体的问题,然后对每一版草稿都问这些问题。

Instead of the standard LLM question of please review this, you turn please review this into specific questions and then ask those questions about every draft.

举个例子,LangChain 做了一个场景:每次都用同样的标准给 AI 智能体打分。

One example of this came from LangChain, which was grading an AI agent's work the same way every time.

Harrison Chase 在 LangChain 中指出,Jev 非常适合做评估,尤其是在线评估,就是那种需要大量打分追踪的场景。

Harrison Chase from LangChain pointed out that Jev is, quote, great for evals, especially online evals where you want to grade lots of traces.

这乍一听可能更像是开发者用的东西,而不是公司里其他知识工作者用的。但考虑到我们大家都会越来越多地将智能体投入生产,或者管理已经存在的智能体——比如客服机器人、研究智能体等等——有一个更好的工具来构建评估系统,用于评判这些智能体的工作,这很可能成为核心基础设施。

Now, this might feel initially like something that is more for devs than for other types of knowledge workers in companies, but given how much all of us are going to start putting agents into production or managing agents that already exist— think support bots, research agents, etc.— having a better tool to build evaluation systems into how we judge those agents' work seems like it could very easily become core infrastructure.

Every 团队展示了如何使用这种能力,大规模、超快速地检查工作是否符合规则,从而改进 AI 写作。

And the team at Every showed how you could use this sort of ability to check work against rules at mass scale and at incredible speed as a way to improve AI writing.

他们在 12 段文字中故意埋了错误,Jev 抓到了 7 个中的 6 个,而 Claude 3.5 Sonnet 抓到了全部 7 个。

They planted mistakes on 12 passages of writing, with Jev catching 6 of 7, as compared to Claude 3.5 Sonnet's catching all 7.

不过 Jev 在 0.35 秒内就抓到了 6 个错误,而 Claude 3.5 Sonnet 用了 8.83 秒。而且 Jev 的成本便宜了大约 580 倍,这意味着你可以反复运行同样的检查很多次,总成本更低、速度也更快,不需要用顶级模型来做同样的审查。

Jev, however, caught 6 in 0.35 seconds, as compared to 8.83 seconds for Claude 3.5 Sonnet. and it did so at about 580 times cheaper, meaning that you could rerun that same check a huge number of times and still have it be both cheaper and faster than using a frontier model for the same sort of review.

那你怎么真正把它变成能用的工具呢?

So how might you actually turn this into something that you would use?

基本上,你需要把你写作规则的翻译过程走一遍。

You'd basically need to go through a translation process for your rules for writing.

假设你有一本风格指南,或者一份你永远不想看到的短语列表。

So let's imagine that you had a style guide or a list of phrases you never want to see.

你可以把每个短语变成一个‘是’或‘否’的问题,然后对你下一篇长文档的每一段都问这些问题,这样就能确保你的关键稿件里没有 AI 风格或其他写作雷区。

You could turn each of those into a yes or no question and then run those yes or no questions on every paragraph of your next long document, as a way to ensure no AI-isms or other writing third rails in your key communications.

我觉得这触及了 Jev 会带来的最大思维转变之一。

And this gets, I think, at one of the biggest rewirings that we're going to need to do with Jev.

Jev 的独特价值不仅仅是能完成一件事。

A lot of the unique value of Jev is not just being able to do a thing.

而是能以如此大的规模完成这件事,以至于它变成了质的区别,而不是量的区别。

It's being able to do a thing at such scale that it actually becomes a difference in kind rather than a difference in scale.

换句话说,真正能够逐句细致地检查 AI 风格,这跟对整个文档跑一次通用语言模型检查,完全是两码事。

Being able to realistically check every single sentence in minute detail for AI-isms in other words, becomes categorically different from just running a generic LLM check across the document as a whole.

第五类用例,很多人也在试验,坦白说它更偏向开发者领域,但我觉得为了完整性,还是值得讨论一下。

The 5th category of use cases that lots of people were experimenting with admittedly does get a little bit closer to the developer realm, but I think at least for the sake of completeness, it's still worth discussing.

这些用例可以总结为‘加速你的 AI 智能体’,很大程度上跟过去几个月我们一直在聊的模型路由有关。

These use cases you might sum up as speeding up your AI agents, and a lot of this is around the sort of model routing that we've been talking about for the past several months.

人们问的问题类似于:哪个模型能处理这个任务?

The kinds of questions people were asking were things like, which model can handle this task?

这一步需要多少推理?

How much reasoning does this step need?

这个请求匹配哪个技能(如果有的话)?

Which skill fits this request, if any?

这个旧工具的输出还有用吗?

Is this old tool output still relevant?

你看,在所有这些场景里,共同点是人们用这些小型的自动化微判断,把智能体路由到正确的智能层级、正确的上下文、正确的技能和工具,以便最高效地完成它的任务。

You can see how in each of these cases, the common thread is people using these small automated microjudgments to route an agent to the right level of intelligence, the right context, the right skills, the right tools to do whatever its job is in the most efficient way possible.

人们在任务前用 Jev 来选择模型。

People use Jev to pick a model before a task.

我让它在 Codex 内部的任务执行过程中动态调整 GPT-6 reasoning 的投入程度。

I made it change GPT-6 reasoning's effort inside Codex during the task.

卡住时多思考,常规步骤少思考。

More thinking when stuck, less for routine steps.

在测试中,他们发现 Astra 成本降低了 50%,运行速度也更快了。

And in their tests, they found 50% lower Astra costs while also getting faster runs.

人们还用 Jev 来管理上下文窗口。

People are also using Jev to manage the context window.

Daniel San 做了一个叫 Jev Skill Suggestion for Claude Code 的工具,它的机制是:每次请求,Jev 都会分类出哪个技能最匹配任务,然后只把那个技能注入到 Claude 的上下文里。

Daniel San built something called Jev Skill Suggestion for Claude Code, where, quote, for every request, Jev classifies which skill best matches the task, then injects only that skill into Claude's context.

结果,Token 和成本都减少了 88%。

This led to an 88% decrease in tokens and cost.

我知道很多以前不是开发者的人,现在已经对 Codex 和 Claude Code 这类工具非常熟练了,还建了很大的技能库。而这类基于 Jev 的工具,可能就是一种方法,让你不必在每次请求时都把所有库发给模型。

I know a lot of even you formerly non-developers have started to be sufficiently proficient with things like Codex and Claude Code that you've built up big skills libraries, and these sorts of Jev-based tools are potentially a way to stop sending all of those libraries to the model with every single request.

这里要提一点,虽然我们现在讨论的很多用例目前还只是个人实验,但你也会看到 Jev 和类似它的判断模型被直接内置到我们使用的工具和框架里。

And one thing to note here is that while a lot of these use cases that we're discussing are at this stage individual experiments, you're also going to see Jev and judgment models like it built natively into the tools and harnesses we use.

比如 AJ Asper 写道,我们构建了一个基于 Jev 的新框架,把重复性工作的成本降低了 90%。

AJ Asper, for example, wrote, we built a new harness using Jev that cuts the cost of repetitive work by 90%.

这个框架在运行过程中会学习任务,把步骤从大语言模型调用逐渐迁移到代码里。

The harness learns the job as it runs, moving steps from LLM calls to code.

他们举的例子是合规警报,并且逐批测量了每批警报的成本,每批包含 50 条合规警报。

The example they gave was compliance alerts, and they measured the cost per alert batch by batch with each batch representing 50 of those compliance alerts.

一开始每条警报的成本大概是 2.95 美元,但到了第 1000 条警报时,成本就降到了只有 0.25 美元。

At the beginning, the cost per alert was about $2.95, but by alert 1,000, it was down to just $0.25.

第六类 Jev 用例是关于即时响应的。

A 6th category of Jev use cases are about instant response.

这个人现在到底想要什么?

What does this person want right now?

很多人都在尝试这方面的各种版本。

A lot of people were experimenting with some version of this.

Marcus Lowe 写道,如果复制粘贴能变聪明会怎样?

Marcus Lowe wrote, what if copy-paste was smart?

复制一份简历,粘贴到申请页面,字段自动填好。

Copy a resume, paste into an application, and the fields fill themselves.

X 平台上的 Norman 也做了类似的事,他把粘贴的文本拆开,判断每个字段是什么,进行匹配,再对照原文,最后只粘贴那些匹配确定的字段。

Norman on X also did a version of this, splitting the pasted text into pieces, working out what each field is, matching them, checking against the original, and pasting only the confident matches.

考虑到工作中很多内容就是把信息从一个格式(比如邮件、PDF、会议记录)挪到它该去的地方(比如客户管理系统、录入表单或模板),这确实是一类非常非常相关的用例。

Given how much work at work is moving details from one format, i.e., emails, PDFs, meeting notes, etc., into other places where that information is supposed to live, like CRMs, intake forms, or templates, this is a category of use cases that feels very, very relevant.

另外,知道 Jev 在哪些地方有用,同样重要的是知道哪些地方你要小心使用。

Now, just as important as knowing where Jev is useful is knowing where you got to be careful with it as well.

我觉得需要更加谨慎的几个领域包括招聘、金钱和安全。

A couple places that I think warrant greater caution are areas like hiring, money, and security.

在招聘方面,你可以想象,如果收到一大堆申请,你可能忍不住想用 Jev 这类工具来给它们排序。

On hiring, you can see how if deluged by a set of applications, you might be tempted to use something like Jev to rank them.

但问题在于,即使你给了它很好的标准,Jev 也只会返回一个数字,没有任何推理过程。

But the problem is that even if you've given it good criteria, Jev is going to return with a number that doesn't have any reasoning attached.

顺便说一句,这正好是一个例子,说明你可能需要构建一个更复杂的系统,结合多种类型的人工智能。

This, by the way, is a great example of where you might want to build a more complex system that uses multiple types of AI.

想象一下,你用 Jev 做同样的排序,然后自动让其他类型的大语言模型去复查,比如那些可能值得再看一眼的临界候选人。

Imagine you do that same sort of ranking with Jev, but then automatically have other types of LLM review, for example, for bubble candidates that might be deserving of a second look.

实际上,Jev 的开发者 TypeSafe 也试图明确说明他们认为 Jev 不擅长什么。

TypeSafe, the makers of Jev, actually even tried to be clear about what they think Jev is bad at.

他们列举的一些情况包括多步骤问题,这类问题每一步都会导致准确率下降。

Some of the things they put include multi-step questions where they see accuracy drop with each hop.

他们指出 Jev 不太擅长计数、数学或日期——它能提取事实,但你得在其他地方做实际的计算。

They point out that Jev isn't really good at counting math or dates, that it can extract the facts, but that you're going to want to do the actual math elsewhere.

他们还指出了其他一些问题,比如一致性问题以及解读意图的问题。

And they also point out some other issues like problems with consistency and problems with reading intent.

所以,如果你想判断某个任务是否适合用 Jev,四个标准可能会有帮助。

So if you're trying to figure out if a task is a good fit for Jev, 4 criteria might help.

第一,你能提前写出答案。

The first is that you can write down the answers in advance.

想想分类、是或否这样的二选一、或者一个等级。

Think categories, a yes or no binary, a scale.

如果答案是一句话或者一个计算,那可能就不太适合 Jev 这种判断模型。

If the answer is a sentence or a calculation, that's probably not a good fit for Jev's sort of judgment model.

第二个标准是关于数量。

Criteria number 2 is about volume.

有一堆或一股数据流,比如几百个工单、几百封邮件之类的。

There's a pile or a stream, such as hundreds of tickets, hundreds of emails, things like that.

第三个标准是关于风险——如果答错了,代价很低,或者很容易发现。

A third criteria is about stakes, where a wrong answer is cheap, or it's easy to catch.

基本上,虽然我把它叫做判断模型,但如果答错的后果严重,你不想完全依赖它的判断。

Basically, although I'm calling it a judgment model, you don't want to leave things up to its judgment alone if getting it wrong has big consequences.

最后一个标准是,你能不能把需要的证据以文本形式输入,并且控制在三万两千个 token 以内。

A final criteria is whether you can give it the evidence that it needs as text in under 32,000 tokens.

一旦你找到了一个好任务,下一步就是写出好问题。

Once you've figured out a good task, the next step is going to be to write good questions.

从这些不同例子里总结出的一些技巧,比如每个问题只做一个判断。

Some of the tips there from across all of these different examples are things like one judgment per question.

换句话说,一个不太好的问题就是:这个线索好不好?

In other words, a not-so-good question is, is this a good lead?

因为这实际上不是一个判断。

'Cause that's actually not one judgment.

而是把一大堆不同的判断揉在一个问题里了。

That's a whole bunch of different judgments embedded in one.

“好线索”可能指的是这个行业和你的服务匹配度有多高、公司规模是否适合你的目标客户、他们的购买意向有多强,等等等等。

Good lead might refer to how good a fit the industry is for your service, whether the company size matches who you like to serve, how strong their buying intent is, et cetera, et cetera, et cetera.

基本上,你需要把问题拆分成它的各个组成部分。

Basically, you're gonna wanna break questions into their constituent parts.

如果你的问题涉及一个量表,你需要用文字描述这个量表的每个等级。

If your question involves a scale, you need to describe each level of that scale in words.

而且,你甚至可以充分利用 Jev 的量表功能,问比你认为需要的更多的问题。

And you also might even wanna take advantage of Jev scale opportunities to ask more questions than you think that you need.

最后,像所有 AI 工具一样,在真正放心投入生产使用之前,你需要做一些测试。

Lastly, like everything with AI, you're going to want to do some tests before you actually trust it in production.

这可能会有点麻烦,但比如说,如果你用那个邮件分类器来按优先级整理东西,那不妨自己先标注 50 封邮件作为测试,看看它的结果和你的判断是否一致,确保它能按你的要求工作。

It might be a pain, but if you're using that email classifier, for example, to try to organize things based on priority, maybe you want to label 50 emails yourself as a test and see how it compares to make sure it's actually going to do for you what you want it to do.

现在,在我们收尾的时候,我会把我正在讲解的这份演示文稿发到本集节目的配套网站上,网址是 aidailybrief.ai。

Now, as we wrap up, I will be posting this presentation that I've been working through on this episode's companion site on aidailybrief.ai.

最后几页更偏向实操,针对六个用例领域中的每个项目都给出了一个具体想法,比如内容存档分析器、研究评级器、收件箱分类器、改写检查器等等。

And the last couple of pages get a little bit more practical with one idea for each project in the 6 use case areas, things like a content archive analyzer, a research rater, an inbox triager, a rewrite checker, etc.

里面还会有一个入门提示模板。

And there'll even be a starter prompt in there as well.

在 Jev 发布后的头十天里,我们从充满炒作和兴奋的概念阶段,极快地进入了真正有价值的实际应用场景。

In the first 10 days since Jev was released, we've gone from buzzy, exciting concept to actually valuable production use cases extremely quickly.

然而,由于这本质上是一种新的基础能力——能够大规模、高速、几乎零成本地应用简单判断——我认为我们还需要一些时间来真正弄清楚,它能有多深入地嵌入到各种不同的用例中。

And yet, because this is at core a new primitive in its ability to apply simple judgment at scale, at speed, and for effectively no money, I think it's going to take some time for us to really figure out just how deeply we can weave this into all sorts of different use cases.

随着越来越多的案例出现,我会回来和大家分享其中最棒的。

As more and more come online, I will come back and share the best of them.

不过今天,这期的 AI Daily Brief 就到这里了。

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

一如既往感谢大家的收听或收看,下次再见,peace。

Appreciate you listening or watching as always, and until next time, peace.
已剔除 2 处广告(点击展开查看)
M1
M10:44广告 · 已剔除

First of all, thank you to today's sponsors KPMG, Blitzy, Harbor, and HyperAgent.

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

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

M1
M114:10广告 · 已剔除

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.

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.

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

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

Blitzy's understanding of massive codebases unlocks autonomous security fixes, modernization, and new features.

So what happens when there's no legacy code at all?

Greenfield is supposed to be the easy part— clean slate, no technical debt.

But even greenfield moves at human speed, one sprint at a time.

Blitzy changes the unit of work from the developer to the project, autonomously planning, building, testing, and validating entire applications from scratch.

Hundreds of thousands of lines of production-ready code.

One Blitzy customer stood up a brand new application, 534,000 lines of code, compressing a 65-week roadmap into 2 weeks.

Another shipped an entire application with no frontend engineer.

Legacy or greenfield, the answer is the same: software at the speed of compute.

Build what's next at blitzy.com.

That's blitzy.com.

Every episode, we cover the competition between OpenAI, Anthropic, SpaceX AI, Google, and Meta.

Chances are you've already formed an opinion about who's leading, but every AI lab is taking a different approach, building different technologies, forging different partnerships, and developing a unique ecosystem.

Harbor Capital Advisors AI Lab Ecosystem ETF suite gives investors a way to gain exposure to the AI ecosystem they believe is best positioned for success.

Search Harbor AI Lab Ecosystem ETFs wherever you invest, or follow @HarborCapital on X to learn more.

Visit harborcapital.com for a prospectus containing investment objectives, risks, fees, expenses, and other important information.

Read and consider it carefully before investing.

Risks include principal loss and artificial intelligence-related risks.

Harbor ETFs are distributed by Foresight Fund Services, LLC.

Harbor is not affiliated with AI Daily Brief, and the funds are not affiliated with, sponsored by, or endorsed by any AI lab.

This is a paid advertisement and not personalized investment advice.

Investing involves risk, including possible loss of principal.

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

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

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

Marketing agents turn competitor moves into landing pages.

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

Ops agent chases the paperwork and tracks the budget.

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

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

Hire yours at HyperAgent.

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