Jev is HERE. How to use it
频道: Greg Isenberg
视频: https://www.youtube.com/watch?v=4mTLpuQpB80
原文语言: en
统计: 共 53 轮 · Greg 18 · Ryan 28
[0:00] Greg
Jev is here and it's a big deal. It was created by Dooo Almeida. Yes, that's the same guy whose research built chatbt. Now, it's such a big deal because it's a whole new way to do AI. So, I brought on my friend Ryan who's on the founding team of Open Code to just come on and clearly explain what Jev is, what are some insane use cases, and break down some startup ideas that are now unlocked. As of publishing this, Jev is invite only. But good news, by the end of the episode, you're going to see how you can get access today. So, you're going to want to like, comment, and subscribe right now so your algorithm knows to bring you content like this to get your creative juices flowing in the future. Happy Jev Day, and I'll see you at the end of the episode.
Jev 来了，这可是件大事。它出自 Diogo Almeida 之手——没错，就是那位研究成果造就了 ChatGPT 的人。它之所以是件大事，是因为它是一种全新的做 AI 的方式。所以我请来了我的朋友 Ryan，他是 OpenCode 创始团队的一员，来给大家讲清楚 Jev 到底是什么、有哪些疯狂的用例，再拆解一些现在因它而解锁的创业点子。截至本期发布时，Jev 还是邀请制。不过好消息是，看到节目最后，你就会知道今天怎么拿到使用权限。所以现在就点赞、评论、订阅吧，这样你的算法就知道以后要多给你推这类内容，让你灵感源源不断。祝大家 Jev 日快乐，我们节目结尾见。
[0:49]
[music]
（音乐）
[0:53] Greg
Ryan Vogle, welcome to the pod. By the end of the episode, what are people going to learn?
Ryan Vogel，欢迎来到节目。看完这一期，大家能学到什么？
[0:59] Ryan
We're going to learn about a new type of AI, a an type of AI that we haven't really seen before, but I think it's good. It's Jev. And uh people are ready for this new type of classifier AI because we've been so used to just learning and using these LLMs which are slow they stream and I think uh as we'll cover today this AI is fundamentally different in so many different ways with quality speed and price that there are so many different usage applications for it that the possibilities are truly endless and It just becomes on the humans again about how creative you can be.
我们会了解一种新型的 AI，一种我们以前真没怎么见过的 AI，不过我觉得它很不错。它就是 Jev。大家已经准备好迎接这种新型的分类器 AI 了，因为我们一直太习惯于学习和使用这些 LLM——它们很慢，还是流式输出的。而正如我们今天要讲的，我觉得这种 AI 在质量、速度和价格等很多方面都有本质的不同，能用上它的场景非常多，可能性真的是无穷无尽，最后又回到了人身上：就看你有多大的创造力了。
[1:40] Greg
Cool. And I So I just have a few things I need from you because I haven't used Jev. I want you to give me the simplest possible explanation to Jev. I want you to ex give me like, you know, three or four insane use cases so that people can walk away from this episode with like productivity, making money, just like, you know, even boring use cases that could become, you know, $10 million businesses, $100 million businesses. And I just want you to put it all together, wrap it in a bow that people understand, you know, if they stick around to the end that they'll be able to understand why should they care about it. Can you commit to that, Ryan Vogle?
酷。我需要你帮我做几件事，因为我自己还没用过 Jev。我想让你用最简单的方式把 Jev 解释清楚。我想让你给我讲三四个疯狂的用例，让大家看完这期能带走点东西——不管是提升效率、赚钱，还是那种哪怕很无聊、却可能做成 1000 万美元、1 亿美元生意的用例。然后我想让你把这些都串起来，打包成一个大家都能听懂的完整故事，让大家知道只要坚持看到最后，就能明白自己为什么该关心这个东西。这些你能保证做到吗，Ryan Vogel？
[2:16] Ryan
I can. I can. And I'll add one better. I'll make it entertaining so that way you can actually get excited about it because first up, I'm just going to start out with a demo. This is my email. I'm not afraid to share it. I uh been working with email. If you know me at all, you know that I love email because it seems unsolved. I mean, like, Greg, how many spam emails do you get every day? Like, there's too many, right? There's too many. You can't reply to all of them. And it's just so frustrating. And some of the email algorithms that exist are good, but it's not the best. But then some people are trying to like take like traditional AI where it's like they're having like a GPT 5.6 Luna like kind of read every email and then score it, but that takes time and it's not like instant and it's just like h I wish we could just have something that could like instantly categorize all the emails. So this is that this is using Jev. And before I run it, uh I'm going to break down Jev in a super simple example. Jev is a classifier at its truest being that's what it is. I won't get into the architecture and stuff like that because honestly I don't even understand it that well, but essentially you define an input. Let's say uh you have this iPhone as an input, right? And that's the input and then the output is a schema. So we could uh have the schema be what color is the iPhone is the question almost. and it has uh blue, orange, red, green, yellow as the output options for that question. And the classifier Jev then looks at this phone in a text uh format and says, "hm, what uh is this orange? Is it is it red? It could be red." But then it says, okay, this is about I'm pretty confident it's 80% orange, but it could be 10% red or it could be 10% blue, which adds up to 100. And it's the probabilities of those choices. So, it's not just going to be a 100% affirmative. This is orange, this is blue, this is red. It's a hey, I'm
能，能。而且我还能做得更好：我会讲得有意思一点，让你真正兴奋起来。首先，我直接从一个演示开始。这是我的邮箱，我不怕公开。我一直在折腾邮件这块——如果你了解我，就知道我特别爱研究邮件，因为这个问题好像一直没被解决。我是说，Greg，你每天会收到多少垃圾邮件？太多了，对吧？多得要命，根本回不过来，真的很让人抓狂。现有的一些邮件算法还不错，但也算不上最好。还有些人想用传统的 AI 来做，比如让 GPT-5.6 Luna 之类的模型把每封邮件读一遍再打分，但那样很费时间，不是即时的，就让人觉得：唉，要是有个东西能瞬间把所有邮件分好类就好了。而这个就是这样的东西，它用的就是 Jev。在跑之前，我先用一个超级简单的例子把 Jev 讲清楚。Jev 说到底就是一个分类器，它的本质就是这个。架构之类的我就不展开了，因为说实话我自己也没太搞懂。但基本上就是你先定义一个输入。比如说，你把这台 iPhone 当作输入，对吧？这是输入，然后输出是一个 schema。我们可以把 schema 设成——差不多就是一个问题——这台 iPhone 是什么颜色？这个问题的输出选项有蓝、橙、红、绿、黄。然后分类器 Jev 会以文本形式去看这台手机，然后说：「嗯，这是橙色吗？是红色吗？也可能是红色。」接着它会说：好，我相当有把握，80% 是橙色，但也可能 10% 是红色，或者 10% 是蓝色，加起来是 100。这就是这几个选项各自的概率。所以它不会给你一个百分之百肯定的结论，说这就是橙色、这就是蓝色、这就是红色。而是说：嘿，我有
[4:25] Ryan
80% confident that this is orange or this is red. And the best way to illustrate that is with this email example. So each one of these rows that you see on the table is a full email object. It's got a subject. It's got a description. It's got a body. It's got a sender. All the the snazzy email jazz. And what the input is is that just entire email object. There's no sugar coding or any special treatment. It's just the email object. And we have four outputs. We've got a category which is an option where basically it can say is this shopping, work, marketing, finance, security, yada yada yada. Then we've got a priority which it can allow to select from I think five different options where it's like low priority, medium, high, important or urgent which is like oh no, you have a missed credit card payment or something like that. That's obviously urgent. You want to be able to nail that right on the head as soon as that comes in. And then we have a spam score. This is what I was talking about with those percentages. Obviously, not every email is going to be a true or false when it comes to spam. It's going to be a percentage. It's it's a it's a range, if you will. So, it's like some emails are more spammy, like this uh Kickstarter one. It's obviously trying to sell me a bunch of stuff and junk. I don't really care about that. I signed up for that Kickstarter thing like two years ago. Still haven't been able to unsubscribe from the list since. And then we've got some uh some like Mercury things. Okay, this is just like a payment thing. It's like, okay, Exxon Enterprise received $22 from Stripe.
80% 的把握这是橙色，或者是红色。最能说明这一点的，就是这个邮件的例子。表格里你看到的每一行都是一个完整的邮件对象：有主题、有描述、有正文、有发件人，所有那些花里胡哨的邮件要素都在。输入就是这整个邮件对象，没有任何包装或特殊处理，就是邮件对象本身。我们有四个输出。一个是类别（category），这是一个选项字段，它基本上可以判断这封邮件属于购物、工作、营销、财务还是安全，诸如此类。然后是优先级（priority），我记得可以从五个选项里选：低优先级、中、高、重要，或者紧急——紧急就是比如，糟了，你有一笔信用卡还款逾期了之类的，这显然很紧急，这种邮件一进来你就得第一时间精准抓住。然后是垃圾邮件评分（spam score），这就是我刚才说的那些百分比。显然，并不是每封邮件在是不是垃圾邮件这件事上都能简单地判个真或假，它会是一个百分比，可以说是一个区间。所以有些邮件垃圾程度更高，比如这封 Kickstarter 的，它明显是想卖给我一堆乱七八糟的东西，我根本不在乎。我大概两年前注册了那个 Kickstarter 的东西，从那以后一直没能从他们的邮件列表里退订。然后这里有一些 Mercury 的邮件。好，这就是一条支付通知，大意是：好的，Exon Enterprise 从 Stripe 收到了 22 美元。
[5:50] Ryan
That doesn't seem spammy. That seems just like it's infor uh informative uh and it's just informing me that uh something happened. And then we've got the reply percentage. This is how much does this warrant your reply. So, if we go back here, and I'm not going to click on this because this is a real email, but 90% account violation possibility. This is a user saying, "Hey, my account seems to be violated somehow." Jev identified, hey, this user seems to be having some trouble. We should probably warrant a response on this. Now, I've already got these all uh categorized, and there are uh 1,700 of these emails. And [sighs] and and this is where we come back where it's it's so sad because it just takes so much time to run all of these and it's probably going to take like 10 hours to do and then and then I'm going to have to go through and probably pick out some of the data and oh my god the price is going to be so expensive and oh it it's done. Oh it didn't cost 18 cents
这看起来不像垃圾邮件，只是一条告知性的通知，告诉我发生了某件事。然后是回复百分比（reply percentage），也就是这封邮件有多值得你回复。我们回到这里——我不会点开它，因为这是一封真实的邮件——但你看这封 account violation possibility（账户可能违规），90%。这是一个用户在说：「嘿，我的账户好像不知怎么被判违规了。」Jev 识别出来了：嘿，这个用户好像遇到了麻烦，我们大概应该回复一下。现在，这些我已经全都分好类了，一共有 1700 封邮件。然后（叹气）……然后就到了让人难过的环节：把这些全跑一遍要花好多时间，大概得跑个 10 小时，然后我还得一封封过一遍，大概还得挑出一些数据来，天哪，费用肯定贵得要命……哦，跑完了。哦，才花了 18 美分
[6:51] Ryan
1,700 emails.
1700 封邮件。
[6:53] Ryan
That is the power of Jev. I can't explain it any better than that. We had 4.2 million input tokens and 500,000 output tokens. The entire cost was 18 cents for each one of those emails. All categorized, all I mean, you can see here they're all categorized. They're all ranked. They're all given that score.
这就是 Jev 的威力，我没法解释得比这更清楚了。我们用了 420 万个输入 token、50 万个输出 token，整个下来只花了 18 美分，这里面的每一封邮件都分好类了——我是说，你看这里，全都分好类了，全都排好了序，全都打好了分。
[7:18] Ryan
So, if you were to imagine like let's say
所以，你可以想象一下，比如说
[7:20] Greg
Ryan, what I'm here's what I'm hearing. I just want to make sure I I I have a good mental model for what Jeb is and correct me you know where I'm wrong. Okay.
Ryan，我听下来的理解是这样的——我只是想确认一下，我对 Jev 是什么有个准确的心智模型，哪里不对你就纠正我。好。
[7:30] Greg
So Jev is basically like an AI decision maker.
所以 Jev 基本上就像一个 AI 决策者。
[7:34] Ryan
Yes.
对。
[7:34] Greg
So you you give it some information. In this case you're giving it you know the contents of the email and like a set of possible choices like is it spam or not? Jev's going to go ahead and look at that information and choose an answer. So, for example, like is this email spam or urgent or no normal? Um, but you can also have it do things like, you know, is this customer likely to buy or unlikely to buy,
你给它一些信息——在这个例子里，你给它的是邮件内容，再加上一组可能的选项，比如是不是垃圾邮件？Jev 就会去看这些信息，然后选出一个答案。比如说，这封邮件是垃圾邮件、紧急邮件，还是普通邮件？不过你也可以让它做别的判断，比如这个客户是有可能购买，还是不太可能购买，
[8:01]
right? Exactly. You're almost there. That's like 90% correct. It makes a it makes a probability of a decision.
Greg： 对吧？
Ryan： 没错，你已经很接近了，大概对了 90%。它给出的是一个决策的概率。
[8:07] Ryan
Okay. So the difference between it making a decision because a decision would be you uh like you uh submit an API or something like that and it tells you buy or not to buy. Technically what happens on the underside is that percentage. So it would be like 83% buy uh 17% no buy type of thing. And obviously the more the the answer that is the stronger percentage would win and that would get returned to you. But it's not a 100% decisive action type of thing.
好。所以区别在于——说它做出决策，意思是比如你调用一个 API 之类的，它告诉你买还是不买。但从技术上讲，底层发生的其实是那个百分比，比如 83% 买、17% 不买这样。当然，百分比更高的那个答案会胜出，然后返回给你。但它并不是那种百分之百果断的动作。
[8:39] Greg
Okay. So instead of asking chat GPT claude whatever read this email and explain what I should do. You're you're the new mental model is use ask Jeb I mean you ask Jev like how you know read this email and choose a set of actions. So like reply or escalate and then you get like a choice from Jev and that gives you some sort of confidence score. Is that the way to think about it?
好。所以，与其让 ChatGPT、Claude 之类的读这封邮件、再解释我该怎么做，新的心智模型是去问 Jev——我是说，你让 Jev 读这封邮件，然后从一组动作里做选择，比如回复还是升级处理，然后你会从 Jev 那里拿到一个选择，外加某种置信度分数。是这么理解吗？
[9:06] Ryan
It's kind of so we've been the the LLMs that we know nowadays have like corrupted our minds so much because there was an interesting point you said. You said ask Jev. You don't really ask Jev because Jev isn't a text model. What's really interesting, if you look at the actual spec of Jeb, it doesn't generate any text at all, which you're like, okay, that's kind of weird. Um, it obviously generated text because how did you get the data for this, right? That was defined in the schema. So, let me let me see if I can uh pull up a little little whiteboard here, a little whiteboard action. Not too good at this. So, we've got our schema, right? We'll call it uh I don't know. Um we'll have our email, right? And this will be our email input. And and then we'll we'll do a circle for Jev. Jev Jev seems like a circle guy. I would say Jev,
算是吧。我们现在熟悉的这些 LLM 真的把我们的脑子带偏得太厉害了，因为你刚才说了一个很有意思的点：你说「问 Jev」。其实你并不是在「问」Jev，因为 Jev 不是文本模型。真正有意思的是，你去看 Jev 的实际规格说明，它根本不生成任何文本，你会觉得：好吧，这有点怪。嗯，它显然生成了文本啊，不然这些数据是怎么来的，对吧？那是在 schema 里定义好的。我来看看能不能调出一个小白板，来点白板操作。我不太擅长这个。好，我们有我们的 schema，对吧？就叫它……我也不知道。嗯，我们有邮件，对吧？这是我们的邮件输入。然后我们给 Jev 画个圆。Jev 看起来就像个圆圆的家伙。我会说 Jev，
[9:55] Greg
there's the entertainment you promised. There we go.
你承诺的娱乐性来了。这就对了。
[9:57] Ryan
Yeah, exactly. Jev seems like a circle guy. That's that's just the type of guy that Jev seems like. Okay. Maybe a tiny circle. There we go. Tiny circle because he's fast, you know? It's fast and cheap. Okay. So, we've got our email and that goes in to Jev. It doesn't get asked to Jev. It doesn't You're not asking Jev, hey, what should I do with this email? It's just an input like an a like a standard API and you define a schema up here and we'll have a we'll have like a a simple little schema and be like is spam and that can be a what they call a new which is a true false but it's a scale. So it could be a one to zero. Let me format this. Yeah, I told you I wasn't good at whiteboards. I don't I don't know about this. So it could be a 1 to 0 which means that it could be 0.31 or it could be I don't know like 90 and that's that percentage. So it's if it were to return is spam uh 0.90 that would be a 90% chance that it is spam type of thing. So it doesn't give those definitive answers but you can infer definitive answers from that sort of choice. And then let me get rid of this. Why are we doing Jason? Um, and then we could have a choice like uh let's see category and that would be like uh marketing uh it could be finance. It could be uh spam and it doesn't generate the categories itself. it looks at the categories that you've passed into it as like a model because like you pass all of these this like essentially this output schema in and you say here's the email here's the output schema I need you to generate the answer for me and
对，没错，Jev 看起来就像个圆圆的家伙，Jev 给人的感觉就是这种类型。好，也许画个小圆圈。好了。小圆圈是因为它快，懂吧？又快又便宜。好，我们有邮件，邮件输入给 Jev。它不是被拿去「问」Jev 的——你不是在问 Jev：嘿，这封邮件我该怎么处理？它就只是一个输入，就像标准的 API 一样。然后你在上面这里定义一个 schema，我们来写个简单的小 schema，比如 is_spam，它可以是他们所说的 noul 类型——就是 true/false，但它是个刻度，所以取值可以是 1 到 0。我来调一下格式。是啊，我说过我不擅长画白板，这个我真搞不太定。所以它可以是 1 到 0 之间的值，也就是说可能是 0.31，也可能是——我随便说——90，这就是那个百分比。所以如果它返回 is_spam 0.90，那就差不多是说有 90% 的概率是垃圾邮件。所以它不会给出那种确定的答案，但你可以从这种选择里推断出确定的答案。然后我把这个删掉——我干嘛要写成 JSON？嗯，然后我们可以加一个 choice（选项）字段，我想想……比如 category（类别），选项可以是营销、财务，也可以是垃圾邮件。这些类别不是它自己生成的，它作为一个模型，会去看你传给它的那些类别——因为你基本上是把这整个输出 schema 传进去，然后说：这是邮件，这是输出 schema，我需要你帮我生成答案，然后
[11:48] Greg
by the way a schema is just a fancy word for how a database is organized right
顺便说一句，schema 就是「数据库是怎么组织的」的一个高大上说法，对吧？
[11:53] Ryan
it's just how the database is organized but not even the database it's just how the output is organized it's just a fancy way which is why the develop why all the developers love it because they're like oh my gosh It's actually type- safe, which is a whole another video on everything like that. But it just means that you can take the output that this Jev model gives you and instantly use it in code because like this new that it returns is a number object. It's not like text that is a number or something weird that you would have to do some additional data processing on. It just basically gives you this object which is the structure of the data. And um so like let's say we pass in this email and we have these two uh classification categories. So then the model would just evaluate okay is this spam and what's the category and it would just return the percentage and the category. So it's not exactly like generating text like in a traditional like LLM like chat GBT. It's not saying hm well I think this is a spam email from Kickstarter so I should probably rate it. Nope. It just says category spam is spam 90%. type of thing. There's no internal reasoning or anything like that. Which is why people are like well I don't know if I can trust it because the whole recent development with AI as you've probably seen is the models are reasoning which is basically just saying the models are speaking out loud to identify possible issues in their sort of thought progression. And Jev doesn't do that at all. Or it might do that, but it might just do it like really fast on the server. We don't really know. But from our point of view, it doesn't reason. It doesn't have any other text output. It just gives you the output. So just shoots it back.
对，就是数据库怎么组织的——其实连数据库都谈不上，就是输出是怎么组织的，只是个好听的说法而已。这也是为什么开发者都爱它，他们会说：天哪，它居然是 type-safe（类型安全）的——这个展开讲又能单独出一期视频了。简单说就是，这个 Jev 模型给你的输出，你可以直接拿到代码里用，因为它返回的这个 noul 就是一个数字对象，而不是一段写着数字的文本，或者其他什么还得额外做数据处理的奇怪东西。它基本上就是直接给你这个对象，也就是数据的结构。那比如说我们把这封邮件传进去，有这两个分类字段，模型就只会评估：好，这是不是垃圾邮件、类别是什么，然后只返回百分比和类别。所以它并不像 ChatGPT 这类传统 LLM 那样生成文本。它不会说：「嗯，我觉得这是一封来自 Kickstarter 的垃圾邮件，所以我大概应该给它打个分。」不会。它只会说：category 是 spam，is_spam 是 90%，就这样。没有任何内部推理之类的东西。这也是为什么有人会觉得：呃，我不知道能不能信任它。因为你大概也看到了，AI 最近整体的发展方向是模型会推理，说白了就是模型把话说出来，在自己的思考过程中找出可能的问题。而 Jev 完全不这么做。或者它也许会做，只是在服务器上做得特别快，我们也不清楚。但从我们的角度看，它不推理，也没有任何其他文本输出，它就只给你输出结果，直接甩回来。
[13:37] Greg
It just gives you a decision.
它就只给你一个决策。
[13:39] Ryan
Exactly.
没错。
[13:40] Greg
That's the way to think about it. That's the way I'm starting to think about it.
应该这么去理解它。我现在也开始这么理解它了。
[13:42] Ryan
It's it's a decision model. And that's what I uh I pointed it out. Um uh like right here, like all of these are just decisions. It's not
它是一个决策模型。我之前也指出过这一点——就像这里，所有这些都只是决策。它不是
[13:52] Ryan
cuz everyone has started to uh assimilate AI with LLMs which is like that next token prediction where it's a conversational agent. This isn't that at all. This is still AI because it's like machine learning but it's a decision model strictly. You can't ask it to be like hey how are you doing today or can you? So, I I like to I like to think around with these ideas a little bit. And I was like, okay, it's a decision model, right? Well, I'm a decision model. When I'm typing on my keyboard, I'm making the decision to type each letter. So, like if I were to type hello, I'm making the decision to type H, E, L, L, O, which is technically text, but I'm also making the decision for each key. So I'm like, what if I can apply that same principle to Jev? So if we go back to our Excalibur whiteboard here, let's say instead of this category, we just have all of the letters A through Z, right? And each one of those letters is a new. So the model can basically predict each letter and say, okay, what's the percentage? What's the decision of this letter based on previous letters? So, if it types H E L L, it's like, okay, my next best decision is to type O to complete the word hello. And I didn't know how it would work, but uh this is how it worked. So, this is me asking it the prompt, what is bigger, a cat or an elephant?
因为大家已经开始把 AI 和 LLM 画等号了，也就是那种做下一个 token 预测的对话式 agent。这个完全不是那样。它依然是 AI，因为它也算机器学习，但严格来说它就是个决策模型。你没法跟它说「嘿，你今天过得怎么样」——还是说，其实可以？我喜欢拿这些想法稍微琢磨琢磨。我当时就想：好，它是个决策模型，对吧？可我自己也是个决策模型啊。我在键盘上打字的时候，每打一个字母都是在做决策。比如我要打 hello，我就是在决定依次敲 H、E、L、L、O——从技术上讲这是文本，但同时我也在为每一个键做决策。所以我就想：如果我把同样的原理用到 Jev 上会怎样？我们回到这个 Excalidraw 白板上，假设不要这个 category，而是放上 A 到 Z 全部字母，对吧？每个字母都是一个 noul。这样模型基本上就能预测每个字母，并说：好，这个字母的百分比是多少？根据前面的字母，这个字母该怎么决策？所以，如果它已经打出了 H、E、L、L，它就会想：好，我下一步的最佳决策是打 O，把 hello 这个词补全。我当时也不知道能不能行，结果是这样的。这是我问它的 prompt：猫和大象，哪个更大？
[15:24] Greg
H,
嗯，
[15:25] Ryan
and this is all real time, by the way. So, it's very fast, but obviously it's not as trained in these sort of next uh letter uh completion stuff, but it's still fun to see because it's just like this is cool, but it also shows this isn't a traditional type of LLM where it like you can talk to it and it's a conversation. It's a decision-based LLM, which we've kind of learned.
顺便说一句，这全是实时的。所以它非常快，不过显然它没怎么针对这种「逐个预测下一个字母」的补全任务训练过。但看着还是挺好玩的——一方面这确实很酷，另一方面也说明它不是那种传统的 LLM，不是那种你能跟它说话、来回聊天的模型。它是一个基于决策的 LLM，这点我们前面也算弄明白了。
[15:46] Greg
So, I guess that I mean that begs the question around what should I use Jeb for? Especially the person listening to this is someone who wants to build a business, who wants to invest in themselves, who, you know, could be a sidetime job, a side hustle or their own thing, and they see this and they're like, I I notice that this is really interesting. I see like I I I believe Ryan when I when I when I hear him talk, and I could see that this is a glimpse into the future, but I don't know how to use it.
所以我想，这就引出了一个问题：我到底该拿 Jev 来干什么？尤其是正在听这期节目的人——他们想做一门生意、想投资自己，可能是份兼职、搞个副业，也可能是做自己的事。他们看到这个会想：我注意到这东西真的很有意思；我听 Ryan 讲的时候，我是信他的，我也能看出这让人窥见了未来，但我不知道该怎么用。
[16:15] Ryan
Right? And [snorts] there's something really interesting about this because this is the first model that's been that can cater to a lot of different applications, which I'll say in a second, but it's also really fast and really cheap. So, it's you don't have this high barrier to entry that we've seen with other AIs where it's like, okay, I've got to dedicate like $1,000 a month to this. you could dedicate like $5. Like when we got uh when our open code team got set up on this account, we had like a $5 like I guess like intro uh credit I guess on the account. We were able to use that for 2 days without hitting it. And we were using it like a ton. Like all of my demos and everything like that, we were using it. So it's extremely cheap. So you could probably like load 10 bucks on it and be good for like maybe 3 months. But some cool things that you could probably use with this is um I already got my girlfriend working on it because she runs a graphic design agency and she gets a lot of inbound and she needs to know if this inbound is high quality or just like if it's just maybe like solicit solicitation spam because she has a contact form on her website. So she's using Jev to essentially say is this a good lead and uh it basically does that same sort of category where it's like is good lead and it ranks that on a percentage. So it's like is good lead and it ranks it from one or 0ero to one.
对。而且这东西有个特别有意思的地方：它是第一个能适配大量不同应用场景的模型——具体有哪些我待会儿说——同时它还非常快、非常便宜。所以它没有我们在其他 AI 上见过的那种高门槛，不用想着「好吧，我每个月得在这上面砸 1,000 美元」，你可能投个 5 美元就够了。比如我们 OpenCode 团队开通这个账号的时候，账户里大概有 5 美元的新手赠送额度，结果我们用了两天都没用完。而且我们用得超级狠，我所有的 demo 之类的全在跑它。所以它便宜得离谱。你大概充个 10 美元，就能用上差不多三个月。至于能拿它做的一些很酷的事：我已经让我女朋友用上了。她开了一家平面设计工作室，找上门的咨询很多，她得分辨这些咨询是高质量的，还是只是推销骚扰之类的垃圾信息，因为她网站上挂了一个联系表单。所以她基本上就是用 Jev 来判断「这是不是一条好线索」。它做的其实就是同一类分类：设一个「是否好线索」，然后给出一个百分比，也就是在 0 到 1 之间打分。
[17:37] Ryan
So if it's like if you get a 98% lead that's a pretty high lead and you're probably going to want to reply to that. But then if you get someone who's like hm yeah I think I might want graphic design but I'm not too sure. they probably don't know what they want and that would probably require more effort from you as a business owner or her as the graphic designer to sort of feel out that client. So, you can use Jev to make a lot of the decisions in your business that you might have to do yourself. So, like going through I I love the email example just because it's such an easy fix. That way, you can go through all your emails and all your historical emails and be like, "Are there any leads I missed? Are there any high-v value clients that I could maybe attack again to see if I can extract more value for them and me? And basically, you can kind of think through your workflow and anything where you're looking at some data. It can be any type of data. If you're looking at some data and thinking, hm, I have to make a decision on this, you should probably think about adding Jev at that layer. Obviously, not for like 100% of interactions and stuff like that. should be a very heavy advisory role, but Jeb is really good because it can make those split-second interactions. Um, if you run a business that has a a contact form or like for issue triage, um, let's say you get a lot of support uh, inquiries and someone comes in and asks you and they're like, "Hey, I need help with XYZ product." Jev can do instant classification and say, "Okay, let's make the decision. What product team does this need to get routed to? Let me route it over here.
所以，如果你拿到一条 98% 的线索，那分数相当高，你多半会想回复它。但如果来的人是那种「嗯……对，我好像是想做点平面设计，但也不太确定」，那他们多半自己都不知道想要什么，这就需要你这个老板——或者说她这个设计师——多花些力气去摸清这个客户的底。所以，生意里很多原本得你亲自拍板的决定，都可以交给 Jev 来做。比如去翻……我特别喜欢邮件这个例子，因为它太好解决了。这样你就能把所有邮件、所有历史邮件都过一遍，问问：「有没有我漏掉的线索？有没有哪些高价值客户，我可以再去跟进争取一下，看能不能为他们、也为我自己挖出更多价值？」基本上，你可以把自己的工作流梳理一遍，凡是你在看数据的环节——什么类型的数据都行——只要你看着数据心想「嗯，这里我得做个决定」，那你大概就该考虑在这一层加上 Jev。当然，不是说 100% 的交互都交给它，它更应该扮演一个以提建议为主的参谋角色。但 Jev 真的很好用，因为它能在一瞬间做出这些判断。比如你的生意有个联系表单，或者要做问题分诊：假设你收到大量客服咨询，有人进来问你：「嘿，我的 XYZ 产品需要帮助。」Jev 能瞬间完成分类，然后说：「好，来做个决定：这个该路由给哪个产品团队？我把它转到这边。」
[19:05] Ryan
Let me route it over here." And there's so many different things where if you say, "Hm, this is a decision. Maybe I can use Jev here." I guarantee you will have good results and it will be super cheap and fast because it takes around 200 milliseconds per query um to Jev no matter like what the input output structure is. So that is something really shocking too because AI can take like up to like 30 seconds for some things and you normally have to do like streaming where then you wait for the response to be done and then you've got to like have a listener and it's all this complex stuff but with Jev you can just do like a boom like quick API call and it just works.
「我把它转到这边。」还有太多场景了：只要你心想「嗯，这是个决策，也许这里可以用 Jev」，我保证效果会很好，而且超级便宜、超级快，因为不管输入输出是什么结构，每次向 Jev 查询大约只要 200 毫秒。这一点也挺让人震惊的，因为 AI 处理有些任务可能要花上 30 秒左右，你通常还得搞流式输出，等响应全部返回，还得挂一个监听器，一大堆复杂的东西。但用 Jev，你只要「砰」地发一个快速的 API 调用，直接就能用。
[19:42] Greg
So Jev is basically this you know AI traffic cop. So there's information that needs to come in and then Jev is going to decide, you know, where it should go and what should happen next. So Jev is basically going to pump out, you know, what what is this information, how important is it, what should happen next, and it's either going to go to a human being in the case of, you know, your girlfriend's agency where it's like, oh my god, this is a lead that she needs to act on like right now. This is like Coca-Cola. This is CMO of Coca-Cola. Um, but if it was, you know, the confidence score was lower, but also like, you know, a local business in Orlando, maybe it's you automate it or use an LLM to to do something, draft something up or or send something or
所以 Jev 基本上就是一个 AI 交警。有信息要进来，然后由 Jev 来决定它该去哪儿、接下来该发生什么。也就是说，Jev 基本上会输出这几样东西：这条信息是什么、有多重要、下一步该做什么。然后它要么交给真人处理——比如你女朋友工作室那种情况：天哪，这条线索她得马上跟进，这可是 Coca-Cola，是 Coca-Cola 的 CMO；但如果置信度分数低一些，而且只是奥兰多的一家本地小商家，那也许你就把它自动化处理，或者用个 LLM 去做点什么，起草点东西、发点东西，或者
[20:28] Ryan
right,
对，
[20:29]
the confidence is so low that you just ignore it. So, am I getting that right? Yeah, that's like spot on where you basically think of anything that you would have to make a decision that would need to be quick and fast and maybe like provide feedback to a user and you can do it with that.
Greg： 置信度低到你直接忽略它就行。我理解得对吗？
Ryan： 对，完全正确。基本上你可以这么想：凡是需要你做决定、而且这个决定得又快又及时、可能还要给用户反馈的地方，都可以用它来做。
[20:48] Greg
So, where my brain goes with you don't know me too well, but like I'm all about like startup ideas. That's what this podcast is about.
所以我脑子里想到的是——你可能不太了解我，但我整个人就是围着创业点子转的，这个播客讲的就是这个。
[20:54] Ryan
Oh, me too. My brain is always always thinking the the next way to to do something like this.
哦，我也是。我脑子也总是、总是在琢磨，下一个能这么玩的路子是什么。
[20:59] Greg
So, I'm kind of like, "Oh, wow. So, Jev now exists. How do I find a business with an expensive queue of incoming information and then just put Jev at the front of that queue? What I mean by that it like what do I mean by a queue? I mean like
所以我就在想：「哇，现在有 Jev 了。那我怎么去找到这样一门生意——它有一条处理成本很高的进线信息队列——然后直接把 Jev 放到这条队列的最前面？」我这话什么意思呢，我说的「队列」指的是什么？我是说，比如
[21:16] Ryan
you got like a lot of inbound coming in where people need stuff from you and you need to get them routed to the correct person.
你有一大堆进线需求涌进来，大家都有事找你，而你得把他们分派给对的人。
[21:23]
Exactly. There's something that immediately
Greg： 没错。
Ryan： 有个想法一下子就
[21:26] Ryan
comes to mind which would require a little bit of architecture but let's say you run a services aggregation business like a like a SAS level on top of a local a lot of local services stuff in your area and you type in and you say hey I need my uh my driveway powerwashed right and Jev could take in that information and then it could take in a lot of the input stuff of like all of the other businesses in the area and it could like return um percentages of which one would probably be the best fit for you. So you type in a form and then you get an instant match with a company that's like near you. Could require it's obviously a little bit more complicated under that, but the Jev could do stuff like that where whenever you uh you know the forms that you always see when you're trying to sign up for a website and it's like get an instant quote and it's never instant and it always is like we'll email you by end of day. then a lot of that stuff could be put into like a classifier and it could genuinely be an instant quote that they could get to say, "Hey, this is a good match. Hey, this isn't a good match." And Jev could be uh used to do that. And that's why the speed of Jev is nice because then that client could see you're not wasting the client's time, which if that client does become your client in the future, that's an insane insanely good uh virtue signal to say, "Hey, we're not trying to waste your time. We're not trying to waste our time. Let's get this done and work on it together." So,
冒出来了，这需要稍微搭一点架构。不过假设你做一个本地服务聚合的生意，相当于在你所在地区一大堆本地服务之上再加一层 SaaS。用户输入：「嘿，我家车道需要高压冲洗一下。」Jev 就能接收这条信息，再把这个地区所有其他商家的一大堆信息也作为输入，然后返回一组百分比，告诉你哪家大概最适合你。也就是说，你填一张表，马上就能匹配到一家离你很近的公司。当然，底层实现显然会更复杂一点，但 Jev 能干这种事。你知道那种表单吧，就是你想在某个网站上注册时总能看到的那种，写着「获取即时报价」，结果从来都不即时，永远是「我们会在今天下班前给你发邮件」。这类东西很多都可以丢进一个分类器，让用户真正拿到即时报价：「嘿，这个很匹配」「嘿，这个不太匹配」。Jev 就可以拿来干这个。这也是为什么说 Jev 的速度很香：客户能看到你没在浪费他们的时间。如果这个客户将来真成了你的客户，这就是一个好到离谱的信号，等于在表态：「嘿，我们不想浪费你的时间，也不想浪费我们自己的时间。咱们把这事搞定，一起推进。」所以，
[22:51] Greg
what other Jev use cases do you want to show?
你还想演示 Jev 的哪些其他用例？
[22:54] Ryan
Let me see. I was I was messing around with this um and it doesn't seem to be doing well, but I wanted to see if I could hook Jev up to a Bitcoin signal. So, basically every minute it would run and it would have this decision mix right here where it would tell me to buy, hold, or sell. And it does not seem to be doing well, which shows that this model is great, but it does have some regressions. I would not put this model in front of like your stock portfolio or Bitcoin or anything like that. This is just for like routing or other sort of decisions like that where it doesn't need insane and model intelligence. Like I did a test with this with GPT6 Astra, the OpenAI's latest frontier model, and it did a little bit better than this because it cross referenced some news information and everything like that. But that's it's it's completely uh it's not apples to apples comparison. it's apples to oranges because it's just a different type of model. So that's where it's like this is something that a classifier and decision maker could be used to do but it's not the best in all of the situations and everything. And
我看看。我之前拿它瞎折腾过一个东西，效果好像不太行：我想看看能不能把 Jev 接到一个比特币信号上。基本上它每分钟跑一次，这里有一组决策选项，它会告诉我该买入、持有还是卖出。结果看起来表现不太好，这说明这个模型虽然很棒，但确实也有短板。我不会把这个模型放到你的股票组合、比特币之类的东西前面去。它就是用来做路由，或者其他类似的、不需要超强模型智能的决策。比如我拿 OpenAI 最新的前沿模型 GPT-6 Astra 做过同样的测试，它比这个表现好一点，因为它会交叉参考一些新闻信息之类的。但这完全……这不是同类之间的比较，而是拿苹果比橘子，因为它们根本就是不同类型的模型。所以说，这是分类器和决策器能拿来做的事，但它并不是在所有场景下都是最优解。然后
[23:59] Ryan
I also let me see if I can find it. Yeah, right here.
我还……我找找看能不能找到。对，就是这个。
[24:02] Ryan
I want to test out. So, I made a little I made a YouTube video here where everyone who makes content is aware of this issue where you make content and you make a like a longer form YouTube video or something like that, but you want clips. And the cool part about this is so this right here, I'm dragging and dropping in a video file. And what this process is going to do, and I'll explain it really quick, is it's going to transcribe the video and get a like word level transcript of it. And then it's going to pass that entire thing into Jev with some different classifier decisions to find the best clips. And we'll get to see how quickly it works. Paste it in. It prepares audio and scores 17 moments and around like 3 seconds. And each of these moments are like one of the interesting parts of the video. They're not like the filler text where I'm like, "So, um, I'm going to set this up." It's like, "Let's go ahead and watch that. It's flying. It's absolutely fine. We've got 1.1 million tokens." yada yada yada. So, it allowed me in this uh demo to be able to find the best clips that I could publish on short form content. So, honestly, and I worked on this for maybe 10 minutes. So, if you worked on this and iterated on this to create your own startup with this type of idea, you could probably get pretty far, especially if you combined it with other different AI agent types. So, that way you could have a really good clipping sort of uh feel on it. But there's so many different uh ideas that you could come up with this. And honestly, the best way that I've thought about it is if you just think about it for like a night in the morning, you'll be buzzing with ideas like, "Oh, I could do this. I could do this." Um, I don't know if I already showed this one, but the browser use uh for browser control with Jev is pretty insane, too. I'm going to play this clip right here. This is in real time uh done by the browser control guys
我想测试一下。我做了个小……我这儿有一条自己做的 YouTube 视频。每个做内容的人都知道这个问题：你做了内容，比如一条比较长的 YouTube 视频之类的，但你还想要切片。这个东西酷的地方在于——看这里，我把一个视频文件拖进来。这个流程要做的事我快速解释一下：它会先把视频转写出来，拿到一份逐词级的文字稿，然后把整份稿子连同几组不同的分类器决策一起交给 Jev，找出最好的片段。我们来看看它有多快。粘进去——它先准备音频，然后大约 3 秒就给 17 个片段打完了分。每个片段都是视频里比较有意思的部分，不是那种「呃，我先来把这个设置一下」之类的填充废话，而是像「我们来看看。跑得飞快。完全没问题。我们有 110 万 token」这种，等等等等。所以在这个 demo 里，它让我找到了最适合发成短视频的片段。说实话，这个我大概只花了 10 分钟。所以如果你在这上面继续打磨迭代，拿这类点子做成自己的创业项目，大概能走得挺远，特别是再结合各种其他类型的 AI agent，这样就能做出手感非常好的切片工具。能用它想出来的点子实在太多了。说真的，我觉得最好的办法就是：你只要琢磨它一晚上加一个早上，脑子里就会冒出一堆点子，「哦，我可以做这个，我还可以做那个」。嗯，我不确定之前有没有给你看过这个，不过 Browser Use 那个用 Jev 做浏览器控制的演示也相当疯狂。我来放这段视频，这是实时录的，是做浏览器控制的那帮人
[25:48] Ryan
or the browser use guys where this is Jev controlling this browser to pick a flight from Zurich to London in 7.1 seconds. Let's watch it. This is all real time by the way. So selecting the dates and it found a flight in 7.1 seconds. If you asked any other sort of like browser use AI agent, this probably would have taken a minute, two minutes, even three minutes in the same type of regard.
——或者说 Browser Use 那帮人做的：这是 Jev 在操控这个浏览器，7.1 秒就选好了一趟从苏黎世飞伦敦的航班。我们来看。顺便说一下，这全是实时的。它在选日期……7.1 秒就找到了一趟航班。要是换成任何别的浏览器操控类 AI agent，同样的任务大概得花一分钟、两分钟，甚至三分钟。
[26:16] Greg
Yeah, that's a big deal. That's a really big deal. Um, if people want to get set up with Jev, how do they do it?
是啊，这可是件大事，真的是件大事。那如果大家想开始用 Jev，该怎么上手？
[26:23]
So, so Jev right now is on a wait list, but by the time this video drops, it might be out in general accessibility. But if you want instant access to it, you can go to the Verscell gateway and they have Jev available on it right away. So you can just instantly start testing it out. They've added some stuff into their AI package so you can start messing around with it. But honestly, if you ask your AI agent and drop it this link and the type safe AI to say, "Hey, how can I start experimenting with Jev, you can probably get started right away." And that's a great way to get started, too. any type of AI agent, you could talk to it about your business and with Jev and say, "Hey, what sort of workflows do I do on the daily basis that could benefit from a decision maker like Jev?" That's a that's a huge tip. I appreciate that. Um, I'll include link uh in the show notes in the description for where you can go and and and play around with this. I'll also include links for where you can follow Ryan. He's got a criminally underfollowed YouTube channel. I think it's like a thousand subs.
Ryan： 嗯，Jev 现在还要排候补名单，不过等这期视频上线的时候，它可能已经全面开放了。但如果你想马上用上，可以去 Vercel AI Gateway，那上面已经直接上了 Jev，你立刻就能开始测试。他们还在自家的 AI 工具包里加了些东西，方便你上手折腾。不过说实话，你只要把这个链接和 typesafe.ai 丢给你的 AI agent，问它「嘿，我怎么开始试用 Jev？」，大概立刻就能上手。这也是个很好的入门方式。随便什么 AI agent 都行，你可以跟它聊聊你的业务，再结合 Jev 问它：「嘿，我每天做的哪些工作流，能从 Jev 这样的决策器里受益？」
Greg： 这个建议太有用了，谢谢。我会在节目说明、也就是视频简介里放上链接，告诉大家去哪儿玩这个东西。我也会放上关注 Ryan 的链接。他的 YouTube 频道关注的人少得离谱，我记得好像才一千个订阅。
[27:23] Ryan
I know.
可不是嘛。
[27:24]
Um, it's crazy. So, uh, I'll include that as well. Ryan, thank you so much for coming on. You know what I'm doing after this? I'm I'm going to this versel link. I'm going to play with Jev. I'm going to start classifying some stuff. I Let me caution you though. It is dangerously It is dangerously addictive. The amount that once you see the speed and once you see the price, you will just be like, "Holy cow." And to all of you guys watching at home or listening, please just try it out. It's so cheap, you won't even notice. It will be like like 1/ one,000 of a cent type of thing to test it out. It is so cheap. Please test it out. This is a new type of AI. If you've ever done any sort of classification or if you just want to build something for your own email or other system, try it out. It's so fun to use and the experience with it. It's just going to be mind-blowing because I don't think we've seen AI this fast in a long time.
Greg： 嗯，太离谱了。所以这个我也会一并放上。Ryan，非常感谢你来上节目。你知道我录完要干嘛吗？我要去点那个 Vercel 链接，去玩 Jev，开始拿它给一些东西做分类。
Ryan： 我……不过我得先提醒你一句：这东西危险地……危险地让人上瘾。一旦你见识到它的速度、看到它的价格，你只会说一句「我的天」。也想对所有在家看视频或者在听节目的朋友说：请一定去试试。它便宜到你根本察觉不到，测一下大概也就千分之一美分的样子。真的太便宜了，请一定去试试。这是一种新型的 AI。如果你做过任何形式的分类，或者只是想给自己的邮件或别的系统搭点什么，都去试试吧。它用起来太好玩了，那种体验……绝对会让你大开眼界，因为我觉得我们已经很久没见过这么快的 AI 了。
[28:17]
All right. Can't wait to play with it. Thanks everyone for your time. Ryan, you're a legend. Uh, and I'll see you next time. See you.
Greg： 好嘞，迫不及待想上手玩了。谢谢大家抽出时间。Ryan，你太牛了。呃，我们下期见。
Ryan： 回见。