王自如讲AI

<b>如果大家有兴趣的话</b> If anyone is interested <b>我们要不要花五到十分钟的时间</b> Shall we take five to ten minutes <b>把我接下来想要</b> to go over what I'm about to <b>以后要正式更新的内容</b> officially update in the future <b>先给大家做一个预览呢</b> Let me give you all a preview first <b>就是分享一下我所理解的</b> Which is to share my understanding <b>这个AI的生态当中</b> Within this AI ecosystem <b>到底是怎样的一个架构</b> Of what exactly the architecture looks like <b>大家可以</b> Everyone can <b>是不是可以进一步的探讨一下</b> Can we discuss this further?

<b>有兴趣听吗</b> Are you interested in listening?

<b>我们用五分钟十分钟讲一讲</b> Let's talk about it for five or ten minutes <b>讲讲可以</b> Let's talk about it <b>好</b> Okay <b>那我们就来讲一讲</b> Then let's talk about it <b>刚才我们讲到</b> Just now we were talking about <b>好</b> Good <b>我们转过来</b> Let's turn our attention <b>刚才我们讲到</b> As we were just discussing <b>现在AI所带来的所有改变</b> All the changes that AI is bringing now <b>其实集中在哪里</b> Actually, where is the focus <b>集中在生成式AI </b> The focus is on generative AI <b>所以我们为了简化的来讲</b> So to simplify things <b>我就不再讲生成式AI了</b> I won't talk about generative AI anymore

<b>我就简称AI好不好</b> Can I just say AI for short?

<b>那么在生成式AI当中呢</b> So in generative AI <b>实际上其实有三个</b> There are actually three <b>非常基础的元素来构成</b> Very fundamental elements that make it up <b>就是不管今天你看到的东西</b> No matter what you see today <b>有多丰富</b> how abundant it may be <b>其实都只有三个</b> there are actually only three <b>第一个呢</b> The first one is <b>底层其实就是数据</b> At its core, it's essentially data <b>你可以叫做Data</b> You can call it Data <b>然后第二个</b> Then comes the second <b>其实非常关键的支柱</b> Actually, a critically important pillar <b>就是模型</b> That's the model <b>第三个非常重要的支柱</b> The third very important pillar

<b>叫做训练方式</b> is called the training method <b>Training</b> Training <b>在计真的技术层</b> At the technical level of computing <b>其实核心的板块</b> The core components are actually <b>只有这三个</b> just these three <b>我们看到的大量的专业术语</b> We see a vast number of technical terms <b>专业的名词</b> Professional terminology <b>基本上都可以被归类为</b> can essentially be categorized into <b>这三个具体的分支当中</b> these three specific branches <b>甚至当他们产生关系</b> even when they establish relationships <b>和互联的时候</b> When connected to the internet <b>比如我们经常所讲到的</b> As we often mention <b>所谓的Cag和Rag</b> The so-called Cag and Rag

<b>都是这三者之间交互协作的结果</b> Are all results of the interaction and collaboration among these three <b>那么这三个东西</b> So these three things <b>我为了方便大家理解呢</b> To make it easier for everyone to understand <b>用最通俗的语言来去比喻</b> I'll use the simplest language to make an analogy <b>比如做菜</b> Like cooking <b>虽然我不会做菜</b> Although I can't cook <b>但是我相信大家能够懂这个概念</b> I believe everyone can understand this concept <b>数据在这样的一个框架当中</b> What does the data mean <b>意味着什么</b> within such a framework <b>特别类似食材</b> Very similar to ingredients <b>你比如说</b> For example <b>不管中国人外国人</b> Whether Chinese or foreigner <b>对吧</b> Right

<b>美国人欧洲人</b> Americans Europeans <b>印度人日本人</b> Indians Japanese <b>只要你是人</b> As long as you're human <b>其实你吃的东西都差不多</b> What you eat is actually quite similar <b>无非是肉、菜、蛋、奶这四类</b> It's nothing more than these four categories: meat, vegetables, eggs, and dairy.

<b>但是我们的加工方式、口味、偏好不同</b> But our processing methods, flavors, and preferences differ.

<b>但是只要是人吃的食材都差不多</b> However, as long as it's food for humans, the ingredients are pretty much the same.

<b>所以你可以理解为</b> So you can understand it as <b>生成式AI的语境之下</b> In the context of generative AI <b>你所需要用来训练的数据</b> The data you need for training <b>其实都差不多</b> is actually quite similar <b>都差不多</b> pretty much the same <b>那么模型这里边</b> So what is inside the model <b>对应的是什么呢</b> What does it correspond to?

<b>其实是菜系</b> Actually, it's cuisine <b>我的字不好看</b> My handwriting isn't very neat <b>大家稍微包容一下</b> Please bear with us a little <b>也就是说</b> That is to say <b>虽然我们吃的东西都差不多</b> Although we eat pretty much the same things <b>但是不同的菜系之下</b> But under different cuisines <b>意味着它所满足的口味不同</b> It means it caters to different tastes <b>我们满足的需求不同</b> We meet different needs <b>你说最近我天天上火</b> You say I've been getting heaty every day lately <b>你还要炒川菜吃辣的</b> And yet you still want to stir-fry Sichuan dishes and eat spicy food <b>你这是给自己找不自在</b> You're just asking for trouble

<b>是不是</b> Aren't you <b>那我倒不如吃粤菜</b> Then I might as well eat Cantonese food <b>口味清淡一点</b> It's a bit lighter in flavor <b>所以说不同的菜系</b> So different cuisines <b>解决的是不同需求</b> address different needs <b>我不知道这样大家能清楚吗</b> I'm not sure if everyone understands this clearly <b>你能清楚吗</b> Do you understand clearly?

<b>能清楚</b> Can understand clearly <b>OK</b> OK <b>那训练方式是什么</b> So what's the training method <b>其实我把它比喻为</b> Actually, I liken it to <b>厨师的手艺</b> The chef's culinary skills <b>就是说你同样的食材</b> That is to say, with the same ingredients <b>咱都是川菜</b> We're all about Sichuan cuisine <b>那你说不同的川菜的厨师</b> Then you talk about different Sichuan cuisine chefs <b>他炒出来的水煮肉片</b> His stir-fried boiled meat slices <b>都是一个味吗</b> Do they all taste the same?

<b>不一定吧</b> Not necessarily.

<b>所以说</b> So that's why <b>这里面就涉到了大量的</b> This involves a significant amount of <b>关键的变量和元素</b> key variables and elements <b>如果我选择了一个菜系</b> If I choose a cuisine <b>我选择了一个</b> I choose a <b>比如说咱们说叫LM</b> For example, let's call it LM <b>叫大语言模型</b> called a large language model <b>OK</b> OK <b>我选择了一个语言类的模型</b> I chose a language-based model <b>然后我所训练的数据</b> and the data I trained it on <b>全部来自于公网</b> All sourced from the public network <b>但是炒菜的人不一样</b> But the chefs are different <b>今天可能是</b> Today it might be <b>腾讯</b> Tencent <b>对吧</b> Right <b>然后这个</b> Then this

<b>ByteDance写BD吧</b> ByteDance writes BD <b>对吧</b> Right <b>或者比如我们今天比较熟悉的</b> Or for example, what we're more familiar with today <b>DeepSeek</b> DeepSeek <b>它的厂家不同</b> It comes from different manufacturers <b>只是意味着</b> merely indicates <b>它对于相同的</b> it applies to the same <b>类似高度相似的数据</b> similarly highly analogous data <b>相同的模型原理</b> The same model principle <b>但是不同的厨子炒出来的口味</b> but different chefs create distinct flavors <b>有所区别</b> there are differences <b>OK</b> OK <b>那两者</b> between the two <b>这三者之间的关系</b> The relationship between these three <b>就搞清楚了</b> is now clear <b>所以</b> so <b>数据里面</b> within the data <b>大家会衍生出</b> People will derive

<b>非常非常多的衍生名词</b> a vast number of derivative terms <b>比如什么叫数据清洗</b> such as what is called data cleaning <b>数据打标</b> data labeling <b>数据站</b> Data Station <b>然后指标管理</b> Then indicator management <b>等等一系列的东西</b> And a series of other things <b>其实都是基于数据</b> Actually, they are all data-based <b>这个板块</b> This section <b>衍生出来的</b> Derived from <b>然后我们看到的</b> Then what we see <b>大语言模型</b> Large language models <b>其实只是一种模型机制</b> It's actually just a type of model mechanism <b>今年所有的生成式AI</b> All generative AI this year <b>语言类的AI</b> Language-based AI <b>都是基于</b> Is built upon <b>这个Google的这个Transformer</b> This Google Transformer

<b>然后诞生出来的</b> then gave birth to <b>然后你所有的这些训练方式</b> and all these training methods of yours <b>为什么大厂都要干这件事</b> why all the big tech companies are doing this <b>是因为大家要抢夺</b> It's because everyone is competing to seize <b>未来的生态的话语权</b> the discourse power over future ecosystems <b>所以这个是在后续的东西</b> So this is something that will be discussed later <b>我们要去讲到的</b> which we are going to address <b>那么你会发现</b> You'll find that <b>食材它不是摘回来之后</b> the ingredients aren't ready to be stir-fried <b>就可以直接用来炒菜的</b> right after being picked <b>它必须得进行处理吧</b> they need to be processed first, right

<b>它必须得洗一洗</b> It must be washed <b>切丝切片等等等等</b> Shredding, slicing, and so on <b>所以数据的预处理</b> So the preprocessing of data <b>变得极为关键</b> Becomes extremely crucial <b>前一段时间</b> A while ago <b>我们看到的</b> what we saw <b>这个叫做什么</b> what's this called <b>AI的血汗工厂</b> AI sweatshops <b>其实讲的就是什么呢</b> Actually, what it's talking about is <b>其实讲的就是有</b> Actually, it's about having <b>无数的这种所谓的数据</b> countless so-called data <b>被拿回来了之后</b> that was retrieved afterward <b>为了用于以后的模型训练</b> For future model training <b>必须要做大量的打标和预处理工作</b> Extensive labeling and preprocessing work is required <b>这个工作高度机械</b> This task is highly mechanical <b>高度重复</b> Highly repetitive

<b>没有任何技术含量</b> No technical expertise required <b>所以很多大的AI公司</b> So many big AI companies <b>就在非洲</b> are in Africa <b>在东南亚</b> and Southeast Asia <b>包括甚至在中国</b> including even in China <b>找了很多的廉价劳动力</b> They found a lot of cheap labor <b>来去帮他们完成</b> to help them complete <b>打标和预处理的工作</b> the labeling and preprocessing work <b>为的是什么</b> What is the purpose <b>为的是能够按照某种菜系的烹饪套路</b> To follow the cooking routines of a certain cuisine <b>结合自家的手艺</b> Combined with one's own culinary skills <b>来去做出自己想要的饭菜</b> To create the desired dishes <b>OK</b> OK

<b>所以为什么说数据都差不多呢</b> So why do we say the data is roughly the same?

<b>大于模型是这样</b> The larger model is like this <b>那比如说咱们举另外一个例子</b> For example, let's take another instance <b>做自动驾驶的车</b> Making autonomous vehicles <b>做自动驾驶的</b> Developing autonomous driving technology <b>虽然说各家采集各家的</b> Although each company collects its own data <b>比如ABC三家公司</b> For example, companies like ABC <b>但是他们采集的方式都差不多</b> but their collection methods are quite similar <b>都是拿一辆车</b> all using a vehicle <b>上面绑一相机</b> with a camera mounted on it <b>绑一雷达</b> and a radar attached <b>满大街跑</b> Running all over the streets <b>跑回来的数据了之后</b> After the data comes running back

<b>进入到数据预处理的过程中</b> Entering the process of data preprocessing <b>切丝切片</b> Shredding and slicing <b>打标</b> Tagging <b>然后做这个备注</b> Then make this note <b>做完备注了之后</b> After completing the note <b>丢到一个相同的模型当中</b> Feed it into the same model <b>去训练出来</b> Go train it out <b>那这里边</b> Then in here <b>手艺和所有的这些都出现了</b> Craftsmanship and all of these have emerged <b>所以那么你会发现</b> So then you'll find <b>在基于这样的一个大的框架</b> Within such a large framework <b>所诞生出来的所有的模型</b> all the models that have emerged <b>就是今天我们所看到的</b> are what we see today <b>比如说DeepSeek</b> such as DeepSeek <b>R1</b> R1

<b>比如说Grok</b> For example, Grok <b>比如说GPT</b> For example, GPT <b>它只是不同的厨子</b> They're just different chefs <b>给它所炒出来的这一盘菜</b> It stir-fried this dish <b>所起了一个不同的名字</b> and gave it a different name <b>就像一碗蛋花汤</b> Just like a bowl of egg drop soup <b>有的人把它往里边扔了一根大葱</b> Some people threw a big green onion into it <b>就起名叫猛龙过江</b> Let's name it "The Dragon Crosses the River" <b>但本质上</b> But in essence <b>它还是一碗蛋花汤</b> It's still just a bowl of egg drop soup <b>OK</b> OK <b>那么在大语言模型的逻辑里面</b> So in the logic of large language models

<b>所有的训练数据</b> All training data <b>基本都来自于公网的</b> primarily comes from publicly available sources on the internet <b>无版权</b> without copyright restrictions <b>无法律纠纷的数据</b> and free from legal disputes <b>所以也就是说</b> So that is to say <b>无论是字节</b> Whether it's ByteDance <b>无论是百度</b> Whether it's Baidu <b>大家训练的基本的语料库</b> The fundamental corpus everyone trains on <b>和原材料都差不多</b> Similar to raw materials <b>但是有没有细小的差别</b> But are there subtle differences <b>有细小的差别</b> There are subtle differences <b>比如说如果字节</b> For example, if Byte <b>想要训练的时候</b> When you want to train <b>是不是可以结合</b> can it be combined with

<b>大量的用户的浏览数据</b> a large amount of user browsing data <b>来生成出对于你</b> to generate results for you <b>更个性化的答案和结果</b> More personalized answers and results <b>这是它的优势</b> This is its advantage <b>那百度是不是</b> What about Baidu <b>对于公网所有的</b> For all public networks <b>这些搜索引擎</b> These search engines <b>所收集到的语言和目录</b> The languages and directories collected <b>更有所谓的优势</b> Have so-called advantages <b>那可能是它的优势</b> That might be its strength <b>它本质上是不是都差不多</b> Is it essentially all about the same <b>本质上是差不多的</b> Essentially, it's about the same <b>所以那么我们就发现</b> So then we realize <b>可以把这个部分叫做什么</b> We can call this part what

<b>叫做底层的技术</b> called underlying technology <b>技术层</b> technology layer <b>然后所生成的</b> then generated <b>大家今天看到的</b> what everyone sees today <b>所有这些语言模型</b> All these language models <b>DeepSeek、Grok、GPT</b> DeepSeek Grok GPT <b>你可以理解为叫做</b> You can understand them as being called <b>叫做初级应用</b> called primary applications <b>OK</b> OK <b>好</b> Okay <b>紧接着讲到了最精彩的部分</b> Then came the most exciting part <b>我们有了</b> We have <b>其实这个底层的数据</b> Actually, this underlying data <b>有了这个初级的应用</b> With this basic application <b>但是你就会发现</b> But you'll soon realize <b>用着用着</b> as you keep using it <b>大语言模型的瓶颈出来了</b> the limitations of large language models become apparent

<b>那就是说</b> which is to say <b>它唯一擅长的</b> The only thing it's good at <b>就是跟你聊聊天</b> is chatting with you <b>跟你提供了情绪价值</b> providing you with emotional support <b>甚至给你算一点</b> or even doing some calculations for you <b>有的没的</b> trivial things <b>但是你会发现</b> but you'll find <b>你的生活当中</b> in your life <b>并不是每一天</b> not every day <b>都有大量的问题</b> there are a lot of problems <b>需要让它去帮你解决的</b> that need it to help you solve <b>你真正需要解决的是</b> what you really need to solve is <b>我希望回家的时候</b> I hope to go home when

<b>外卖就在我门口</b> The takeout is right at my door <b>对吧</b> Right <b>我希望我导航换电池的时候</b> I hope when I navigate to change the battery <b>他能够让我提前导航到电池的换电站</b> It can let me navigate to the battery swap station in advance <b>再给我导航到我的目的地</b> Then guide me to my destination <b>而不需要我手动的将两段导航联系在一起</b> without requiring me to manually link the two navigation sections <b>我需要的是</b> What I need is <b>如果我明天就开始放假</b> If I start my vacation tomorrow <b>想规划一个五一小长假</b> I want to plan a five-day May Day holiday

<b>我只需要跟他讲一句</b> I just need to say one word to him <b>他就能够帮我订票</b> and he can book the tickets for me <b>安排行程解决所有的问题</b> arrange the itinerary and solve all the problems <b>这是你想要解决的现实问题</b> This is the real-world problem you want to solve <b>而不是去问</b> Instead of asking <b>对吧</b> Right <b>说岳飞是在哪一年</b> in which year Yue Fei <b>对吧</b> Right <b>这个被奸臣所害</b> was betrayed by treacherous officials <b>且被斩首的</b> and beheaded <b>这个没有意义</b> This makes no sense <b>对我们的生活当中</b> In our daily lives <b>大部分是没有的</b> Most of it doesn't exist

<b>所以大语言模型的局限性</b> Thus the limitations of large language models <b>就会越来越大</b> it will grow larger and larger <b>这个时候我们就想了一个新的问题</b> At this point, we thought of a new question <b>如果大语言模型</b> If large language models <b>想要解决具体问题的时候</b> want to solve specific problems <b>是不是必须要连接具体的服务</b> Is it necessary to connect to a specific service?

<b>服务一</b> Service 1 <b>服务二</b> Service 2 <b>服务三</b> Service 3 <b>如果想连接服务的时候</b> When you want to connect to the service <b>过去在消费互联网的逻辑当中</b> In the past, within the logic of the consumer internet <b>大家非常清楚一个概念</b> Everyone was very clear about one concept <b>叫做API</b> called API <b>也就是说</b> That is to say <b>服务这个所谓的</b> This so-called service <b>这个所谓的服务提供商</b> This so-called service provider <b>我们可以理解为</b> We can understand it as <b>美团好了</b> Meituan then <b>比如说或者叫去哪儿</b> For example, or called Qunar <b>等等等等</b> and so on <b>你可以理解为</b> You can understand it as

<b>它是服务的提供商</b> It is a service provider <b>那么如果你想调用这个服务</b> If you want to call this service <b>你必须要去通过美团</b> you must go through Meituan <b>所规定好的这个API接口</b> using the specified API interface <b>接入它的商家数据</b> to access its merchant data <b>还有它的物流数据</b> And its logistics data <b>才能够让你在你的APP里边点餐</b> Only then can you order food in your app <b>或者完成所谓的订票的工作</b> Or complete the so-called ticket booking task <b>对吧</b> Right <b>也就是说</b> That is to say <b>在过去的消费互联网</b> In the past consumer internet <b>过去概念当中</b> In past concepts <b>所有的服务的调用</b> all service invocations <b>是高度的格式化的</b> were highly formatted

<b>这个词是关键</b> this term is key <b>是高度格式化的</b> is highly formatted <b>但是在AI的语境之下</b> but in the context of AI <b>你就会发现</b> you will find <b>需求是什么</b> what the demand is <b>需求侧</b> Demand side <b>是高度</b> is highly <b>定制的</b> customized <b>能理解我的意思吗</b> Do you understand what I mean <b>高度定制</b> Highly Customizable <b>这两者之间形成了一个巨大的鸿沟</b> This creates a vast chasm between the two <b>当你的需求高度定制</b> When your needs are highly customized <b>但是调用服务却非常非常格式的时候</b> But the service invocation is extremely rigid <b>是不是两者就连不上了</b> Is there no connection between the two?

<b>他没办法把它连接在一起</b> He couldn't connect them together.

<b>这个时候我们就出现了一个AI时代下的新的概念</b> At this point, we encounter a new concept in the AI era.

<b>叫做MCP</b> It's called MCP <b>全称叫做Model Context Protocol</b> <b>（模型上下文协议）</b> The full name is Model Context Protocol (MCP) <b>也就是说</b> That is to say <b>以大模型的上下文和需求为特征</b> Characterized by the context and requirements of large models <b>来去调用服务</b> To call services <b>我举个非常简单的例子</b> Let me give a very simple example <b>当你的需求是高度定制</b> When your needs are highly customized <b>但是服务本身是格式化的</b> But the service itself is standardized <b>就好像你是一个Lightning接口 雷电接口</b> It's like you're a Lightning connector <b>遇到了一个USB接口</b> Meeting a USB port <b>你俩插不上</b> You two can't get a word in

<b>对不对 无法对接</b> Right? Can't connect

<b>那这个时候我需要一个高度兼容性的一个接口</b> At this point I need a highly compatible interface <b>叫做什么呢</b> What's it called <b>就是大家非常熟悉的USB-C</b> It's the very familiar USB-C <b>所以我们今天不管买手机</b> So today, whether we're buying a phone <b>买电脑</b> a computer <b>还是买所有的东西</b> or anything else <b>我只要看到USB-C</b> Just seeing USB-C <b>我就知道我所有的都能用</b> I know everything will work <b>对不对</b> Right?

<b>你可以理解为MCP这样的这种相调用的方式</b> You can think of it as MCP's phase modulation approach <b>就是我们所说的大模型和服务之间的</b> what we call the large model and the service between <b>USB-C接口</b> USB-C port <b>哇</b> Wow <b>这事儿就有趣了</b> This is interesting <b>如果大模型过去自己解决不了的问题</b> If the large model couldn't solve problems by itself in the past <b>通过带用服务可以解决</b> it can be resolved by utilizing services <b>它的可用性将会大幅度的提升</b> its usability will be significantly improved <b>而这个时候</b> and at this point <b>如果我们还能够加上一个决策机制</b> If we could also add a decision-making mechanism

<b>也就是说当我给它一个具体讯息</b> That is to say, when I give it a specific piece of information <b>比如说啊</b> For example <b>我想一会儿喝一杯咖啡到家</b> I think I'll have a cup of coffee when I get home <b>帮我点杯咖啡送到家</b> Order me a coffee to be delivered home <b>哇</b> Wow <b>你现在回家的路上</b> You're on your way home now <b>AI模型听到这句话了之后</b> After the AI model hears this sentence <b>首先要获取你的位置信息</b> First, it needs to obtain your location information <b>首先要知道你喜欢喝的咖啡口味</b> First, it needs to know your preferred coffee flavor

<b>你经常点在哪一家店</b> Which store do you frequently order from <b>然后以及你家的地址</b> And also your home address <b>并且结合你现在离家的距离</b> Combined with your current distance from home <b>知道OK</b> Got it OK <b>可能30分钟以后下单</b> Probably placing the order in about 30 minutes <b>结果会更好</b> The outcome will be better <b>在这一系列的分析的过程当中</b> Throughout this series of analyses <b>AI必须要做若干决策</b> AI must make several decisions <b>和优先级的排序</b> and prioritize accordingly <b>才能决定什么时候调用某一个外卖服务</b> determines when to call a particular food delivery service <b>帮你把外卖送到家里</b> delivers the food to your home

<b>这个逻辑应该是非常容易理解的</b> this logic should be very easy to understand <b>那么在这个优先级排序和所谓的决策过程当中</b> then in this prioritization and so-called decision-making process <b>它的这个机制就是我们所说的</b> This mechanism is what we call <b>AI智能体</b> AI intelligent agent <b>或者叫做AI Agent</b> or AI Agent <b>或者我们叫做加个形容词</b> or we can add an adjective to it <b>叫做Agentic AI</b> <b>（代理式人工智能）</b> called Agentic AI (Agentic Artificial Intelligence) <b>所以就是通过Agentic AI</b> <b>（代理式人工智能）</b> So it's through Agentic AI (Agentic Artificial Intelligence) <b>所谓的智能体</b> the so-called intelligent agent <b>把大语言模型</b> to empower large language models <b>和所谓的服务连接在一起</b> connected with so-called services

<b>来解决很多很多的具体问题</b> to solve many, many specific problems <b>那么我们来试想一下</b> Now let's imagine <b>如果Agentic AI</b> <b>（代理式人工智能）</b> if Agentic AI <b>或者叫AI Agent</b> <b>（人工智能体）</b> or called AI Agent (Artificial Intelligence Agent) <b>可以通过MCP调用的服务几乎是无限的</b> <b>（MCP：模型上下文协议）</b> The services that can be invoked through MCP are almost limitless (MCP: Model Context Protocol) <b>也就意味着</b> which means <b>AI助理未来能够帮你解决的问题</b> the problems that AI assistants will be able to help you solve in the future <b>几乎是无限的</b> Almost limitless

<b>那么如果把Agentic AI的规模乘以N</b> <b>（代理式人工智能）</b> Then if we multiply the scale of Agentic AI by N (Agentic Artificial Intelligence) <b>可能就会带出一个新的概念</b> It might give rise to a new concept <b>叫做AGI</b> Called AGI <b>叫做General AI</b> <b>（通用人工智能）</b> General AI (Artificial General Intelligence) <b>也就是我们所说的叫做强和高度人工智能</b> Also known as strong or highly advanced artificial intelligence <b>特别典型的例子</b> A particularly classic example <b>就是大家看到钢铁侠里边那个Jarvis</b> Would be Jarvis from Iron Man that everyone's seen <b>说钢铁侠说Jarvis帮我看一看</b> Iron Man said, "Jarvis, take a look at this for me."

<b>那个什么数据</b> That something data <b>那个什么数据</b> That something data <b>其实就是一个所谓的AI Agent</b> Actually, it's just a so-called AI Agent <b>后边再加一个大语言模型</b> Add a large language model behind it <b>再加背后数据具体服务调用的这样一个机制和流程</b> Plus a mechanism and process for invoking specific data services in the background <b>所以为什么最近AI Agent非常的火</b> That's why AI Agents have been so popular recently <b>就是因为在大语言模型和AI Aagent的配合之下</b> It's precisely because of the collaboration between large language models and AI Agents <b>让未来所谓的通用和强人工智能的可能性</b> The possibility of what we call general and strong artificial intelligence in the future

<b>有了具体的技术路径</b> has a concrete technological pathway <b>而不再是大家的空法想象</b> and is no longer just empty imagination for everyone <b>好</b> Okay <b>讲到这儿</b> Speaking of this <b>我再给大家讲一个问题</b> Let me address another question for everyone <b>回答刚才那个小伙的问题</b> To answer the young man's question from earlier <b>他问</b> He asked <b>为什么这两年才火</b> Why has it only become popular in the last couple of years <b>非常重要</b> Extremely important <b>因为在大语言模型没有出现之前</b> Because before the emergence of large language models <b>人类和机器之间的交流</b> the communication between humans and machines <b>没有一个重要的翻译官角色</b> lacked a crucial translator role

<b>你所有的交流必须要高度的机械化</b> All your communication must be highly mechanized <b>比如说</b> For example <b>你好</b> Hello <b>请帮我开灯</b> Please turn on the light for me <b>对吧</b> Right <b>而不是你回家随口一感叹</b> Instead of you casually exclaiming when you get home <b>哎呦</b> Oh dear <b>这家里边怎么这么闷呢</b> Why is it so stuffy inside the house <b>自动帮你把窗帘打开</b> Automatically opens the curtains for you <b>所以有了这个翻译官</b> So with this translator <b>就不仅仅简单是你开窗帘</b> it's no longer just about opening your curtains <b>或者点外卖这么简单的事了</b> or ordering takeout as simple as that

<b>它很可能在一些复杂的场景里边</b> it could very well be in some complex scenarios <b>就对你的需求有了非常精准和高度分析的能力</b> It possesses a highly precise and analytical capability for your needs <b>所以也就是说</b> So that is to say <b>因为大语言模型的出现</b> Due to the emergence of large language models <b>有了翻译官的能力之后</b> After acquiring the abilities of a translator <b>他的通用性</b> Its versatility <b>AI潜力的通用性</b> The Versatility of AI's Potential <b>得到了大幅的增强</b> Has Been Significantly Enhanced <b>在这个通用性的大幅增强之下</b> With This Substantial Enhancement in Versatility <b>AI的空间才真正被打开</b> The True Potential of AI Has Finally Been Unleashed <b>所以在它出现之前</b> Before it appeared

<b>所谓的带动每个行业重做一遍</b> The so-called "redefining every industry" <b>天方夜谭</b> A pipe dream <b>解决具体问题</b> Solving specific problems <b>天方夜谭 鸿沟巨大</b> A Thousand and One Nights, a vast chasm <b>所以未来真正的机会在哪里</b> So where do the real opportunities lie in the future <b>如果你想问我想做的事情是哪里</b> If you want to ask me what I aspire to do <b>其实就是做一个AI的</b> It's essentially about creating an AI <b>‌Agentic System</b> <b>（自主代理系统）</b> Agentic System (Autonomous Agent System) <b>或者叫Agent System</b> <b>（智能体系统）</b> Or called Agent System (Intelligent Agent System)

<b>然后来针对我们所说的产业互联网</b> Then targeting what we call the industrial internet <b>企业经营 渠道管理</b> Business operations, channel management <b>或者是说上下游所有的这些</b> Or in other words, all these upstream and downstream <b>企业所共同拥有的困境</b> challenges commonly faced by enterprises <b>来通过AI的方式</b> can be addressed through AI <b>大幅的提升企业的运作效率</b> to significantly enhance operational efficiency <b>这就是未来我想做的事</b> This is what I want to do in the future.

<b>当然这是第二件事</b> Of course, this is the second thing.

<b>在AI内容之外的事</b> Things beyond AI content.

<b>所以真的机会在哪里</b> So where is the real opportunity?

<b>是在这儿而不是在这儿</b> It's here, not here <b>所以我认为大部分的消费者</b> So I think most consumers <b>和普通的网友</b> and ordinary netizens <b>其实并不需要理解里边</b> actually don't need to understand what's inside <b>具体细节问题</b> Specific details and issues <b>有了这样一个认知体系</b> With such a cognitive framework <b>大家迅速的明白</b> Everyone quickly understands <b>说有一个人出了一个什么东西</b> That someone has come up with something <b>他是什么</b> What is he <b>他其实就是个Agent</b> He is actually just an Agent <b>然后比如说又出现一个新的概念</b> Then for example, a new concept emerges <b>比如说我们在数据清理里边</b> For instance, in data cleaning <b>又加了一个什么新的技术</b> What new technology has been added again

<b>让数据打标和预处理</b> To make data labeling and preprocessing <b>做的高效</b> More efficient <b>其实你就可以知道</b> Actually, you can already tell <b>原来发生在这儿</b> It happened right here <b>然后突然某一天</b> Then suddenly one day <b>如果是谷歌或者是</b> If it's Google or <b>比如说Apple</b> For example, Apple <b>或者是任何一家公司</b> or any company <b>比如说可能马斯克的公司</b> such as possibly Musk's company <b>推出了一个全新的大模型</b> launched a brand-new large model <b>对吧</b> Right <b>训练成本极低</b> with extremely low training costs <b>它的准确度又高</b> Its accuracy is high <b>OK 那可能就知道了</b> OK, then we might know <b>突破是在这儿</b> The breakthrough is here

<b>所以那我讲到这儿</b> So, that's what I've talked about here <b>我就可以问问大家的问题</b> I can ask everyone's questions <b>讲了这么多</b> After talking so much <b>我想问问大家</b> I want to ask everyone <b>DeepSeek真正带来的创新和突破</b> the real innovation and breakthroughs that DeepSeek has brought <b>是在数据层面呢</b> At the data level <b>还是在模型层面</b> Or at the model level <b>还是在训练层面</b> Or at the training level <b>弹幕飘过一篇字</b> A barrage of text floats by <b>让我看一看</b> Let me take a look <b>在哪里</b> Where is it <b>有没有弹幕</b> Are there any bullet comments <b>其实是在训练维度</b> Actually, it's training dimensions

<b>他把模型的训练的成本</b> He reduced the training cost of the model <b>大幅降低</b> significantly <b>但其实数据和模型的底层原理</b> but in fact, the underlying principles of the data and the model <b>是差不多的</b> are quite similar <b>是差不多的</b> It's about the same <b>所以带来的价值是在这儿</b> So the value it brings is here <b>突破是在这儿</b> The breakthrough is here <b>好吧</b> Alright <b>这是个简单的预览</b> This is a simple preview <b>这边还有很多的基础的概念</b> There are many fundamental concepts here <b>我们又展开讲</b> Let's elaborate further <b>我这里就</b> I'll just <b>具体的视频当中</b> in the specific video <b>我们再说吧</b> let's talk about it later

<b>但是这里边</b> but in here <b>我们其实已经Touch到了</b> we've actually touched upon it <b>已经谈到了非常多的</b> A great deal has already been discussed <b>核心的概念</b> Core concepts <b>比如说</b> For example <b>一些所谓的</b> Some so-called <b>其底层的技术认知的框架</b> The underlying framework of technological understanding <b>初级的应用层</b> Basic application layer <b>不同的模型</b> Different models <b>不同的用法</b> Different usage methods <b>不同的之间的局（限）性</b> Limitations between different approaches <b>它的特点在哪里</b> What are its features <b>以及大预言模型出现了之后</b> And after the emergence of large predictive models <b>所担任的翻译官和通用性的</b> The role of translator and universality <b>这个可能性被打开</b> This possibility has been unlocked <b>才有了</b> Only then

<b>Agent的出现</b> The emergence of Agent <b>通过MCP调用服务</b> <b>（MCP:模型上下文协议）</b> Invoking services through MCP (MCP: Model Context Protocol) <b>所以今天开播之前</b> So before starting the broadcast today <b>我看到了一个</b> I saw one <b>有趣的一个留言</b> an interesting comment <b>有趣的一个留言</b> an interesting comment <b>就是说</b> which means <b>会不会推出MCP的服务</b> Will there be an MCP service?

<b>其实我想告诉大家</b> Actually, I want to tell everyone <b>就是MCP它不是一个服务</b> that MCP is not a service <b>MCP是个接口</b> MCP is an interface <b>接口怎么能成为服务呢</b> How can an interface become a service?

<b>对不对</b> Right?

<b>OK 所以讲到了这个</b> OK, so we've come to this point.

<b>AI的这个逻辑和基本的框架</b> The logic and basic framework of AI.

<b>所以这个</b> So this..

<b>我想以后可能我们</b> I think in the future, perhaps we <b>探讨很多东西的时候</b> when discussing many things <b>可以迅速的归位和总结</b> can quickly get back on track and summarize <b>好吧</b> Alright