AI 会抢你饭碗？焦虑、合作，还是共存？

【转录3/20】【3:43-8:37】 Yeah, I mean, I totally agree with him because AI is everywhere. Not just in the filmmaking. So why do you think it's so scary? Well, just like what we talked about at that gathering, everyone was talking about AI. You know, like how fast AI tools are evolving, like from script writing to editing. It's already, like you said, everywhere. Some people were super hyped about it. Others were a bit more on the fence. But the overall vibe, pretty clear AI is not just some fancy tool on the side anymore. It's really starting to change how we think about the whole creative process. So it's already everywhere, for the filmmaking industry. I bet it's the same for a lot of business as well. Yeah, but I'm still kind of curious. Just now you mentioned actually AI is changing how we think about the whole creative process. So did it happen to you before? Well, it definitely helps me speed things up. I can write scripts faster. I can let AI fine tune what I wrote. Sometimes AI can find better words than I do. But sometimes AI found some really strange words. But it helps, especially I like to use it as an assistant to me. So I can brainstorm with AI. Oh yeah, you mentioned that before. It is like our assistant or our partner. Sometimes we can just share our ideas and brainstorm with them. And sometimes actually they will give us some kind of very good ideas, right? Yeah, sometimes. It sometimes gives us some really good sparks. Yeah, totally. 对啊，就是很多时候，我自己在创作的时候，其实我喜欢让AI跟我一起脑爆，就有点像我们的助手一样，然后这样子可能会有激发更多的灵感，我猜就是我那个朋友的朋友，在戛纳AI电影节上面拿奖的那个人也是一样，据说他花了两年时间，去做这个电影 哇，那所以他其实还是很有前瞻性，他两年前就是在AI还没有特别火的时候，他已经利用AI来帮助他去完成这个写稿，还有拍摄。我不太清楚啊，但是起码的话，他一直跟AI合作，然后最后还拿奖了，这非常厉害 对啊，我看过那个预告片，其实那感觉还是蛮像AI出来的东西，有点就是假假的感觉 等一下，Wing，我想问一下，他那部影片的话，他是以真人的那种模式，还是说动画模式的 他偏真实偏写实的感觉，但是还是那种超写实的感觉，就有点过于，就是还是有点，可能那个高光，或者什么反光位，是有点假假的感觉 哦 但整体已经挺不错的了 对啊，所以他那些角色，那些人物的话，他都是虚拟的，对吧？ 他不是真人的，不是真人的 哇 对啊 那以后演员都会失业了，是吗？

【转录4/20】【8:39-12:27】 很难说，现在估计就是新的模式，就是我们在那个电影人聚会的上面，是我们聊了很多嘛，那有些人就会提出来说，其实我们要改变一种模式去想，就以前可能就我们要做很多casting，就要做很多试镜，那可能现在呢，因为AI可以直接用演员的脸，然后生成一个视频嘛，那将来的趋势可能也会这样子，那这个时候就是看大家怎么去包装自己，包括你的社媒啊，看哪个可以引来更多流量，或者引来更多人去关注，那可能就会挑你，就看这个演员本身，在社媒上面够不够charming，够不够吸引人，就换了一种方式，跟以前casting这种就不太一样 那这样子，岂不是变成了以貌取人？ 也不一定哦，就是，你很难说你在社媒上，你的活跃度，或者你有没有观众缘，其实是，不是完全是因为外貌的，那观众缘这个东西，其实是一门玄学 那也是，面由心生，那些很有观众缘的演员的话，说不定他们其实内在也是有的，那也不一定 对啊，很难说的，但是AI一定会改变我们的思维模式，这是肯定的，我觉得 我也觉得是，其实我个人的话，我有时候很多时候，把那个AI，除了在我的工作上面会跟它有交流或者是沟通，或者是在这个方面应用的话，有时候我也会把他当作是我的情感导师 免费的心理辅导师 对，我就说我最近在经历什么事情，然后你首先你帮我剖析一下为什么我会有这样子的心情，为什么我会有这样子的心态，然后我应该怎么去做，有时候他给我的还是挺好的一个建议 不就是跟那些祷告士是一样的吗，就是神父在那里，只是听你说的东西，然后其实不用给你什么反馈，你只要有人听就好了 那不一定，我就是很喜欢，就是挖它的根本原因，就是说你帮我分析一下，为什么我会有这样子的心态，为什么我会有这样子的感受，然后首先他帮我分析完这个深层原因之后，然后我就说那我该怎么去做 希望他可以帮到你，不要让你误入歧途 Well, sometimes it helps. Good. So, Wing, talking about the AI filmmaking, are you diving into this one now? Yeah, but I'm not totally sold yet. To be honest, writing a solid prompt, that takes me a bunch of tries. It's like playing a weird game of trial and error. Just nudging words around and hoping the AI finally gets what I'm after. For example, I used Gemini's VEO before. I gave it a shot. It's super slick, don't get me wrong. It's super cool. But to actually get something close to what I imagined, I had to keep tweaking the prompt like I was tuning a guitar that just wouldn't stay in tune. Just a little bit, just a little bit. I want to get it. It's very close, but not there yet. Yeah, when I keep tuning that, every time you generate anything, it costs tokens, which means real money. Oh, no. So, yeah, it's not cheap. That's how it earns money, huh? Yep, so if you want to get something serious, potentially good, it just means you need to get ready to burn through your wallet. Wow, sounds like a headache. Yeah, so, you know, if you are trying to generate something on the level of that Cannes award-winning film, AI film, that's not gonna be cheap.

【转录5/20】【12:29-15:38】 You probably burn through a ton of tokens just to get the right look, right feel, or right vibe. It's not going to be cheap. It's not like you just type a prompt and win an award. It's more like type, tweak, pay, repeat. And then you probably can get something close to what's in your head. Okay. You know, just now, when we talked about the AI filmmaking, and I was thinking about one thing. Right now, if they don't need to hire anyone in the films, they don't need to, you know, rent a place to shoot or whatever. I think it saves a lot of money. So, because we have a saying in China, right, like the actors cost two million something per day.二百零八万. So that's why I was like, oh, wow, so AI filmmaking probably saved a lot of money and then probably it also had a really good result. But what you said just now, just like, okay, so it's still very costly. 但是你想想，那怎么都比那种绿幕抠图的演员出来的效果要好，像我们之前聊过的一部片，全程靠抠图，那这时候这个AI多少都会比那个要好 对 那些演员也不便宜 那当然 所以其实就据我所知，就目前的情况来说，如果你要生成像刚才说的戛纳获奖的那个AI电影，其实它是不便宜的。因为我自己用过Gemini 那个VEO,其实它每出一个片段, 你要不断地试,就是看你怎么写那个提示词,其实提示词不太好写,写得好的人其实不太多。 对,我就很明显,就算我不是生成那个视频, 但是有时候我去跟那AI 去对话的时候,我想要他生成 一些东西的话,都会有一点偏差, 所以真的不太好写。 真的是。But anyway, I sort of have come to terms with it, you know, this whole AI thing. Instead of being afraid of it, and I know it's not just a fad, it's the future. So a few months back, I had this epiphany.就是几个月之前, 我突然有一个顿悟。 What's that? Instead of trying to fight the AI wave, you know, like paddling against the surf, I might as well learn to ride it.就是与其跟那个浪潮对抗, 不如我就乘风而去。 Well done. That's very good. Because I think when people come across the AI wave, they were like, okay, I just go with the flow, but you try to learn how to ride it. Well done, huh? Yeah. So what have you done? Remember I mentioned I started a new company? Yeah. It's actually AI-related. Wow, cool. So now, instead of stressing about being replaced by machines or AI, I'm figuring out how to build with them. It's pretty wild, but I'm kind of excited too.

【转录6/20】【15:41-19:08】 Well, what is it about? Wing, I know, is the DeepSeek founder, Liang Wenfeng, your relative? Both of you also have the same surname, Liang. Probably he gave me the epiphany. Like, you know, somehow in the back in the history, we're from the same ancestor. Yeah, same root. So what is it about? So it's about AI agents. You know what is AI agent? I don't know. Why AI needs agents? It's actually called AI智能体, 就可能中文你有听过。 So can you explain more? Let me put it this way. All those big AI models like DeepSeek by Liang Wenfeng, my remote, very, very remote relative. You know, those big models, AI models are like the brain, like our central nervous system.就是有点像我们中枢神经系统。But actually do something like, say, pick up a cup, you need your brain to send signals to your hand through the peripheral nervous system.就是你，你要拿一个杯子的话，其实你就需要你的中枢神经，就是你的大脑传输一个信号给到你的手上的周边神经系统，然后你才能拿起那个杯。 Yes. So AI agent is like the nervous system of our hand. They are the part that actually does things based on what the AI model decides. Just like your hands for grabbing, feet for walking. AI agents can be designed for different purpose. You might have one agent that handles research, crawling through documents and summarizing info. Another might handle task management like scheduling, coordinating stuff across tools, etc. Others are built for customer service like writing emails or even automating creative workflows like code writing. 所以我可以这么理解吗，那个AI智能体的话，就是它会有不同的智能体，不同的功能。就好像我们身体的部位那样子的话，手是用来抓取东西的，脚是用来走路的，然后所以AI它不同智能体的话，也是会有，根据不同的需求，然后来进行不同的设计，那是不是可以理解为，像我们现在市面上面有很多很多的那些AI工具，比如说有些是豆包，豆包专门用来对话，然后Deepseek，专门用来做一些信息的搜索之类的，是这样子吗? Deepseek它本身是一个大模型，豆包应该也是一个大模型，我记得，它其实更多像是基于豆包的一些。比如说，我要小红书生图，那我就基于豆包的生图功能，然后做一些设定，比如说，小红书要求的是3:4的比例，然后我其实小红书的格式，包括那个图片，是有一定的规律的，它就将这种规律写进去这一个Agent智能体里面，让你每生成一张图都符合这个要求 哦，就更加的专业化 对 哦 是的 哇，那所以的话就是，其实一个AI可以做很多的事情，那好像我们人一样，一个人可以做很多事情，但是我们需要分工合作，然后我们需要更加的精细化的一个分工 嗯，对的呢 That's cool. So what's your agent like?

【转录7/20】【19:10-21:59】 Well, what we are building is an AI agent that helps with sales, especially in private spaces like WeChat. You know how most CRM tools don't really fit into those chat-based platforms. CRM is client resources management. 其实就是你的客户关系管理的一个系统，那其实目前大部分的CRM系统，其实不太不会是这种聊天为基础的一种平台。So that's where our agent comes in. It helps with chatting to customers, following up, and just keeping the conversation going in a natural, human-ish way. Wait, Wing. I have a question. Is your AI like the robots in the e-commerce? Like the 客服机器人。I just give them some keywords and they give me the answer. But usually the answer is not related to my question. Stuff like that. That's a great question. That's what we found, and that's how we started this AI agent. The cool part of it is every conversation our AI agent generated is kind of tailor-made. So even if the original command or situation is the same, just like there's some keywords that you send, the way it responds depends on the context. Like what the person said before, how they've interacted in the past. It's not really Q&A replies. Ah, I see. So I think your AI is more advanced than those kinds of robots because it's more customized. Yeah, as a customer, when I use e-commerce, I hate those. Like for example, when is my product delivered, they always send me the exact same context. And I hate that. Yeah, just the general answer. Three to seven days. But you don't know when exactly. Yeah, and then no matter how I ask, they just give me the exact same answers over and over and over. And I hate that. So every time when I talk over e-commerce, I would just keep typing I need a real person, like 人工服务,人工服务,人工服务. Same here. Now I learned. Whenever I open the chat box, okay, I just type 人工服务. Me too. It's really annoying, you know. Yeah. So that's why we build this AI agent. We don't want the sales or after sales service work like a robot. We want some warmth in our conversation. Just a little bit tailor-made conversation.

【转录8/20】【22:00-24:56】 Know the context. At least, please know the context. Yeah. 需要点人味 对，你至少联系一下上下文，包括我们找那些客服，人工服务也是。人工服务回答我之后，我第一句话就“请看聊天历史记录” Yeah. 我真的是觉得现在人工服务，他也是非常的AI化,你知道吗? 我就给他发一个问题,譬如说是售后的问题,然后他就会说“你稍等一下,我先看一下。好的,我现在先分析一下。现在帮您转接一下售后团队。”然后售后团队来了就说“你好,我先看一下上面的问题是什么。”我真的觉得这个沟通非常的低效,然后就完全就很浪费我时间。可能再等了五分钟的时间的话, 然后他都没有回答我问题。That's really troublesome. Yeah, I know. That's why we have this AI agent. Yeah. So tell me more about it. Well, besides the tailor-made conversation, we can do mass messaging to a lot of people, but every message is different, based on the context. For example, if you want to send out, like you, teaching English, there are some keywords you want your students, like 10, 20 students to know today's keyword, what is AI, what is artificial intelligence is, and then you send like, I want the students to know this word and expand it a little bit. When you do mass messaging, you will send exactly the same message to everyone, right? Yes, yes. But through these agents, every message is different. Some kids might already know what is AI, so you probably wouldn't send the message, or just send a message like, this is a word reminding, help you to re- remember the word, or some students don't know this word before, you never mention it, then you would expand the word and give some examples this student might be interested in. So every message can be different. Wow, that sounds very effective. Yeah, we are pretty proud of it. You should, you deserve it. 所以Wing，你现在那个AI的话，其实真的是非常的智能，而且的话，也可以根据客户的需求，然后来制定。 我就想，刚才你在讲的时候，我就想起，什么海王海后的话，经常就会说什么，诶宝宝你吃了吗 诶，现在不是了，可以叫名字了 不会再搞错了 Yeah. Okay, just kidding. And one neat feature of it, we train the AI agents to say I don't know, which is really important. Because if usually when you ask the AI model to give you an answer, it's very likely it wouldn't say I don't know. It won't. Right? So we train the AI agent to say I don't know, especially when this AI agent is supposed to be a sales rep for makeup products.

【转录9/20】【24:59-28:27】 This sales rep shouldn't know how to write code. But for most AI models, they would just help you write code, no matter what your previous question is, or what the previous contact is, it doesn't matter. Whatever you ask, they just do it for you. But Wing, I'm kind of curious why you try to teach AI to say I don't know. Because as we all know, AI is powerful, omnipotent. They know everything. So that's why we try to ask AI everything. So why you try to teach them to say I don't know? It's a chat-based agent. We want the conversation to have some warmth, like human being. So actually help you, if this AI agent is representative of you, you wouldn't want it to know everything, right? This is not real. You want it to represent you. It just helps you to do more conversation. It's not going to replace you. It's just more like you, another you on internet world. So we want this agent to know when to say I don't know, just like us. We don't know everything. We will say I don't know. That makes us feel more real. That's a very interesting insight. 所以AI是很全能的，但有时候因为它太全能，我自己倒觉得有点不太真实，当然我们现在目前的使用场景来说也无所谓，但是如果我们在一个对话里面，其实当你对方是一个known-it-all Walking dictionary Usually this kind of people are not really likeable. 对呀，就是太全能的话，或者什么都知道的话，可能还会跟人产生了距离感 对呀，一般人都不可能知道知道所有事情，专家也只会知道他专家的那个部分，他不可能所有东西都知道 对 你不可能让一个销售去写代码，你也不可能是让一个写代码的人去教你化妆 对 当然有些全能的人，是另外说了，你说一般大众 对，确实是，所以你说，你要教AI说我不知道的话，那确实是会更加有人味，就起码的话，也会更加让我们没有那么大的一个距离感，不然的话，就会让我觉得好像如果对方是什么都知道的话，我就很害怕跟他接触，因为我就很害怕暴露我自己有多蠢 所以我们自己的定位，其实就是想把这个产品变成是不同角色的一个助理或者助手，减轻大家去做繁琐的工作，让我们去做一些更有意义，更需要我们去动脑去做的事情，那些繁琐的事情那就交给AI去做，那我们也可以解放我们双手，解放我们的时间 解放我们大脑，可以去创造，更加有意义的事情 是啊。That's the whole point. Hopefully. That's our hope. So to do that, you know, one of the biggest things for us is to kill the hallucination in AI. That's super important. Hallucination? You mean 幻觉? 嗯，是的，就是会乱说东西、乱编东西，就是AI有时候会乱编东西 Yeah, I heard about that. Tell me more about it. How did you hear about that?

【转录10/20】【28:29-31:28】 Sometimes I ask them to give me the links so that I don't need to search every page or every website, every video by myself. I hope it can save me time, right? Then it sent me the links. I was very happy. Oh, wow! AI is very intelligent. 404. Probably not. 就是我当时去叫他，我说我需要这些知识，我需要这些信息，然后你可不可以把那个链接发我，我想说，那我就可以省很多时间，我就直接看那些链接，我就可以知道，他在讲什么东西，对吧？然后他就给我发了链接，但事实上，我每一次打开的话，他就说页面不存在，或者是页面已过期，反正就是404 Yep，那就是幻觉，那叫做AI幻觉，AI hallucination Okay, just try to fool me. I know. Well, imagine if that happens in a customer service. If you want to return something, it says you can return it, you know, just give you a random date like 30 days. In fact, you can only return it within 7 days. That would cause huge problems for the clients and for your business. You cannot do that, right? Yes. That could be really problematic. Like before, another famous example is AI image generator makes weird stuff like hands with extra fingers. It's pretty creepy. Oh, it happens. I remember last year when AI was so popular like everybody talked about AI and sometimes when we're trying to identify if the image is generated by AI, we just count the fingers. AI couldn't create a person with 5 fingers, with 10 fingers. Yep, it's really creepy. It is. Because AI doesn't really understand what a hand is, it can only fake it. Basically, I can use a Chinese term to describe how AI learn about what a hand is. It's 填鸭式学习, 其实你硬塞给他很多资料，然后他说，哦，这就是手，但是很多时候可能就是，你知道我们准备高考，经历过这个填鸭式学习，不代表我们所学的所有东西可以消化得了 It's just root learning. I just teach you the tricks, and then you learn it. And after the exam, then you will leave everything behind. Probably just don't remember it at all or something like that. So Wing, what causes these hallucinations? Based on what I learned, there are four main things. So here comes a lot of jargons. A lot of terms and jargons. So number one is training data limitations. AI models are trained on vast data sets.

【转录11/20】【31:29-35:19】 But if the data is biased, incomplete, or inaccurate, the AI can produce flawed outputs. It's like the internet has everything on it, but it doesn't mean everything is correct. A classic example of this is how AI used to be really good at telling the difference between a husky and a wolf. 就是之前有一个比较经典的案例，关于这一个数据局限性，就是AI很擅长去区分哈士奇还有狼，其实这些人都分不清楚这两种动物 对啊，然后，所以AI是怎么做? Well, it turns out most of the pictures of wolves are coming with snow. AI was linking the white snow in the image with the idea of wolf. So later it was discovered that AI would mistook something else with snow for wolf. 所以AI的话，它会对狼跟哈士奇进行一个辨别的话，就是根据背景有没有雪，因为它AI很多时候会把雪跟狼联系在一起，就有雪出现的地方的话，就是狼。对，因为训练数据，有时候你很难说可能这堆数据里面，大部分我要训练AI去区分哈士奇跟狼的时候，刚好狼的照片里面都有雪，然后AI就误以为有雪的都叫狼，然后它就把这个做了一个关联。那我们人是不知道的嘛，那一开始我拿这堆数据去训练它，我没有察觉到狼的图片里面都是雪，对啊，那对于我来说，可能我根本就不知道，谁知道之后在试一些其他的数据去验证它的时候，就突然发现怎么有雪的地方都叫狼，所以就出现这种问题 Makes sense. Well, actually, it reminds me of something. Do you know there is a very famous website called Reddit? So it's like BBS, right? So many people leave the comments or post something on that platform. And that's why the AI company wants to get the data from that website. However, actually, in that platform, many people post, I cannot say many people, like some people post something prejudiced, racist, or stuff like that. So sometimes the result comes out as inaccurate if the AI tries to learn from that. Yeah, the internet is not always right. Yeah. 所以我刚才举了一个例子，就我之前有了解到的国外有一个网站就叫Reddit,然后它是非常有名嘛,它就好像我们之前的那些BBS那样子,很多人在上面发帖呀,很多人在上面评论。就其实很多时候的话,它都可以帮助我们解决很多问题。但事实上,它有一些人在上面,因为我们没有办法控制每个人他的发言,所以有些人他可能发言的会比较有偏见, 可能会带有种族歧视之类的。所以如果AI去学习这一部分不太好的数据的话,那最后它可能出来的结果就是会有偏差,然后就不太正确。 对啊,我最近听了一个博客,他就说到说那个Claude又出新模型嘛,就是一个大模型公司。然后不同版本呢,它的对于,我忘了是关于哪个问题,的针对性是不一样的。有一个模型是特别特别注重动物的权益。估计就是它训练数据的问题,就它用了不同的数据去进行训练。结果出来的结果就有点不一样。 Oh, very interesting huh? Yea, so the second question is ambiguity and context. You know, Language is messy and full of double meanings. It doesn’t matter if it’s Chinese or English.

【转录12/20】【35:22-39:09】 It caused misunderstanding. If AI doesn't fully get the context, it might totally mess up a response. Like that classic joke in advertisement. A client would say, 'I want colorful black.' Humans can’t even figure that out, how AI knows what to do? Yeah. I wonder how.就是本来它语言,就不管是哪一门语言的话,它都会有多重含义。就好像我们那个什么英语听力还有什么中文听力,就网上经常讲的那些。 段子。 对,那些段子。然后它都是一个单词或者一个短语的话,它都有多重的含义嘛。所以如果AI它不能完全理解上下文的话,它可能就会完全就会搞错。 对啊,就是广告圈之前最出名的那个笑话,就是客户说我要五彩斑斓的黑。所以这个我就觉得是个很经典的例子。你这样说这种,就不管是哪一种语言,哪一种场景,其实都不太makesense。 对啊,那其实是什么意思我都还没有搞懂。 就是可能他觉得黑不是他想要的东西,就是一种感觉。就很多时候我们都说一种感觉。当我们在给AI输入一些提示词的时候,如果是说一种感觉,AI可能很有可能就无法理解了。就过于感性的东西。 对,因为有时候就好像说,我叫AI帮我生成一段话的时候呢,它会非常的,就它逻辑性非常强嘛,但是呢它又非常的冷冰冰,然后又说请你带点人味。但是呢我又觉得它可能生成出来的效果,反正永远都不是我想要的,就是它就没有什么人味。它就没有办法理解我这个人味是什么意思。 就这个时候你要说得更具体一些,单纯说人味是不够的。 对对。 也就是说我喜欢用哪些词,或者是我请参考某一段文字的风格之类的,那它会更容易去理解。 对啊。好,学习了,下次就这样去做。 Okay, so what's number three? Well, the number three is lack of grounding. Sometimes AI doesn't have access to real-time or accurate info, so it just makes stuff up to fill the gaps, like when an essay footnote cites random sources that didn't even exist. Yeah, just exactly what I mentioned just now, I asked them for some links, they gave me the links already expired, or probably like never existed. Yep. 所以就有时候AI的话,它没有实时或是准确的信息。因为我记得好像有一些模型的话,它的数据来源已经是可能是一年前或是两年前,对吧?然后所以它有时候它就会,我们问它一些实时的问题的时候呢,它就会凭空编造一些东西,然后来填补这个空白,就好像it knows everything. Yep. 就好像我刚才有举过一个例子,就说我去叫它帮我去找一些网页,它给我那些网页全都已经不存在了,我都不知道为什么不存在它还能找得到。That's ridiculous. Yep, that's called making stuff up. 一本正经在胡说八道。 Yeah. They did a very good job, huh? Yeah. Although a lot of models right now can have internet search with it, but still I try it, sometimes it still make up something. Yeah. Sometimes. Okay, so what's next? The last one is overfitting, which means if AI gets trained on specific data too much, it starts to stick to patterns that might not always apply. Let's say it's trained to recognize faces with two eyes, one nose, and one mouth.