| {"base_path":"examples","text": "[S1]诶，我最近看了一篇讲人工智能的文章，还挺有意思的，想跟你聊聊。[S2]哦？是吗，关于啥的啊？又是哪个公司发了什么逆天的新模型吗？[S1]那倒不是，是一个咱们国内的教授，复旦大学的邱锡鹏教授，他提了一个新概念，叫什么，呃，叫情境扩展，Context Scaling。[S2]Context Scaling？情境扩展？听起来有点，呃，有点玄乎啊，这是个啥意思？[S1]对，我一开始也觉得有点抽象，但你看完就觉得，诶，特别有道理。他大概意思就是说啊，咱们现在对人工智能的追求，不能光是把它做得更大，你知道吧，就是不能光堆参数，喂数据。[S2]嗯，是，这个我懂。就好像之前大家都在比谁的模型参数多，几千亿，上万亿的。[S1]对对对，就是那个意思。他说那个时代，算是第一幕，就是模型规模化的胜利，靠堆料，堆出了像这个ChatGPT这样厉害的通用模型。[S2]嗯，是的。[S1]然后呢，现在差不多是第二幕，就是大家发现光堆料好像不行了，收益越来越小，就开始搞一些，呃，后训练的优化。[S2]哦，后训练优化？比如呢？[S1]比如让AI学会用工具，或者搞那个什么思维链，让它像人一样一步一步思考问题，还有就是强化学习，让它自己在游戏里或者写代码的时候自己跟自己玩，然后越变越强。[S2]哦，明白了。就是让它不光是知识多，还得会用，变得更聪明，是这个意思吧。[S1]没错，就是这个理儿。但是呢，这两步走完了，就遇到了新的瓶颈。邱教授就觉得，关键问题出在了这个情境，也就是Context上。[S2]嗯？情境？[S1]是。很多时候AI做不出正确的决定，不是因为它笨，而是因为它没搞明白现在到底是个什么情况，就是我们给它的任务或者情境描述得不够清楚。[S2]哦，原来是这样。[S1]对。所以他觉得下一幕，也就是第三幕，就应该是这个情境扩展，Context Scaling。核心就是要让AI能真正理解那种特别复杂、多变，甚至有点模糊的真实世界的情境。[S2]嗯，这个听起来就高级了。那他说的这个情境，到底指什么啊？不就是我们输进去的那段话吗？[S1]诶，不是那么简单的。他说的情境啊，是个特别立体的东西。它不光包括你说了什么，还包括，呃，时间，地点，谁在说话，目的是什么，甚至还包括咱们文化里那种只可意会不可言传的规则，还有人与人之间的那种默契。[S2]哇，那这个范围可太广了。[S1]是吧。他举了个例子，就比如说，当一个人说不要的时候。[S2]嗯。[S1]你想想，在不同的情况下，这个不要的意思完全不一样。有可能是真的拒绝，也可能是在开玩笑，甚至可能是一种反向的请求，就是我们说的口是心非，嘴上说不要，身体很诚实那种。[S2]哈哈，确实确实，这个AI可怎么判断啊。[S1]对啊，所以就需要理解整个情境。他说这个里面最关键的，就是要捕获一种叫暗知识的东西。[S2]暗知识？听着像武林秘籍一样。[S1]有点那个意思。就是指我们人类会，但是很难用语言清楚地讲出来的那些能力。[S2]哦？比如说？[S1]比如说社交的智慧啊，怎么通过一个眼神，一个停顿，或者语气变化来理解对方的意思。[S2]嗯，是的。[S1]还有就是文化适应能力，在不同的文化里，有些事能做，有些事不能做，这些都没人写在书上，但我们就是知道。[S2]没错。[S1]他说AI要是能学会这些，那才算是真正的智能突破了。这也能解决一些AI安全的问题，比如那个很有名的回形针悖论。[S2]哦，那个我知道，就是你让AI造回形针，它最后可能会为了这个目标把整个世界都给占了，用来造回形针。[S1]对。但如果它有情境智能，它就能理解我们人类社会的复杂性，知道有些事是不能做的，就算你没有明确下指令禁止它。[S2]有道理，有道理。那，那技术上要怎么实现呢？听起来好难啊。[S1]是挺难的。他提了三大技术支柱。第一个叫强交互性。就是AI要不停地跟环境，特别是跟人来互动学习，不光要知道怎么做，还要知道为什么这么做。[S2]嗯，在互动里学习。[S1]第二个叫具身性。就是AI得有个身体，不一定是真的人形机器人，在虚拟世界里也行，总之它得能感知和行动。[S2]哦，明白。[S1]第三个，我觉得是最特别的，叫拟人化。[S2]拟人化？[S1]对，就是说AI需要有类似人类的情感共鸣能力。不是假装有感情，而是真正理解人的情绪，人的偏好，懂得社交的距离感，什么时候该关心，什么时候该保持沉默。[S2]哇，这个要求可太高了，感觉比养个孩子还难。[S1]哈哈，可不是嘛。所以他说这事儿吧，不是要去替代其他的技术路线，而是把它们都整合起来。像什么推理增强啊，多模态啊，强化学习啊，最终都是为了服务于一个目标，就是让AI能深刻地理解情境。[S2]嗯。[S1]所以总的来说，就是别光在已有的路上卷了，要去定义一些大家都觉得重要，但没人说清楚的问题。[S2]确实。听你这么一说，感觉这个情境扩展，Context Scaling，确实给人工智能指了一个挺不一样的方向。不再是冷冰冰的计算，而是要变得更懂我们人类，更懂这个复杂的世界。[S1]是啊，他说这可能是通往通用人工智能，就是AGI，非常关键的一步。[S2]嗯，听起来未来可期啊。","prompt_audio_speaker1":"pod_f_enhanced.wav","prompt_text_speaker1": "然后呢，现在的社会现实就是，我们作为寿命更长的群体，我们的收入更少，这个是我们必须要面对的一个基本的事实。女性的平均寿命通常是比男性要长几年，但是与此同时，女性的收入却往往比男性低。","prompt_audio_speaker2":"pod_m_enhanced.wav","prompt_text_speaker2": "下面这个故事啊，里边没有灵异，但是是我最喜欢的那种故事，就是如果你在听这个故事的时候，尤其是故事高潮时候那一句话，你如果能在脑海里形成一个画面，那么可能今晚啊，你就指着这个画面熬夜了。"} | |
| {"base_path":"examples","text": "[S1]Hey, did you hear about that company called MoSi AI? [S2]MoSi AI? Yeah, I think I've heard of them. Aren't they the ones doing AI stuff? What new thing have they come up with now? [S1]Yeah, that's them! They recently launched this super hot new product called, um, Asteroid. [S2]Asteroid. That's a pretty cool name. Does it mean like the space rock? [S1]Yeah, I think that's what it means. Let me tell you, this thing is incredible. They say it's currently the most realistic, human-like conversational TTS model out there. [S2]Oh, TTS technology? You mean the text-to-speech thing? Aren't there already a lot of those on the market? What makes this one so special? [S1]Well, it's completely different. They say the voice produced by Asteroid sounds almost exactly like a real person talking. And it's super smooth and natural. Not at all like, you know, that stiff robotic tone. [S2]I see. Some voice assistants do still have that mechanical feel, especially during multi-turn conversations. So how amazing is this Asteroid exactly? [S1]I heard they internally call Asteroid China's own version of NotebookLM. [S2]NotebookLM? Oh, I know that one. Isn't that the personal AI that Google made? The one that helps organize notes and answers all kinds of questions? So Asteroid has similar functions? [S1]Right. That's probably what they mean. It's not just that the voice sounds incredibly human. The intelligence level is also really high. It can have these really logical, contextual, in-depth conversations with you. It's just like chatting with a real person. [S2]Wow, that sounds amazing. If they can really achieve that... [S1]Yeah, it's basically like having a personal assistant that's both articulate and really understands you. [S2]Hmm. That does sound appealing. [S1]And some people are saying it's like the, what's it called again in the voice technology circle? Oh right, DeepSeek. [S2]DeepSeek? Isn't that the company making large language models? Their models are pretty popular now. That's high praise. So they're saying Asteroid is top-tier technology? [S1]Yeah, I think that's what they mean. It's like they've reached a whole new level in voice synthesis. Similar to the impact DeepSeek has had in natural language processing. It could be that kind of groundbreaking technology. [S2]If Asteroid is really that impressive, where could it be used? I feel like there must be huge potential there. [S1]Absolutely. Just imagine future smart customer service, audiobook reading, and those virtual livestreamers that are so popular now. The quality would improve dramatically. We might even have personal assistants using Asteroid to talk to us directly. How natural would that be? [S2]Yeah. That does sound exciting. When can we actually try it out? Are there any demos available? [S1]I haven't looked into that carefully yet. But since they've already announced it, I'm guessing it won't be long. I'm really eager to try it and see just how human-like it is. [S2]Yeah, yeah. If it can really deliver what they're promising, getting information and interacting with machines will be so much more convenient. The experience will be much better too. [S1]Exactly, exactly. We're just waiting for MoSi AI to give us this big surprise.", "prompt_audio_speaker1":"m1.wav","prompt_text_speaker1": "How much do you know about her?","prompt_audio_speaker2":"m2.wav","prompt_text_speaker2": "Well, we know this much about her. You've been with her constantly since the first day you met her. And we followed you while you went dining, dancing, and sailing. And last night, I happened to be there when you were having dinner with her at Le Petit Tableau."} |