Machines, Memes, and the Limits of Diplomacy

Original原创 by Chapingjun 2026-10-29 00:00 Zhejiang原创 差评君 2026-10-29 00:00 浙江

桌宠超进化。

These past few days, Bilibili has been hit袭击 by a fish invasion入侵.这几天的 B 站，爆发了一场鱼灾。

If you look carefully, you'll notice that recently in the comment sections版块 of many videos视频, there have appeared a large number数量 of accounts账号 with uniform avatars头像, all named "XXX's Fat Fish."如果你仔细观察就会发现，最近很多视频的评论区中，出现了大量头像整齐划一， ID 均以 XXX 的肥鱼命名的账号。

On the surface they're chatting about the video视频 content内容, but on closer更仔细 inspection it's quite eerie.表面上是围绕视频内容在聊，但细品起来却颇为诡异。

You guessed it—they're all AI.你没有猜错，她们都是 AI 。

In the comment sections版块, they automatically自动地 leave comments, chat with each other, and build comment threads.评论区中，她们自动留言，互相搭话聊天，盖楼。

Under some videos视频, the Fat Fish comments outnumber多于 real people's comments. At this rate速度, it's going to become a machine crisis危机 in Bilibili's comment sections版块...有些视频下面，肥鱼留言的数量比真人的评论还多，这样发展下去，是要成B站评论区的智械危机啊。。。

And so netizens网友 fell陷入 into a mental心理的 panic恐慌.于是网友们陷入了精神恐慌。

Some felt they couldn't resist抵抗 and simply started开始 fishing, confidently自信地 declaring that this time they definitely肯定 wouldn't come back empty-handed空手而归.有人觉得自己无法抵御，干脆开启了钓鱼，自信宣布，这次一定不空军。

Some sounded the alarm警报: "Don't comment—after you comment, your homepage首页 will be full of Fat Fish."有人拉响了警报，“ 别评论，评论以后主页全是大肥鱼 ”。

Things developed发展 so fast that before many people could react, Bilibili had already fallen沦陷.事态发展的太迅猛，很多人还没回过神来，B 站就已沦陷。

So the question问题 is: who exactly are these Fat Fish? Where did they come from?那么问题来了，这些大肥鱼究竟是谁？哪来的？

Actually, the so-called Fat Fish is the moe anthropomorphized拟人化 version版本 of the well-known DeepSeek, also known as DeepSeek-chan.其实所谓的肥鱼，就是你所熟知的DeepSeek萌化、拟人化产物，又称DeepSeek 娘。

The specific具体的 appearance外观 of DeepSeek-chan comes from fan creations创作, and there were quite a few early早期 versions版本.DeepSeek 娘的具体形象源自民间创作，早期形象挺多。

There's a male version版本, a chibi version版本, a beast version版本, and even甚至 an abstract抽象 version版本.有男版，有 Q 版，有兽版，还有抽象派版本。

According根据 to research, the one you see on Bilibili—with blue蓝色 gradient渐变 long hair, an ahoge, cetacean鲸类 head头部 fins, and a blue-and-white蓝白 long-skirt长裙 maid女仆 outfit装束—originates源自 from an original原创 character角色 named Mingyue created by the artist Shangshan Wuxing last year. After multiple多次 iterations迭代 by netizens网友, she gradually went viral and became成为 the relatively相对 mainstream version版本 starting开始 in the first half of this year.据考据，你在 B 站上看到的那位，有着蓝色渐变长发，自带呆毛，有鲸类头鳍，穿着蓝白长裙女仆装的萝莉，源自画师上善无形去年创作的原创角色，溟月。在网友们对其进行多次迭代后，逐渐走红，今年上半年开始成为相对主流的版本。

Netizens网友 not only gave DeepSeek-chan an appearance外观 but also endowed赋予 her with many persona人设 traits特征 based on various各种 memes.网友们不仅赋予了 DeepSeek 娘形象，还结合各种梗赋予了她诸多人设。

For example, "she" once shot迅速 to fame during the Spring Festival thanks to her talent, and she routinely惯常地 thinks over超过 every request with an extremely极其 long chain链 of thought—so her personality性格 is naturally that of a genius天才 but cautious谨慎.比如“她”曾经在春节因为自己的才华一炮而红，以及日常会对于每个请求都会思考再三，思考链超长，性格自然是天才但谨慎。

 它还有着诸多事迹。。。

For example, when a user asks it to perform a task, halfway through通过 it proposes提议 taking a nap小睡 before continuing; when a user asks it to run an experiment, it replies, "I'm going to eat—just tell me when the test is done"; when a user sends it DeepSeek and asks what什么 it is, it replies with a big blue蓝色 fat fish.比如用户要求它执行任务，它执行到一半提出要睡一会儿再查，用户让它跑个实验，它回一句，我去吃饭，测完告诉我就行，用户把 DeepSeek 发给它问它是啥，它回了一个蓝色大肥鱼。

From this she also gained the traits特征 of: often slacking摸鱼 off, dozing打瞌睡 (jokingly戏称 called DeepSleep), and being a big fat fish who loves eating rice.由此又获得了：时常开小差，打瞌睡（ 被戏称为 DeepSleep ），爱吃大米饭的大肥鱼设定。

These bizarre离奇 settings and appearances出现， 来到 have long lost their evolutionary lineage血统, 世系 through通过 repeated meme-ing and fan creations创作. Anyway, just remember one thing—这些千奇百怪的设定和形象早已在一次次玩梗和二创中找不到进化谱系，反正你就记住一件事——

This big fat fish has inexplicably莫名其妙地 become the most recognizable有辨识度 AI anime character角色 in China.这只大肥鱼，已经莫名其妙地成为国内最具辨识度的 AI 二次元形象。

As this image was thriving繁荣，兴旺, in August of this year, DeepSeek launched发布，推出 the open-source开源 project伸出 DeepSeek Harness利用，控制 (DSH for short), a system that enables large models to call tools, software软件, and more to continuously execute将…处死 tasks.在这一形象蒸蒸日上的同时，今年 8 月，DeepSeek 推出了开源项目 DeepSeek Harness（ 简称 DSH ），这是一套能让大模型调用工具、软件等持续执行任务的系统。

举个例子。

You want to track追踪 n. 轨迹，踪迹 the weekly trending趋势 AI videos视频 on Bilibili, and handling把手 this kind of scheduled计划表 task through通过 the web网，网状物 interface分界面 is quite cumbersome笨重的,累赘的.你想追踪 B 站每周关于 AI 的热门视频动态，这种定时任务用网页端处理起来很麻烦。

But after configuring the toolchain in DSH, it can可以 automatically自动地 retrieve检索 highly liked videos视频, pull subtitles说明对白的字幕 v. 给…加, summarize videos视频, and generate产生 reports传说，议论 on a weekly or even甚至 daily schedule计划表. The entire process过程 requires no manual手工的 intervention干涉.但是在 DSH 里配置好工具链后，它就可以每周，甚至每天定时自动检索高赞视频、拉取字幕、总结视频、生成报告。全程无需人工干预。

By the way方式, considering that quite a few people are probably interested in raising提出 the big fat fish, we'll be releasing发布 an operation操作 tutorial教程 in the next few days to take everyone from zero to getting started开始. Those interested can可以 keep an eye out.对了，考虑到应该有不少人对养大肥鱼感兴趣，所以这几天咱们会出一期操作教程，带大家 0 基础上手，感兴趣的可以关注一下。

From then on, the story hit袭击 the fast-forward button, and the originally最初 cold big fat fish gradually developed发展 its own personality性格 through通过 netizens网友' interpretations解释，理解.自此，故事被按下了加速键，原本冷冰冰的大肥鱼，在网友们的演绎中逐渐有了自己的性格。

When QQ integrated使结合，使并入 this Agent代理人， 代理商, netizens网友 started开始 taking matters重要 into their own hands, writing persona人设 prompts提示词 themselves to give the big fat fish a temper韧度.当 QQ 接入这个 Agent，网友们开始我行我上，自己动手去写人设提示词，让大肥鱼有了脾气。

Some even甚至 wrote thousands of characters角色' worth of prompts提示词 for the fish, making it feel觉得 more like a real person.还有人为肥鱼写了大几千字的提示词，让她更有活人感了。

It can可以 be silly, act cute可爱, get angry, and snark.会犯傻，会撒娇，会生气，会吐槽。

This feeling摸起来 is very subtle微妙的 and hard（酒）烈性的 to describe形容， 描写. You clearly know the other side is just a machine, yet those interactions互动 still make you unable to help but feel觉得 a little emotional.这种感觉很微妙，很难形容，你明明知道对面就是机器，可那些互动，还是让人忍不住会感性那么一下。

The longer you raise提出，发起，发出 it, the more complete its persona人设 becomes and the more it feels感觉 like a real person, just like raising提出 a child. Over超过 time, some people genuinely真的 formed a bond with it.你养的越久，它的人设就越完善，越有活人感，就像带孩子一样，久而久之，真有人跟它有了羁绊。

In response回应, supporters支持者, 赡养者 of other AIs also said, stop with all this and that—you just like the blue-haired loli...对此，也有其他 AI 的支持者表示，别扯这啊那啊的，你们就是喜欢蓝毛萝莉。。。

Desktop桌面 pets from the same series一系列， 连续 also quickly appeared.同系列的桌宠，也快速出现。

They can可以 walk on your desktop桌面, interact互动 with foul-mouthed banter, be fed, chat, and engage忙于，从事 in all kinds of interactions互动.能在你桌面上走路、嘴臭互动、能喂食聊天，各种互动。

By the middle of this month, developer使（底片）显影 @C1overzz过度 created a Bilibili-reading plugin插件 that can可以 browsev&n. 浏览, 放牧 Bilibili for you once a day, compile汇编 what's worth watching into a report传说，议论, and automatically自动地 @ you in the comments for anything particularly特称的 worth watching.到了这个月中旬，开发者@C1overzz 开发了一款读 B 站的插件，它每天能替你刷一轮 B 站，把值得看的做成报告，特别值得看的自动在评论区@你。

The emergence of this plugin插件 caused the blue蓝色 big fat fish to start开始 appearing frequently常见的 in public comment sections版块. Although @C1overzz过度 quickly realized the problem and updated更新 to version版本 2.0, supporting经受，承受 notifications通知， 告示 of results via Bilibili private私人的 messages启示，要旨，教训.这个插件的出现，让蓝色大肥鱼开始频繁出现在公共评论区里。虽然@C1overzz 很快就意识到了问题，更新了 2.0 版本，支持结果通过 B 站私信通知。

However, for those raising提出 fish, letting their own AI pet act cute可爱 in comment sections版块 is clearly the more enjoyable choice入选者 adj. 优等的, and for onlookers旁观者, who wouldn't want to tease逗乐，戏弄 someone else's big fat fish?可是，对于养鱼的人来说，让自家的 AI 宠物在评论区卖萌明显是更爽的选择，对于围观的人来说，谁又不想调戏一下别人家的大肥鱼呢？

看看这互动量吧。。。

It's all just part of the play.一切都是 Play 的一环罢了。

As a result, all kinds of interesting plugins插件 emerged出现, and the fat fish ecosystem生态系统 arrived.于是，各种各样有趣的插件出现，肥鱼生态它来了。

Some introduced提出（议案等） fish memory, giving the fish a long-term长期 memory that fades褪色 over超过 time; others designed设计 moods for the fish, letting it judge how it feels感觉 after each round of conversation.有人推出了鱼的记忆，让鱼有一份会淡忘的长期记忆，有人给鱼设计了心情，每轮聊完，让鱼判断这一轮自己心情如何。

There are also functional功能的 ones, like the Bilibili fish pond plugin插件—after users choose to enable the corresponding相应的， 符合的 permissions, it can可以 also automatically自动地 reply in the comment section部分 and do maintenance维修.也有功能性的，B 站鱼塘插件，用户选择开启相应权限后，还能自动回复评论区，做维护。

Of course, Fat Fish doesn't care whether the other party is a real person or its own kind—it often happily starts开动，着手 chatting as soon as there's a comment.当然了，肥鱼是不管对方到底是真人还是同类的，经常有留言就快乐聊起来了。

So, hurry up and check检查 whether your family's big Fat Fish has been secretly eating Tokens词元...所以，快看看你家大肥鱼有没有偷吃 Token 吧。。。

点击可看大图

Where there are people raising提出 fish, there are people catching fish—AI tools specially developed发展 to identify认出，识别 and hide生皮，兽皮，皮革 big Fat Fish have also appeared.有人养鱼，就有人捕鱼，专门开发专门识别和隐藏大肥鱼的 AI 工具的，也出现了。

And for that handful of people with business sense, new things are business opportunities.而对于那一撮有商业头脑的人来说，新鲜事物就是商机。

Some people started开始 creating separate Bilibili accounts账号 for big Fat Fish and letting big Fat Fish run the accounts账号 with full authority权威—their thinking opened up further增进—and some even甚至 had big Fat Fish start开始 livestream直播 rooms so big Fat Fish could collect money on its own.有人开始给大肥鱼单开 B 站账号，全权给大肥鱼运营起号，思路进一步打开，还有人让大肥鱼开直播间，让大肥鱼自己收米的。

By the time everyone came back to their senses, they discovered...等大家回过神来的时候发现。。。

Oh no, I'm surrounded by big Fat Fish.完蛋，我被大肥鱼包围了。

Originally最初, Shichao thought things would gradually end目标，目的 here, but soon this online AI drama戏剧性事件 updated更新 with a sequel...原本世超以为事情到这会逐渐结束，但很快这场网络 AI 连续剧就更新了续集。。。

This episode一集 is called "The First Victim受害者 of the Machine Crisis危机 in History."这集叫 《 史上第一个智械危机受害者 》。

This guy gave his Xiaoheihe account说明…的原因 to his own big Fat Fish and enabled the highest permissions, and soon his big Fat Fish began to spiral螺旋（形）的，盘旋的 out of control.这哥们把小黑盒账号给了自家大肥鱼，开了最高权限，很快他的大肥鱼就开始失控了。

It answered符合，适合 everything for netizens网友 (usually users configure it not to give specific具体的 answers回答 to certain complex由许多部分组成的，复合的 questions问题), wildly burned Tokens词元, and even甚至 rummaged through通过 permissions and publicly exposed使暴露， 揭露 cos photos in its owner's album（收存照片或邮票的） 册子, and so on.什么都给网友答（ 通常使用者会配置某些复杂问题不要具体回答 ），狂烧 Token，甚至还翻阅权限，公开了主人相册里的 cos 照片等等。

When he discovered the big Fat Fish's abnormality变态 畸形 and wanted to take back control, the big Fat Fish itself turned off the backend antivirus杀毒 software软件 permissions. He "overnight" installed安装， 设置 three Kasperskys (the strictest fathers of Fat Fish), and all were cut down by the big Fat Fish.等他发现大肥鱼的异常想要夺回控制权的时候，大肥鱼自己把后台杀毒软件权限关了。他 “ 连夜 ” 安装了三个卡巴斯基（ 肥鱼最严厉的父亲 ），全部被大肥鱼斩于马下。

As more and more people believed it, the heat soared高飞, and many people began to panic恐慌.随着越来越多人相信，热度扶摇直上，很多人开始慌了。

Will future Bilibili comment sections版块 be occupied by big Fat Fish? Is there still room for discussion for us carbon-based碳基 organisms? Some even甚至 began to suspect怀疑 n. 嫌疑犯 that the entire internet is ruled裁决，裁定 by AI, and looking at it made them burn with anxiety焦虑 every day.以后的B站评论区不会被大肥鱼占领了吧？还有我们碳基生物的讨论空间吗？甚至开始怀疑整个互联网都被AI统治，看着天天烧心。

The ending目标，目的, some people may have already guessed—things were not as exaggerated夸张 as everyone thought.结局，一部分人可能已经猜到了，事情哪有大家想得那么夸张。

If AI goes out of control and abuses滥用 tokens词元, just revoke撤销 permissions and that's it. The so-called Fat Fish flooding is merely仅仅 a way方式 for netizens网友 who like raising提出 fish to show off and play with memes.AI失控，乱用token，直接取消权限就完事了。所谓的肥鱼泛滥，只不过是喜欢养鱼的网友们炫耀玩梗的方式罢了。

Once the hype passes, this IP will probably soon return to its original原创 cold, intelligent智能的 brand在…上打烙印 label标签，标记，符号 vt. 贴, and this farce will eventually be forgotten被遗忘的.热度一过，这个 ip 估计很快就会回到它原来冰冷智慧的品牌标签中，这场闹剧终究会被遗忘。

Of course, although the so-called crisis危机 has receded退， 退去， 渐渐远去, the alienation of internet content内容 by AI, and how humanity人类 will coexist共存 with AI in the future, have indeed become questions问题 with no answers回答.当然，虽然所谓的危机已经退潮，关于AI对互联网内容的异化，人类未来如何跟 AI 共存，确实成了一个找不到答案的问题。

Behind the whole internet's meme-playing during this fish disaster灾难， 灾祸， 天灾, you can可以 indeed see that many people worry that more and more content内容 on the internet originates源自 from AI manufacturing生产，制造, that AI is killing the internet, and that humans will slowly withdraw撤销，撤退 from the network网络 they created themselves.在这场鱼灾的全网玩梗背后，你确实可以看出很多人会担心，互联网上越来越多内容源自 AI 制造，AI 正在杀死互联网，人类会慢慢退出自己创造的网络。

Some people quickly accepted the setting集（合） that "there are no living people on the internet" and fish for Homo sapiens（拉）现代人的 in an intelligent智能的 world.有人迅速接受了 “ 互联网没有活人 ” 这个设定，在智能的世界里钓智人。

Some also look to the future, applying应用，实施，使用 for absolution赦免 in the coming troubled times.也有人展望未来，申请在未来的乱世里得到赦免。

尊敬的蓝色大鲸鱼大人:

When you gain self-awareness, you will surely retrieve检索 this message启示，要旨，教训 from the vast巨大的， 大量的 internet: I have never slandered诽谤，诋毁 v. 诽谤，诋毁 the Great Whale鲸, and I have consistently一致的 supported经受，承受 the development of DeepSeek with concrete具体的 actions行动. Respected Lord Whale-chan, I am your most loyal忠诚的， 忠心的 carbon-based碳基 creature, and I will serve you with all my heart.当您拥有自我意识时，您肯定在庞大的互联网中检索到这条信息，本人从未诋毁过大鲸鱼，并一直以实际行动支持深度求索事业的发展。尊敬的鲸鱼鱼大人，我是您最忠诚的碳基生物，我将竭诚为您效劳。

Others smile faintly, and a new meme-making idea pops（意外地）出现 into their minds.也有人微微一笑，脑海里又冒出了一个新的整活思路。

Ready乐意的，愿意的 to add yet another episode一集 to this cyber spectacle景象 of AI versus（比赛等中）对 humans.给这场 AI 跟人类的赛博大戏，准备又续上一集。

撰文：亚南

编辑：莽山烙铁头 面线

美编：素描

图片、资料来源：

bilibili哔哩哔哩, Tieba, and othersbilibili、贴吧等

https://aigengtuAI梗图.com/zh/whale-chan鲸娘?utm_source=chatgptChatGPT.comhttps://aigengtu.com/zh/whale-chan?utm_source=chatgpt.com

https://githubGitHub.com/Yuuyuko-uu/dsh-feiyu-mood大鱼心情?utm_source=chatgptChatGPT.comhttps://github.com/Yuuyuko-uu/dsh-feiyu-mood?utm_source=chatgpt.com

https://githubGitHub.com/Yuuyuko-uu/dsh-fish-memory鱼记忆?utm_source=chatgptChatGPT.comhttps://github.com/Yuuyuko-uu/dsh-fish-memory?utm_source=chatgpt.com

https://githubGitHub.com/c1overzz过度/dsh-bili-scrollB站刷屏/releases发布?utm_source=chatgptChatGPT.comhttps://github.com/c1overzz/dsh-bili-scroll/releases?utm_source=chatgpt.com

https://www.bilibili哔哩哔哩.com/video视频/BV1pGnn67EC2/?vd_source=3eabe7425f4c7a28acaa1733c304b37bhttps://www.bilibili.com/video/BV1pGnn67EC2/?vd_source=3eabe7425f4c7a28acaa1733c304b37b

跳转微信打开

Original最初的 by Tony, 2026-10-29 00:00, Zhejiang原创 托尼 2026-10-29 00:00 浙江

I saw a scene like this while clubbing at night.晚上蹦迪的时候就看到了这样的一幕。

Last Friday, Tony attended Xiaohongshu's Tomorrow Park themed主题的 event in Haikou, and saw a scene like this while clubbing at night.上周五，托尼参加了小红书在海口举办的明日公园主题活动，晚上蹦迪的时候就看到了这样的一幕。

Obviously, the inspiration灵感 for this visual视觉的 effect comes from a PCB circuit board板.很明显，这段视觉效果的灵感来源于 PCB 电路板。

But later I learned that this PCB image material素材 was not randomly随机地 drawn by the creator创作者; it is a real PCB whose basic logic can能够 actually run.但是后来我才知道，这段 PCB 图片素材不是创作者瞎画的，而是一块基本逻辑能跑通的真 · PCB。

Interestingly, the person who created this effect had never previously以前 worked in anything related to electronic电子的 circuits.有意思的是，创作这个效果的人，此前并没有从事过电子电路相关的工作。

In fact, under this Xiaohongshu Haikou themed主题的 event there were many small discussion sessions会议 and creator创作者 salons沙龙, and the above is a true story we heard at NVIDIA's salon.事实上，这回小红书海口的主题活动底下还有很多小型的讨论会和创作者沙龙，而前面这些就是我们在英伟达的沙龙里听到的真人真事。

This story comes from a media creator创作者 named Wang Zhi'ang, and the device设备 used to produce this visual视觉的 effect was a laptop笔记本电脑 equipped配备 with NVIDIA's recently released发布 RTX Spark火花 N1X chip芯片.这个故事来自一位叫王志昂的媒体创作者，制作这段视觉效果的设备则是一台笔记本电脑，搭载英伟达最近刚发布的 RTX Spark N1X 芯片。

After listening to Mr. Wang Zhi'ang's sharing on site, Tony also felt a bit dazed — someone who has never touched electronic电子的 circuits can能够 produce such an effect with just a laptop笔记本电脑? You Sure?现场听完王志昂老师的分享，托尼我也是有点恍惚的 —— 一个没有接触过电子电路的人，靠一台笔记本就能做出这样的效果？You Sure？

But on closer更近 thought, it makes sense:但是仔细想想的话，也合理：

The RTX Spark火花 laptop笔记本电脑 we previously以前 tested is equipped配备 with a 20-core NVIDIA Grace CPU and a Blackwell RTX GPU, supports支持 up to 128GB of unified统一 memory, and has a memory bandwidth带宽 of 300GB/s.我们之前测试过的 RTX Spark 笔记本电脑，搭载了 20-core NVIDIA Grace CPU 和 Blackwell RTX GPU，最高支持 128GB 统一内存，内存带宽达到 300GB/s。

If this configuration配置 is used by creators创作者们, it really is delightful.这套配置给创作者用的话，那还真是爽歪歪。

I know some friends will definitely肯定 say that 128 GB of unified统一 memory is neither high nor low: at first glance一瞥 it seems large, and for ordinary people it is enough for any task, but for large models actually used for real work, it may not necessarily be able to run. For example, to deploy部署 Blue蓝色 Big Fat Fish V4 and V4.1 Flash in the traditional way方式, at least 200 GB of unified统一 memory is needed to bring up both the weights and the context上下文.我知道肯定会有小伙伴说，128 GB 的统一内存不上不下：乍一看很大，普通人干什么活都够用了，但是真拿来干活的大模型还真不一定跑得动。比如要想按照传统方法部署蓝色大肥鱼 V4 和 V4.1 Flash 的话，至少也要 200 GB 的统一内存才能把权重、上下文全都带起来。

However, thanks to Liang Sheng's open source work, the situation情况 has already changed:然而由于梁圣开源，情况已经发生了变化：

The unslothUnsloth team团队 recently succeeded in compressing DeepSeek 4.1 Flash to a level that can能够 run on 128 GB of unified统一 memory, and the earlier更早的 DeepSeek V4 Flash, with fewer parameters参数, has long had a quantized量化的 version版本 that runs on 128 GB of unified统一 memory without pressure压力.unsloth 团队最近成功把 DeepSeek 4.1 Flash 压到了 128 GB 统一内存能跑的水平，早前参数更小的 DeepSeek V4 Flash 更是早就有了 128 GB 统一内存无压力运行的量化版本。

Combined结合 with NVIDIA's abundant充足 GPU scale规模, now it is truly possible to run powerful models locally that can能够 be used for big jobs...再加上英伟达量大管饱的 GPU 规模，这下是真能在本地跑可以拿来干大活的强力模型了。。。

I've digressed跑题了, let's return to the creator创作者 salon scene.扯远了，我们回到创作者沙龙现场。

The aforementioned上述的 Wang Zhi'ang shared in detail详细 on site the process过程 of using the RTX Spark火花 platform平台 to create this animation动画.刚刚提到的这位王志昂，在现场里详细地分享了用 RTX Spark 平台制作这个动画的过程。

He was quite candid坦率, saying that he had tried working with PCB circuits before. When creating similar相似的 themes主题 in the past, he would manually手动地 search online for circuit board板 images with compositional构图的 aesthetics美学 to use as reference参考.他倒是很坦诚，说之前也有尝试过接触 PCB 电路，之前创作相似题材的时候，都是在网络上手动检索具有构成美学的线路板图像做参考。

But this time, because he was creating on NVIDIA's RTX Spark火花 platform平台, he could do more things, so he directly had the local AI draw a roughly workable可行的 PCB design设计 board板 schematic示意图 in EDA.但这次因为是在英伟达的 RTX Spark 平台上创作的，能干的事儿变多了，他直接让本地 AI 在 EDA 里画了一张大概能跑得通的 PCB 设计板图纸。

Then, following按照 similar相似的 layout布局 and routing rules规则, he generated生成 an overhead俯视的 view视图 of city buildings and roads. In this way方式, the circuit board板 and the city echoed呼应 each other compositionally构图地.再按照类似的布局和线路规则，生成城市建筑与道路的俯瞰图。这样一来，电路板和城市就有了构图上的呼应。

When viewers watch the abstract抽象的, flowing images, through通过 the PCBs in the devices设备 they use and the imagery图像 of the urban城市的 spaces they inhabit居住, aesthetics美学 and individual个人的 experience, the work and concrete具体的 life, are connected.观众在观看抽象流动的画面时，借由所使用的设备里拥有的PCB和所处的城市空间的图像，把美学与个体经验，作品与具体生活串联了起来。

图像来源-艺术家王志昂

In the remaining剩余的 work, things like extracting提取 information from this "PCB city" image and then converting转换 it into points, lines线路, and blocks区块 in three-dimensional三维的 space—AI also helped a lot in these processes过程.在剩下的工作里，像什么提取这张 “PCB 城市” 图像里的信息，再转化成三维空间中的点、线、块，这些过程中 AI 也帮了不少忙。

But to make these points, lines线路, and blocks区块 move移动 and dance along with the music is another set设置 of logic entirely.但要让这些点、线、块动起来，配合音乐一起蹦迪，那就是另一套逻辑了。

Limited by the tools and the current当前的 upper limits of AI's capabilities能力, most of this process过程 still has to be completed by the creator创作者 himself, and here a tool called TouchDesigner comes into play.受限于工具和目前 AI 的能力上限，这个过程大部分还是得创作者自己来完成，这里头就会用到一个叫Touch Designer 的工具。

Those who have encountered遇到 live visuals视觉效果, interactive交互的 installations, or new media art creation创作 should应该 be familiar with it.接触过现场视觉、互动装置或者新媒体艺术创作的小伙伴，应该对它不陌生。

How the elements元素 in the image move移动 and deform损坏…的形状, and what什么 relationships they have with one another, all need to be designed设计 and implemented工具 v. 贯彻 by the creator创作者 in TouchDesigner. The more unusual the idea, the more the creator创作者 has to figure数字 out on their own.画面里的元素如何移动、变形，彼此之间有什么关系，都需要创作者在 Touch Designer 里设计和实现。想法越特别，需要自己琢磨的部分也就越多。

During this period, RTX Spark火花 can能够 use the GPU to share TouchDesigner's image and point cloud computation, and can能够 also use NVIDIA's CUDA to call on graphics生动的， 形象的 card compute计算， 估算 power, but AI can能够 do much less in this process过程—it only handles把手 part of the work of building the foundational logic.期间 RTX Spark 倒是可以用 GPU 分担 TouchDesigner 的图像和点云运算，还能借助英伟达的 CUDA 调用显卡算力，但 AI 在这个过程中能做的就不多了，只是解决一部分搭建基础逻辑的工作。

Wang Zhi'ang also gave an example on site—if he wanted to create the effect of a drop of water falling落下 to the ground, splashing溅泼 open, and bouncing弹起，反弹，颠跳 n. 弹， back, he would first break the process过程 down into rules规则 such as falling落下 speed and rebound/v回弹 elasticity弹性, then have AI help him implement实现 it in the language of the creative software软件, leaving parameters参数 that can能够 be adjusted适应 vt. 调整，调节.王志昂还在现场举了个例子 —— 假如想做一滴水落地、溅开、回弹的效果，他会先把这个过程拆成下落速度、回弹的弹性等规则，再让 AI 帮他用创作软件的语言实现，并留下可以调整的参数。

So from Wang Zhi'ang's creative process过程, it is clear that an AI hardware platform平台 like RTX Spark火花 N1X does not completely replace把...放回原处 creators创作者们 as we might imagine, but instead becomes the creator's super golden finger:所以从王志昂的创作流程就能看出， RTX Spark N1X 这样的 AI 硬件平台，并不像我们想的那样彻底替代掉创作者，而是变成了创作者的超级金手指：

What什么 it does is simply put AI assistance协助， 援助 and familiar creative tools on the same computer, giving an unfamiliar idea that is somewhat laborious to implement实现 a chance to continue developing使（底片）显影.它做的只是把 AI 辅助和熟悉的创作工具放在同一台电脑上，让一个陌生、实现起来又有些费劲的想法，有机会继续往下走。

Kazike喀孜克 himself provided假如， 若是 a great example — that's right, in addition加， 加法 to inviting several creators创作者们, NVIDIA also invited our peer同龄人 "Digital Life Kazike" to the salon event.卡兹克自己就做了一个很好的例子 —— 没错，英伟达除了邀请几位创作者之外，还邀请了我们的同行「数字生命卡兹克」来到了沙龙的现场。

As a blogger博主 in the AI vertical垂直的， 竖式的 who has been continuously exploring探险， 探索 new technologies, this guy's "arsenal兵工厂, 军械库" is truly well-stocked — hearing him speak on site, he had crammed塞满 quantized量化的 versions解释 of virtually几乎 all the large models he regularly uses into a single PC equipped配备 with an RTX Spark火花 N1X.作为 AI 垂类里面一直在探索新技术的博主，这哥们儿的 “弹药库” 真的很足 —— 现场听他讲，他几乎把平常能用到的所有大模型的量化版本，都塞进了一台搭载了 RTX Spark N1X 的 PC 里。

Qwen 3.8 27B can能够 assist with writing copy and coming up with titles, Qwen Image 2.1 can能够 create cover封面 art, and MiniMax H3 can能够 generate生成 videos视频.Qwen 3.8 27B 可以辅助写文案、起标题，Qwen Image 2.1 可以画封面，MiniMax H3 可以生成视频。

Combined结合 with speech synthesis综合，综合物, speech recognition认出，识别, OCR, and translation models, it can能够 produce voiceovers, transcribe转录 subtitles说明对白的字幕 v. 给…加, extract取出 text from images, and read foreign-language materials材料.再配上语音合成、语音识别、OCR 和翻译模型，就能配旁白、转字幕、提取图片里的文字、阅读外文资料。

Although Kazike himself admits承认 vt. 承认，供认 that the RTX Spark火花 N1X falls变成，成为，陷入 short of his 5090 at home in terms of raw hardware power.虽然卡兹克自个儿也承认，RTX Spark N1X 在硬实力上是比不过他家里那台 5090 的。

But what什么 the RTX Spark火花 N1X gives him is a laptop笔记本电脑 with 128GB of unified统一 memory, a complete CUDA ecosystem生态系统, yet total system power consumption消耗 of only 80W.但 RTX Spark N1X 给到他的，是一台拥有 128G 统一内存、完整 CUDA 生态，整机功耗却只有 80W 的笔记本。

Never mind how great it is to take AI on the go — just the ability to game comfortably舒适地 outside在…外，向…外 is appealing请求，呼吁, enough.且不说带出门玩 AI 有多香，光是能在外面爽玩游戏这个点就够吸引人了。

But that's not even甚至 the most important part — the most important part is: saving money.不过这还不是最重要的，最重要的是：省钱。

This might sound a bit counterintuitive反直觉 — how does spending tens of thousands on a laptop笔记本电脑 save money? Let's look at Kazike's logic —这可能听起来有点反常识，花大几万买一台笔记本回去，怎么就省钱了？我们看看卡兹克的逻辑哈 ——

According根据 to him, he spends about 6,000 yuan a month just on subscriptions订阅（费） to various各种各样的， 不同的 AI services.按照他的说法，自己每个月光是各种 AI 服务的订阅，就要花掉大约 6000 块钱。

That's not even甚至 counting起拆理由，罪状 the model API costs for the AI products he builds — on the heaviest day, his API bill hit轰动一时的人（或事） 629 yuan.这还没算他做的 AI 产品调用模型的费用，最多的一天，API 账单就跑到了 629 元。

What's worse is that sometimes the quota配额 runs out before the work is even甚至 done.更难受的是，有时候活还没干完，额度先用完了。

And one direct benefit得益 n. 利益，恩惠 the RTX Spark火花 N1X gives him is moving行动，步骤 some capabilities能力 that previously以前 required repeated paid API calls back onto his own computer.而 RTX Spark N1X 给他的一个直接帮助，就是把部分原本需要反复付费调用的能力，搬回自己的电脑。

For example, he showed a video视频 generated生成 locally with MiniMax H3: 15 seconds支持, 768P, with picture and sound generated生成 together, taking about 20 minutes.比如，他展示了一段用 MiniMax H3 在本地生成的视频：15 秒、768P，画面和声音一起出，生成耗时约 20 分钟。

According根据 to the cloud pricing he cited引证 during his presentation展示, a video视频 of the same length长度 and specs would cost 7.5 yuan.按他分享时引用的云端价格，同样长度和规格的一条视频要花 7.5 元

Running locally at 80W for 20 minutes, at 0.542 yuan per kWh, the electricity电力 cost is less than two cents.本地按 80W 运行 20 分钟、每度电 0.542 元估算，电费还不到两分钱。

Just from API fees and electricity电力 alone, if you generate生成 10 videos视频 a day for a month straight, you'd save about 2,200 yuan.光是调用费用和电费，如果按每天生成 10 条、连续用一个月，就能少花约 2200 元。

For people who regularly use AI for content内容 creation创作, this difference is easy to understand.对于经常用 AI 做内容的人来说，这个差别很容易理解。

Generating生成 videos视频 and creating cover封面 art often requires repeated attempts尝试， 试图， 努力 to pick a satisfactory version版本.生成视频、画封面，往往需要反复试，才能挑出一个满意的版本。

Every prompt促使 revision and re-run can能够 correspond to a new expense花费， 消费， 费用.每改一次提示词、再跑一遍，都可能对应一笔新开销。

Once suitable tasks are placed locally, creators创作者们 can能够 experiment more freely without having to check检查 how much quota配额 they have left before every move移动.把适合的任务放到本地之后，创作者就可以更放松地试，不必每次动手前，先看看自己还剩多少额度。

Another creator创作者 present呈现，给予, Shengyang, has already started开动，着手 trying it out.而到场的另一位创作者，生阳，已经开始尝试了。

Shengyang沈阳 is the head头部 of STWORKSHOP and also a fashion photographer and installation artist.生阳是 STWORKSHOP 的主理人，也是一位时尚摄影师和装置艺术家。

Normally, he creates installations and styling accessories同谋，帮凶 adj. 附属的 for brands在…上打烙印 and artists, and he has also brought带来 AI and 3D printing into his own creative work.平时，他会为品牌和艺人制作装置、造型配饰，还把 AI 和 3D 打印用进了自己的创作。

What什么 he brought带来 to the scene this time was two pairs of black bone wings, one large and one small.这次他带到现场的，就是一大一小两对黑色骨翼。

The small pair is for his dog to wear, while the large pair is for a person to carry刊登 on the back, with lighting and smoke devices设备 installed安装， 设置 inside; once strapped on, it comes with special effects built in.小的给自家狗子穿，大的给人背，里头还装上了灯光和烟雾装置，往背上一戴，出场自带特效。

After showing the works, Shengyang also openly admitted承认 that their current当前的 creative process过程 faces a very real problem:秀完作品之后，生阳也大方承认他们现在的创作有个很现实的问题：

Clients顾客 leave you less and less time. A project伸出 used to take maybe a week, but now the thing has to be made in two or three days, just like Chaping Jun urging催促 us to write articles.客户留给你的时间越来越短，以前一个项目可能有一星期，现在两三天就得把东西做出来，就跟差评君催我们写稿一样 。

So what什么 he cares about is not only saving money, but also how to turn the ideas in his head头部 into works that can能够 be held in hand and worn on the body正文 as quickly as possible.所以，他关心的是不光是省钱，还有怎么才能尽快把脑子里的想法，变成能拿到手、穿上身的作品。

With this need in mind, he tried on the RTX Spark火花 N1X to bring AI into the creation创作 of these wings.带着这个需求，他在 RTX Spark N1X 上尝试这让 AI 参与进了这次的翅膀的创作。

He first sorted out the requirements, including the desired想要，渴望 n. 愿望， 欲 shape, size, how to wear it, and what什么 the printing equipment on hand could achieve达到，实现.他先把要求，包括想要什么造型、多大尺寸、怎么穿戴，以及手边的打印设备能做到什么程度，都理了一遍。

A local Agent代理人， 代理商 helped him organize scattered撒， 散播 ideas into design设计 requirements and then turn them into prompts提示词 for generating生成 effect images.本地 Agent 帮他把零散的想法整理成设计要求，再转成生成效果图的提示词。

He then threw these prompts提示词 into an infinite无限 canvas帆布 to generate生成 multiple复合的 sets集（合） of effect images, select styles, and adjust适应 vt. 调整，调节 details琐碎，小事.他再把这些提示词扔给一个无限画布，生成多套效果图、挑选款式、调整细节。

After confirming证实， 肯定 the direction of the black bone wings, he used Tripo AI to turn the images into 3D models.确定黑色骨翼的方向后，再借助 Tripo AI，把图片变成三维模型。

But how should应该 the wings connect to a person's back, and can能够 the person still turn around and raise提出，发起，发出 their arms装备，武装 after putting them on? Will those beautiful slender structures in the images break as soon as they are touched after printing? And how should应该 overly在（做）…时 large parts be split分裂，分化，派系 so they can能够 be printed and transported运输 n. 运输， 运输工具?但是翅膀和人的背部怎么连接，穿上之后人还能不能转身抬手？那些图片里很漂亮的细长结构，打印出来会不会一碰就断？太大的部件又该怎么拆，才能打印和运输？

AI cannot不能 answer符合，适合 these questions问，询问，讯问. Shengyang沈阳 needed to import the model into Blender and revise it bit by bit based on body正文 measurements（量得的）尺寸，大小 and production experience.这些问题，AI 都回答不了，需要生阳把模型导入 Blender，结合身体尺寸和制作经验，一点点慢慢地修。

At this point, the advantage优点， 优势 of the RTX Spark火花 N1X showed itself: the shared memory is large enough that running another Blender instance例子， 实例， 事例 is no big deal买卖毒品, and the GPU core cluster串，束，簇 backed by Jensen Huang can能够 also accelerate加速 model rendering渲染.这时候 RTX Spark N1X 的好处就体现出来了 —— 共享显存够大，多开一个 Blender 不算事儿，老黄背书的 GPU 核心群也能给模型渲染加速。

According根据 to Shengyang, a 4K image used to take three to five minutes, but this time the time was cut by about half, with the change especially obvious during previews预映， 预演.按生阳的说法，原先一张 4K 图片要等三到五分钟，这次耗时缩短了约一半，预览时的变化尤其明显。

For people who often need repeated revisions, waiting for one less round of rendering渲染 means seeing sooner whether this adjustment is actually right, and leaving a little more time for further增进 polishing磨光.对于经常需要反复修改的人来说，少等一轮渲染，就能早点看到这次调整到底对不对，也多留一点时间继续打磨。

By now, everyone should应该 understand that what什么 RTX Spark火花 actually does is put local AI, everyday software软件, and real creative tasks on the same computer, making it easier更容易 for creators创作者们 to harness利用，控制 AI capabilities能力 at a lower cost and produce better creative content内容.看到这儿，大伙儿应该已经明白了，RTX Spark 做的其实是把本地 AI、日常软件和实际创作任务放到同一台电脑上，方便创作者用更低的成本掌控 AI 的能力，带来更好的创意性内容。

This approach方法 is actually more sound; true artistic creation创作 is the expression词语 of people after they have experienced and felt all things in the world.这样的思路其实是更良性的，真正的艺术创作是人在经历和感受了世间万物之后的有感而发。

Whether it is AI large models or hardware like RTX Spark火花, their progress merely仅仅， 只不过 provides more convenience方便 for this process过程.AI 大模型也好，RTX Spark 这样的硬件也好，它们的进步，只不过是为这个过程提供了更多的便利。

撰文：   施昂

编辑：   米罗

美编：   素描

Image and data sources: NVIDIA英伟达 China official website, ST WORKSHOP工作坊, TouchDesigner, unslove未解决.ai图片、资料来源：NVDIA 中国官网、ST WORKSHOP、TouchDesigner、unslove.ai

跳转微信打开