Genie 4:世界模型开始出声,评论区开始逐帧验货 — HN 讨论摘要
原文概要
Google DeepMind 在 10 月 27 日发布 Genie 4,署名作者是 Raia Hadsell、Diego Rivas 和 Hang Qi。定位是「新的世界模型」(world model),核心卖点是原生多模态:画面、环境音和交互对话是同时生成的。官方说 Genie 4 建立在 Gemini 家族能力之上——Gemini Omni 负责把模拟锚定在真实世界的外观与行为上,其他 Gemini 模型负责维持场景的连贯与连续。
交互形态是键盘控制移动与镜头,再用文本 prompt 实时改写场景:改天气、移动镜头、加入新的人或物体,Gemini 会结合当前画面调整你的指令;你不出手的间隙,Gemini 仍在监控并推动场景继续往前走。官方列了四项能力:实时塑形;物理与人类动作的真实感(空气动力学、碰撞,人物动作自然、能被实时引导);原生音频——这是第一个实时生成音频的世界模型,风声、海浪、角色对话都能实时生成并对输入作出反应;以及把 Gemini 拉进来模拟复杂事件,其中一个展示是把 Gemini 4 Argon、Genie 4 和 Gemini Robotics 组合起来,模拟 Apptronik 的 Apollo 人形机器人下国际象棋,并把任务拆成实时步骤执行。
演示样例包括:湖面浮台上吹萨克斯的纪录片质感镜头(提示词里写着 16mm 胶片、低反差、不要 HDR)、全 stained-glass 材质的玻璃世界、骑在龙背上的第一人称飞行、1980 年代的日本车站、桌上开箱露出一只孔雀的魔术师场景,以及从一张 2D 档案照片几秒内长成可以走进去的模拟。
发布形式是 research preview,先给 trusted testers,之后先给 Project Genie 里的 Google AI Ultra 订阅者,以及 Gemini API 和 Google AI Studio 的早期接入伙伴;所有生成的音视频都带 SynthID 水印。官方自陈的局限同样具体:短时间内能保持连贯,但跑得久了、或经过多次复杂 prompt 编辑之后,视觉细节和物体恒常性会漂移;复杂物理交互、实时音频准确性、渲染出可读文字、空间准确度、长期一致性,都还是进行中的研究。
内部测试者反馈集中在几个方向:电影人把它当预演用的「draft mode」,能实时跑好几条 take,随时改机位、灯光和台词;照片记者把一张历史照片变成几秒可进入的模拟,用来推演山洪或山火怎么扩散;游戏开发者在写引擎代码之前先用自然语言摸出可探索的原型。HN 上这条帖子拿到 489 分、287 条评论。
讨论焦点
一、这到底是世界模型,还是给视频生成加了个键盘
第一条质疑指向可控性。sodacanner 的观察最直接:
“I don’t know if I’m missing the point, but Genie 4 almost looks worse than Genie 3 in terms of player control? It doesn’t even seem to respond to inputs, it just looks like they took some AI generated videos and put the key widgets on top.” — sodacanner [c:50261519] (译文:也许是我没看懂,但单看玩家操控,Genie 4 好像还不如 Genie 3?它甚至不太像是对输入有反应,看起来就像拿 AI 生成的视频,上面贴了几个按键。)
ainch 认了这一点,并把 Genie 的来历讲了一遍:
“The original Genie was a great piece of work with a real sense of purpose - to enable agents to learn about the real world in the way Dreamer or Schmidhuber’s paper sketched. But now most of the people that drove that mission have left Deepmind (Jack Parker-Holder, Tim Rocktaschel, Demis) and it’s become this instead.” — ainch [c:50264841] (译文:最初的 Genie 是一件很了不起、目标感明确的作品——它要的是让智能体按 Dreamer 或 Schmidhuber 论文设想的方式去理解真实世界。但现在推动那个使命的人大半已经离开了 DeepMind(Jack Parker-Holder、Tim Rocktaschel、Demis),它变成了现在这个样子。)
这条「原始使命被换掉了」的判断,官方致谢名单可以提供部分对照——Jack Parker-Holder 仍在核心贡献者名单里,Demis Hassabis 也出现在受谢名单末尾。但更尖锐的质疑来自命名本身。jackb4040 把 Google 的用法和 LeCun 的定义直接对撞:
“It’s hard for me to understand Google’s redefinition of “world models” (basically: a txt2videogame modality for LLMs) as anything other than an attempt to bury Yann Lecun’s definition of “world models”, which includes an implicit bear case against LLMs and thus Google’s stock price.” — jackb4040 [c:50261898] (译文:Google 对「世界模型」这词的重新定义(说白了:给 LLM 加一个文本生成电子游戏的模态),我很难理解成别的什么,除了想埋葬 LeCun 对「世界模型」的定义之外——后者内含对 LLM 的看空逻辑,也就是对 Google 股价的看空逻辑。)
而把「世界模型」这个词重新校准的,是评论区里这句:
“This is a reflection of a complaint I frequently make about so called “world models”. It’s more accurate to call them video game models. They don’t simulate the world, they simulate video games. For that reason a common issue with them is that the physics are video game physics rather than real world physics.” — Bjorkbat [c:50264089] (译文:这是我经常抱怨的一点:所谓「世界模型」更准确的说法是「电子游戏模型」。它们不模拟世界,它们模拟电子游戏。所以这类模型常见的问题就是,物理是电子游戏的物理,不是真实世界的物理。)
二、逐帧验货:一份穿帮清单
演示视频被当成送测固件一样拆开看。第一个被抓住的是物理。
“In the clip with a marble bust dropped on a watermelon, at this point https://imgur.com/a/IHLnkUs the bust should’ve toppled over, if the physics were properly modeled.” — mlmonkey [c:50261683] (译文:那段大理石胸像掉到西瓜上的片段,到这个位置的时候胸像本该已经翻倒了——如果物理被正确建模的话。)
“It also failed to properly model the physics of knights riding wyverns into battle.” — sobellian [c:50262893] (译文:它同样没有正确建模骑士骑着双足飞龙冲入战场的物理。)
“A big chunk of the right side of castle disappeared, including catapult, when camera turned back while flying on dragon. See 1:25 to 1:29. I was not expecting this to happen in a “world model”.” — smusamashah [c:50261864] (译文:骑着龙飞的时候镜头转回来,城堡右侧一大块连同投石机直接消失了。看 1:25 到 1:29。我没想到这种事会发生在「世界模型」里。)
“There’s also a whole extra bonus castle that appears to the left when the camera looks left around 0:22. Which is amusing both because it wasn’t there on the initial approach, and also because nobody puts two castles each with their own distinct curtain walls right next to each other like that.” — dantillberg [c:50263958] (译文:另外还有一整座附赠城堡,在 0:22 左右镜头往左看时冒出来。搞笑的地方有两点:它一开始进场时并不在那儿;而且没人会把两座各自带着城墙的城堡并排放得这么近。)
整个龙背飞行片段被评价为「廉价货」:
“That entire clip looks like it’s straight out of a complete bargain-bin videogame. The weird smoothness of the ride, the pivot in mid air, the wings apparently don’t have collision with the poles they fly past.” — OkayPhysicist [c:50262497] (译文:整段片子的观感就像是从 bargain-bin(廉价货)电子游戏里直接剪出来的。飞行过程诡异地顺滑,半空中做出翻转,而翅膀显然跟它飞过的杆子完全没有碰撞。)
“Definitely wouldn’t have been one of my picks to show off my fancy new model.” — OkayPhysicist [c:50262497] (译文:如果是我拿自己那台新模型做演示,这段绝对不会选。)
为这些穿帮辩护的人也有,理由是「那段本来就是在演示 steering」:
“They were demonstrating steering.” — dsissitka [c:50262444] (译文:他们那段本来就是在演示操控。)
另一位把物理穿帮读成了机器人该有的心智模型:
“We all have a world model in our heads of what we think is going to happen, we adjust quickly (or not) based on what does. This is pretty incredible stuff.” — peddling-brink [c:50262823] (译文:我们脑子里都有一套关于「接下来会发生什么」的世界模型,再根据实际发生的事快速调整(或者不调整)。这东西确实了不起。)
这段辩护立刻被人怼回去:
“Busts rarely hang in midair at a 45° angle.” — bigwheels [c:50263646] (译文:胸像很少会以 45 度角悬在半空。)
“As an aside, why fall into the role of being an apologist for google-hype and Big Tech?” — bigwheels [c:50263646] (译文:顺带一问,为什么要跑去当 Google 吹鼓手和大科技公司的辩护律师?)
真正致命的追问是:这种物理到底能用来干什么。modeless 把游戏和机器人训练分开算:
“Which is fine for a video game. Not at all fine if you want to train robots with it. Trying to combine both use cases into a single model is doomed to failure IMO.” — modeless [c:50263619] (译文:拿来当电子游戏没问题。但要拿它训练机器人就完全不是那么回事了。在我看来,想把这两种用途塞进同一个模型里,是注定要失败的。)
三、Stadia 2.0:世界模型会不会变成新 DRM
整条帖子里最有历史感的争论,是有人把 Genie 4 直接接到了 Stadia 身上。
“As I said before, World models will become simulation engines of AAA games where there is no need to release the actual game artifact and even if the game exists as a binary artifact, it will never be released and will be available on their hosted cloud back-end, then simulated and streamed back to you.” — rvz [c:50261386] (译文:正如我之前说的,世界模型会变成 3A 游戏的模拟引擎——你根本不需要拿到游戏本体;即便游戏以二进制形式存在,它也永远不会发布给你,只存在于他们的云端后端,模拟完再流回给你。)
“So, Stadia + World Models = New generation of DRM for games.” — rvz [c:50261386] (译文:所以,Stadia + 世界模型 = 游戏 DRM 的新一代。)
Stadia 到底怎么死的,评论区吵了一轮,三种解释都有。verdverm 归给体验:
“Stadia is dead, the reason is playing games over the internet sucks, input lag and serious graphics degradation - this is why I cancelled” — verdverm [c:50261466] (译文:Stadia 死了,原因是通过网络玩游戏体验糟糕——输入延迟、画质严重退化。我就是这么退订的。)
ZiiS 归给算力去向:
“Stadia is dead because if you have a data-center of GPUs you are using them for AI. It has nothing to do with input lag and graphics degradation.” — ZiiS [c:50261591] (译文:Stadia 死了,是因为如果你手握一个 GPU 数据中心,你会拿它跑 AI。这跟输入延迟和画质退化毫无关系。)
ActionHank 认为这三方都记错了时间线:
“Oh no my guy, you are misremembering, stadia died a quiet death long before AI.” — ActionHank [c:50261654] (译文:兄弟你记错了,Stadia 在 AI 之前很久就已经安静地死了。)
抛开死因不谈,钱的问题更实际。perchedonstilts 直接算账:
“None of this makes sense. Think of the economics involved here: is it cheaper to run tens/hundreds of thousands of GPU clusters to simulate a unique game world for millions of players, or would it be cheaper to just write the god damn code once, and then host it for download, or to allow them to be streamed?” — perchedonstilts [c:50264746] (译文:这一套都说不通。想想要摊进去的经济账:为了给几百万玩家各自模拟一个独一无二的游戏世界,跑成百上千个 GPU 集群更便宜,还是干脆把代码写一遍、放上去托管让人下载或串流更便宜?)
paxy 的反应更直接:
“Because it wasn’t bad enough to need a $5000 5090 to play a game, now you’re going to need a full rack of them!” — paxy [c:50261609] (译文:玩个游戏要一台 5000 美元的 5090 还没被骂够,现在你得配满一整个机架了!)
但也有评论指出这跟玩家的关系另有算法。SomeonesAccount 解释了「为什么非确定性引擎对游戏更好」这个问题的答案:
“It wouldn’t be better for players. It would be better for companies because the user no longer owns the computer that runs the game.” — SomeonesAccount [c:50261552] (译文:那对玩家并没有更好。对公司才更好——因为运行游戏的电脑不再属于用户。)
moralestapia 顺着这句话算出了内容行业的账:
“And a month or so of development time. Pure IP re-selling.” — moralestapia [c:50261628] (译文:再加上一个月的开发时间。纯粹的 IP 转卖。)
“October? Yeah, just release the “Spiderman saves Halloween” prompt.” — moralestapia [c:50261628] (译文:十月档?好啊,直接发布「蜘蛛侠拯救万圣节」这条提示词就行。)
Matl 则提醒 Stadia 当年的商业模型本身也不成立:
“But the model was free version of the service for a limited resolution AND you have to buy games there OR a subscription based option and you still have to buy games there.” — Matl [c:50262151] (译文:但那个模式是:服务只提供限画质的免费版本,而游戏你还得在那边买;或者走订阅制,但游戏照样还得在那边买。)
“That made no sense and wasn’t a viable business model.” — Matl [c:50262151] (译文:这说不通,也不是可行的商业模式。)
游戏行业自己的意见更谨慎。causal 泼了一盆冷水:
“Game industry is still struggling to find a way to incorporate regular LLMs without suffering catastrophic edge cases. A long way to go before these can do more than fancy tech demo, I think.” — causal [c:50261480] (译文:游戏行业到现在都还在挣扎:怎么把常规 LLM 塞进来而不遭遇灾难性边界情况。要让这些东西不只是花哨的技术演示,我看还早得很。)
也有人直接说:我家不买。qup 的判断被补充成了一整条独立游戏路线:
“They can try, my household won’t be purchasing those.” — qup [c:50261528] (译文:他们可以试,我家不会买这种东西。)
“The problem will be separating the wheat (people with good stories and concepts enabled by Ai) from the chaff (people with no taste slopping out trash with Ai)” — verdverm [c:50261722] (译文:真正的问题是把麦子和糠分开——有 AI 加持、故事和概念都不错的那部分,和没品味、靠 AI 批量倒垃圾的那部分。)
四、真正的目的可能不是游戏
讨论进行到一半,有人把话题从「游戏」拽到了「机器人」。inshard 的问题把三家的牌都摊开了:
“So world models capture inherent complexity of physical reality to help teach AI about hidden laws which probably helps train both LLMs and multimodal models used for autonomous vehicles and robots? And AMD acquiring World Labs is a move to outdo Nvidia’s work in this space. And Google probably sees this as a gateway to a new platform of autonomous vehicles and humanoids that they could rent out as a CSP? Competing with Nvidia, Tesla and complementing Waymo. Am I getting that right?” — inshard [c:50261674] (译文:所以世界模型是抓取物理现实的内在复杂性,好让 AI 学到隐藏的规律,这大概能同时帮到用于自动驾驶和机器人的 LLM 与多模态模型训练?而 AMD 收购 World Labs,是想在这个问题上压过 Nvidia?Google 大概是把世界模型当成通往自动驾驶与人形机器人新平台的入口,可以像 CSP 那样租出去?既和 Nvidia、Tesla 竞争,又补上 Waymo。你说得对吗?)
认同的回答只有一句,但很关键:
“The point is simulation for autonomous vehicles and robots.” — empath75 [c:50261771] (译文:重点就是给自动驾驶和机器人做模拟。)
NitpickLawyer 把这条路补全成一条实验室到现实的链条:
“One possible use for these is to facilitate lab -> this -> real world for robot models. Basically better than lab / algorithmically designed environments and cheaper / easier to train on than real life. Something that “looks” like real life and can scale, to apply RL for robotics to.” — NitpickLawyer [c:50261708] (译文:这东西的一个可能用途,是给机器人模型搭出「实验室 → 这里 → 真实世界」这条路。它比实验室环境或算法设计的环境更好,又比现实生活更便宜、更好训练——一个「看起来」像真实世界、又能扩展规模的东西,用来对机器人做强化学习。)
paimapi 则看到了另一面:
“Why hire crew or actors when you could have them be fully AI? Why film a commercial with fluffed, oil dusted, well-lit hamburgers when you could fully render something? Why actually have a vehicle capable of off-roading when you could just AI-render what it looks like off-roading (or vertically taking off, for eg)?” — paimapi [c:50261919] (译文:既然演员可以全用 AI,为什么还要雇剧组和演员?既然广告里的汉堡可以全渲染出来,为什么还要真的拍一个撒了油、打了光的汉堡?既然车看起来越野就行,为什么不直接用 AI 渲染出它在越野的样子(或者垂直起飞)?)
“It’s a dystopian business use case that I hope will get obliterated in the near future as the generations that believe in this stuff and are fascinated by it age out of the population.” — paimapi [c:50261919] (译文:这是个反乌托邦的商业用例。我希望随着相信这套东西、被它迷住的那一代人退出人口,这个用例能在不久的将来被彻底碾掉。)
五、闭源到什么程度:连参数量都不给
第二个负面主题是「不告诉你任何技术细节」。kouteiheika 把三条不满一次列完:
“Looks cool, but it’s very hard to get me excited about yet another proprietary model for which absolutely zero technical details are released (not even the number of parameters), which is safetymaxxed to the limit, and whose usage quota will (most likely) be so limited that it’ll be effectively useless for anyone who doesn’t work at Google or doesn’t have infinite money.” — kouteiheika [c:50261912] (译文:看着挺酷,但要我为又一个专有模型激动起来实在很难:它绝对零技术细节(连参数量都没有),安全限制拉满,而使用额度(大概率)会小到,对任何不在 Google 工作、也没有无限预算的人来说基本等于没用。)
有人觉得这套期待本身就不合理:
“What an insane set of requirements for a frontier technology and your excitement by it.” — estearum [c:50262205] (译文:对一项前沿技术提出这么离谱的一套要求,还能对此兴奋,也是难为你了。)
alex_sf 反驳的角度是 Google 的产品体验一向如此:
“I mean this doesn’t seem very insane? Basic info about the model, the ability to actually use it. Google in particular is awful about getting access to their models from the UX side: I don’t even bother trying them out unless it’s via openrouter.” — alex_sf [c:50262240] (译文:我倒觉得这样一点也不离谱。要基本信息,要真的能用得上。Google 在模型入口这件事上用户体验尤其差——不通过 openrouter 我都懒得试。)
价格问题则被两位读者当场算了一遍:
“Would be fun to muck around with world models. Are there any that I can try for under £200 p/m without being part of a special project?” — HughParry [c:50261886] (译文:能随手玩弄世界模型应该挺有意思。有没有不用参加什么特别项目、每月 200 镑以内就能试的?)
“Techincally, isn’t the less expensive sub for Gemini Ultra £75 p/m?” — msabalau [c:50263111] (译文:严格来说,Gemini Ultra 便宜一档的订阅不就是每月 75 镑吗?)
还有两个很实际的追问。一个是工程问题——这东西到底能不能落到游戏引擎里:
“So does this export 3d assets or an Unreal/godot/unity project so it can be turne dinto a game or is it limited to their runtime?” — pdntspa [c:50261637] (译文:那这东西能导出 3D 资产,或者 Unreal/Godot/Unity 工程,好转成真正的游戏吗?还是只能在他们自己的运行时里用?)
另一个是长度和成本:
“Question I have is when will they be able to render something more then 30 seconds.” — BatFastard [c:50261842] (译文:我有个问题:他们什么时候能渲染超过 30 秒的东西?)
“And how much will it cost? There has to be a HUGE amount of computing power behind this. As a video game developer I live in fear and anticipation.” — BatFastard [c:50261842] (译文:那要花多少钱?这背后肯定有巨量的算力。作为一个电子游戏开发者,我的生活里既有恐惧也有期待。)
六、演示本身的翻车:字幕、口音和产品经理
第三条负面主题指向 Google 自己的发布物料。第一个被抓的是 YouTube 自动字幕:
“It’s kind of amazing: the juxtaposition of such advanced generative AI and such terrible speech-to-text coming from the same parent company.” — Zambyte [c:50261538] (译文:说实话挺有意思:这么先进的生成式 AI,和这么烂的语音转文字,出自同一家母公司,这个并置相当惊人。)
被吐槽的字幕原文是 “Genie form [sic] can simulate scenarios”——[sic] 是字幕自己标的,Genie 被写成了 form。fwip 转头发现官方博客正文也有同样的问题:
“Like, yeah, I’d sure hope the music is similar to audio.” — fwip [c:50261598] (译文:是啊,我当然希望音乐和音频是差不多的。)
他指的是博客第二句里的 “enabling audio-like music, ambient noises, and interactive dialogue to be simultaneously generated with the world”——让「类音频的音乐」、环境噪声和交互对话与世界同时生成。artisin 把这句玩成了段子:
“When you’re high on vibes, the line between music and audio-like music tends to blur.” — artisin [c:50262311] (译文:当你 vibe 拉满的时候,音乐和类音频的音乐之间的界线就模糊了。)
出镜演员和音频的走样也没逃过。ytxm 挑的是第一个视频里那句台词:
“Surprised that “do you really think we’re going to make it?” scene made the cut of the first video, her expression and tone is totally out of sync with the scene and demonstrates the problem with almost every AI gen video I’ve come across when it comes to capturing humanity.” — ytxm [c:50261677] (译文:真没想到「你真觉得我们能做成吗?」那场戏会被剪进第一个视频。她的表情和语气跟场景完全对不上,这正好暴露了我见过的几乎所有 AI 生成视频在「抓住人」这件事上的通病。)
JDups 记下了另一个片段的细节:
“The 80’s Japanese train station clip also had the conductor speaking clear english in an accent I can’t quite place. Slight British tinge to it but not really? Not sure. The ambient audio also makes no sense.” — JDups [c:50264203] (译文:80 年代日本车站那段,售票员说的是一口清楚的英语,口音我说不上来,带一点英式的味道,但又不完全是?不确定。环境音也完全不合逻辑。)
zardo 把这归为文化上的偏袒:
“Showing off an anglo-bias with the Japanese train station in the 1980s, with a man speaking english.” — zardo [c:50261765] (译文:拿 1980 年代的日本车站做展示,还让里面的人说英语,这是在展示 Anglo 偏见。)
至于为什么这些穿帮能留在官方演示里,JDups 给了一个让评论区笑了的猜测:
“Maybe they don’t really care? “Gemini choose 10 scenes that would be impressive for our announcement and add them to the post.”” — JDups [c:50264203] (译文:也许他们根本不在乎?「Gemini,挑 10 个适合我们发布会、看着够唬人的场景,加到文章里。」)
更深一层的问题被 doctorpangloss 摆到了台面上——做这种产品的人是谁:
“but instead, they’re trying to get 100 google PMs, who don’t give their kids screentime and who don’t play any games the objective of making a product for Bart Simpsons.” — doctorpangloss [c:50261578] (译文:但他们现在的做法是找 100 个 Google 产品经理——这些人不让孩子看屏幕,自己也不玩游戏——却把目标定成「做一款给巴特·辛普森那种小孩玩的产品」。)
他原本的建议更狠:Google 花十分钟就能把最好的游戏开发者招进来,成本大概 200 万美元,九牛一毛,却没这么做。
七、模拟会不会被绕过
最后是这场讨论里最有理论含量的一条线。spwa4 说,世界模型这条路本身有个洞:
“That’s the thing about machine learning over time, and the big hole in the whole world model idea. Time and time again we make simulations and train models and we think that loop will work. It never does.” — spwa4 [c:50262307] (译文:这就是机器学习这些年来的规律,也是整个世界模型思路最大的那个洞。我们一次又一次地做模拟、训模型,以为这个闭环会奏效。它从来不会。)
“And why not? Because the first model that can actually learn to handle the simulation not only can do the simulation internally too, but it can do it better than out simulation.” — spwa4 [c:50262307] (译文:为什么?因为第一个真正学会处理这种模拟的模型,不只能在内部自己做模拟,而且做得比我们的模拟更好。)
“So we get to a point and people keep making complicated simulations of one kind or another … and then someone (some model) who’s never used the simulation, and doesn’t need it for model training gets by far the highest score on every benchmark and everyone agrees the simulation was kind of beside the point.” — spwa4 [c:50262307] (译文:于是我们走到某个节点,人们不停造着各种复杂的模拟……然后某个(某个模型)从没用过这个模拟、训练时也不需要它,却在每个基准上拿到远高于别人的分数,然后所有人一起承认:这个模拟好像有点多余。)
Closi 认为这取决于目标是什么:
“It depends if the goal is a more accurate simulation, or if the goal is a tool for creatives to build games or content.” — Closi [c:50264274] (译文:这要看目标是更准确的模拟,还是给创作者做一个做游戏、做内容的工具。)
“If it’s the latter, I don’t see a ‘hole’.” — Closi [c:50264274] (译文:如果是后者,我没看出有什么「洞」。)
sobellian 从技术路径上给出了反驳的反驳:
“I would not be so opinionated, modern coding agents derive from chatbot technology. Very different problem domain. They share language but language is universal so that doesn’t say much. Then they figured out that the pretraining is a good starting point for RLVR. Who’s to say a more cinematic imagination can’t be a good starting point for a world model? It doesn’t need to be the whole solution.” — sobellian [c:50264207] (译文:我不至于那么绝对。现代编码 agent 是从聊天机器人技术长出来的,问题域很不一样;它们共享语言,但语言是通用的,所以这说明不了什么。后来他们发现预训练是 RLVR 的好起点。谁能说更具电影感的想象力不能当世界模型的好起点呢?它不需要成为全部的解法。)
spwa4 在后面那条长回复里补上了更近的证据——语音领域的同类故事正在重演:
“This has already happened to other ML fields. For example, there’s not much point anymore to ASR models (just feed audio into LLMs) and the space that TTS models occupy is fast disappearing.” — spwa4 [c:50264404] (译文:这在别的机器学习领域已经发生过了。举例说,ASR 模型现在意义不大了(把音频直接喂给 LLM 就行),TTS 模型占的那块地盘正在快速消失。)
“And now there’s also hints of world models just being bypassed. Just giving a finetuned model the last 50 video frames (surprisingly, that fits in context, especially at reduced resolution), movement statistics, motor power, status, … of a robot and asking what to do actually works (with heavy finetuning, but I think it’ll start working without).” — spwa4 [c:50264404] (译文:而现在也有迹象表明世界模型同样会被绕过去:只要把最近 50 帧视频(出人意料地,压缩分辨率后居然塞得进上下文)、运动统计、马力、状态之类的信息喂给一个微调过的模型,然后问它该怎么做,这件事真的可行(需要重度微调,但我认为以后不微调也能行)。)
这场「标准到底在往哪挪」的争论,还牵出一句更尖的观察:
“The labs are the ones moving them. We’re all just measuring against their own, exponentially growing, claims.” — culi [c:50262721] (译文:挪标准的是那些实验室。我们所有人不过是在拿他们自己指数级膨胀的说法当尺子量。)
八、被打动的部分:彩绘玻璃、GTA 6 和一条辩护
不能只写骂的部分。TGower 盯着 stained glass 那段看了很久:
“That stained glass landscape video is pretty cool. I wonder if it would be possible to implement that effect in a real-time game engine.” — TGower [c:50261592] (译文:那段 stained glass(彩绘玻璃)风景视频挺酷。我在想这个效果有没有可能在实时游戏引擎里实现出来。)
pacman1337 的反应更短:
“this is the gta6 I am was hoping for” — pacman1337 [c:50261868] (译文:这就是我期待的 GTA 6。)
甚至那位批「廉价货电子游戏」的人,也给了自己的对照组:
“Oh yeah, don’t get me wrong “bargain-bin videogame” is a wild step up from “accidental amorphous horrors that defy description”. It just jumped out to me as looking unusually bad compared to the other samples they provided. The stained glass deer, for example, was awesome.” — OkayPhysicist [c:50262728] (译文:别误会,「廉价货电子游戏」相对「意外生成的、连形状都说不清的恐怖物」已经是天壤之别了。只是它比其他几个样片看起来格外差而已。比如那个 stained glass 鹿,就很棒。)
最后是一段关于 HN 本身的观察。empath75 一句话把这场讨论的性质点破了:
“It is always incredible to me how quickly we go from “this entire technology is pure science fiction” to “nitpicking over details”” — empath75 [c:50264539] (译文:让我一直觉得不可思议的是,我们从「这整个技术还纯属科幻」到「挑细节挑刺」的速度有多快。)
Pedro_Ribeiro 一句话把它归因到论坛本身:
“It’s not, it’s just that hacker news is a hyperbolic bikeshedding chamber.” — Pedro_Ribeiro [c:50264615] (译文:不是,只是因为 Hacker News 就是一个爱夸张的 bikeshedding 场(bikeshedding 指把讨论无限沉进细枝末节的闲扯场合)。)
典型观点一览
| 立场 | 用户 | 一句话 |
|---|---|---|
| 玩家操控反而不如 Genie 3 | sodacanner | 像是给 AI 视频贴了按键 |
| 原始 Genie 的目标是智能体学世界 | ainch | 使命的人走了,产品变成现在这样 |
| 这是给 LeCun 的「世界模型」挖坑 | jackb4040 | 名字背后是股价逻辑 |
| 叫「电子游戏模型」更准确 | Bjorkbat | 物理是游戏的物理 |
| 胸像该翻倒却悬在半空 | mlmonkey | 物理没建模 |
| 骑双足飞龙的骑士也算物理失败 | sobellian | 战斗物理也没建对 |
| 城堡右侧带投石机凭空消失 | smusamashah | 世界模型不该这样 |
| 还凭空多出一座城堡 | dantillberg | 进场时它还不在 |
| 整段像廉价货电子游戏 | OkayPhysicist | 翅膀穿过杆子没有碰撞 |
| 官方在演示 steering 而已 | dsissitka | 穿帮可以理解 |
| 现实本来就不完全按预期走 | peddling-brink | 机器人心智模型该有的样子 |
| 别当 Google 的辩护律师 | bigwheels | 胸像很少悬在 45 度角 |
| 游戏可以,训机器人不行 | modeless | 两种用途塞一个模型必败 |
| 世界模型 = 新一代游戏 DRM | rvz | Stadia 加世界模型的组合 |
| Stadia 死于延迟与画质 | verdverm | 我就是这么退订的 |
| Stadia 死于 GPU 改去跑 AI | ZiiS | 跟延迟无关 |
| Stadia 在 AI 之前就死了 | ActionHank | 你们都记错了时间线 |
| 云端生成算不过账 | perchedonstilts | 还是把代码写一遍更便宜 |
| 对公司有利,对玩家无益 | SomeonesAccount | 玩家不再拥有跑游戏的机器 |
| 纯粹是 IP 转卖 | moralestapia | 「蜘蛛侠拯救万圣节」提示词 |
| Stadia 当年的商业模式就不成立 | Matl | 游戏还得另买 |
| 游戏行业连常规 LLM 都没接住 | causal | 现在只是花哨技术演示 |
| 我家不买,我们等独立游戏 | qup | 麦子和糠要分清 |
| 重点是给机器人和自动驾驶做模拟 | empath75 | 一句话点题 |
| 实验室到现实的机器人训练链 | NitpickLawyer | 看起来像现实又能扩容 |
| 反乌托邦的商业用例 | paimapi | 请随这一代人退出而消失 |
| 零技术细节,连参数量都没有 | kouteiheika | 安全拉满、额度小到没用 |
| 这套要求本来就离谱 | estearum | 难为你还能兴奋 |
| Google 的模型入口体验一向差 | alex_sf | 不走 openrouter 我都懒得试 |
| 每月 200 镑内能试吗 | HughParry | 想要一个能玩的入口 |
| Gemini Ultra 便宜档是 75 镑 | msabalau | 价格当场被算出来 |
| 能导出引擎工程吗 | pdntspa | 还是只能在自己的运行时里 |
| 超过 30 秒要多少钱 | BatFastard | 开发者又怕又期待 |
| 世界模型这条路有个洞 | spwa4 | 模型会绕过模拟 |
| 目标是创作者工具就没有洞 | Closi | 看目标是什么 |
| 编码 agent 也是从聊天机器人长出来的 | sobellian | 起点不同不代表走不通 |
| 挪标准的是实验室自己 | culi | 我们只是在量他们的说法 |
| 彩绘玻璃那段确实酷 | TGower | 想在引擎里实现这个效果 |
| 这就是我期待的 GTA 6 | pacman1337 | 期待值直接拉满 |
| HN 挑刺速度惊人 | empath75 | 从科幻到抠细节只要一瞬 |
| 挑刺是 HN 自己的问题 | Pedro_Ribeiro | 这儿就是个夸张的闲扯室 |
总体情绪
这场发布最特别的地方,是它同时被两套标准验收。一套来自技术视角——原生音频、实时 steering、把 Gemini 拉进来做多步事件模拟,这些都算实打实的推进,不是给视频生成套个壳;另一套来自评论区,他们不否认这东西值得拆开看,但拆开之后满屏是穿帮。
两边的分歧点其实只有一句话:世界模型到底该按「模拟真实世界」的标准,还是按「好用的创作工具」的标准来验收。持前一种标准的人盯着胸像悬空、城堡消失、双足飞龙穿模,持后一种标准的人说画面漂亮、可 steering、能出原型就够了。而官方博客自己其实两边都承认了——它把「复杂物理交互、实时音频准确性、可读文字渲染、空间准确度、长期一致性」列为进行中的研究,又同时把生成电影感的镜头当成主要卖点。
真正被低估的那条线是 spwa4 提的洞:模拟会不会被绕过。如果这个洞成立,那么今天被逐帧验货的这些穿帮,其实是无关紧要的细节——因为下一个更强的模型会直接从视频帧和状态里学,根本不需要一个手工搭出来的、物理正确的世界。于是,一份 Bug 清单到底是在讨论一个产品的质量,还是在讨论这条技术路线本身的必要性。
引用帖子
| # | 标题 | URL |
|---|---|---|
| 1 | Genie 4: A new generation of world simulation(官方博客) | https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ |
| 2 | HN 讨论 | https://news.ycombinator.com/item?id=50261319 |
| 3 | Genie 3: A new frontier for world models(2025 年 8 月) | https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models |
| 4 | Genie 模型页 | https://deepmind.google/models/genie/ |
| 5 | How to create effective prompts with Genie | https://deepmind.google/models/genie/prompt-guide/ |
| 6 | Stadia 为何关闭(评论中引用的背景文章) | https://www.denofgeek.com/games/why-google-stadia-failed-shut-down-explained/ |
免责声明
本摘要由 AI 模型辅助生成:minimax-cn-coding-plan/MiniMax-M3