返回首页

对比阅读

对比阅读:DeepMind's Genie World Model Replicated on One RTX 5090 — Compute Barrier Cracks 与 DeepMind 世界模型 Genie 被开源者复刻到单张 5090 — 造游戏世界的算力门槛被打下来

AEN
DeepMindGenieRTX 5090·

DeepMind's Genie World Model Replicated on One RTX 5090 — Compute Barrier Cracks

This week on Reddit's LocalLLaMA programming forum, we spotted a trending post: developer juanviera23 ran Google DeepMind's Genie world model on a single RTX 5090 (19GB VRAM used), hitting 720p at 16 FPS — worth flagging that this is the first time a world model has left the data center and reached consumer-grade hardware.

What This Is

Genie, released by DeepMind in early 2024, is a "world model" — AI that generates interactive virtual spaces in real time from text prompts or video, functioning like a game engine powered by imagination. DeepMind's original demos required cloud clusters, and the industry consensus at the time was that replicating it would need at least 8 H100 GPUs — a toy exclusive to big labs.

juanviera23's open-source release runs on a single NVIDIA RTX 5090 (consumer flagship, retail price around ¥10,000+ RMB), using only 19GB of VRAM (out of 32GB on the card) to produce playable 720p output at 16 FPS. In other words, a ¥20,000 PC can now let AI build worlds locally and in real time.

Industry View

Supporters see this as a landmark moment for the open-source community. From "8x H100" to "one consumer card," the replication cycle came in under six months; game developers are already discussing using it for prototyping — a level demo that used to take a month may now be possible in a day.

But the counterarguments are equally clear. First, 16 FPS is still short of "actually fun" — DeepMind's original at 30 FPS was already considered borderline. Second, world-model-generated content lacks coherence: the AI doesn't "remember" the characters it created in previous frames, and long-term consistency is a widely acknowledged hard problem. Third, beyond gaming, what Silicon Valley is really betting on is synthetic data — using world models to generate training scenarios for robotics and autonomous driving, a commercialization path much closer to reality. We've also noted another divergence: open-source able-to-run does not equal productization; usability, ecosystem, and compliance are the real hurdles ahead.

Impact on Regular People

For enterprise IT: No immediate action needed. If you're already doing local AI inference, the hardware threshold for adding a Genie-style model is just one 5090; what needs real evaluation is the stability and maintenance cadence of open-source inference frameworks.

For individual careers: Design, product, and copywriting roles can start watching for proof-of-concepts — future interactive prototypes might be built directly from prompts, eliminating the back-and-forth with engineering teams.

For the consumer market: We're still in the "it runs" demo phase; "you can't stop playing" is likely half a year to a year away.

BZH
DeepMindGenieRTX 5090·

DeepMind 世界模型 Genie 被开源者复刻到单张 5090 — 造游戏世界的算力门槛被打下来

本周我们在 Reddit 编程论坛 LocalLLaMA 上看到一条热帖:开发者 juanviera23 用一张 RTX 5090(显存占用 19GB)跑出了 Google DeepMind 的 Genie 世界模型,720p、每秒 16 帧——值得标记的是,这是世界模型第一次离开数据中心,进入消费级硬件。

这是什么

DeepMind 在 2024 年初发布的 Genie 是一种「世界模型」(World Model)—— AI 能根据文字提示或一段视频,实时生成可交互的虚拟空间,像一个靠想象运转的游戏引擎。DeepMind 原版演示需要云端集群,业内当时普遍判断复刻它至少要 8 张 H100,是大实验室的专属玩具。

juanviera23 这次放出的开源版本,单张 NVIDIA RTX 5090(消费旗舰,零售一万多元人民币),显存只用 19GB(卡上总共 32GB),就能跑出 720p、16 帧的可玩画面。也就是说,一台两万元的 PC,现在可以本地让 AI 实时造世界。

行业怎么看

支持方认为这是开源社区的标志性时刻。从「8 卡 H100」到「1 张消费卡」,复刻周期不到六个月;游戏开发者已经在讨论用它做原型——以前一个月才能出的关卡 demo,现在可能一天就能跑起来。

但反对意见同样清晰。第一,16 FPS 离「真正好玩」仍有距离,DeepMind 原版 30 FPS 也算勉强;第二,世界模型生成的内容不连贯,AI 不会「记住」自己上一帧造的角色,长期一致性是公认的硬骨头;第三,比起游戏,硅谷真正押注的方向是合成数据——用世界模型给机器人和自动驾驶造训练场景,这条赛道的商业化更近。我们也注意到另一种分歧:开源能跑不等于产品化,可用性、生态、合规才是后面真正的坎。

对普通人的影响

对企业的 IT 部门:短期不必行动。如果已经在做本地 AI 推理,加这张 Genie 模型的硬件门槛就是一张 5090;真正要评估的是开源推理框架的稳定性与维护节奏。

对个人职场:设计、产品、文案岗位可以开始留意概念验证——以后做可交互原型,也许直接用提示词,省掉和工程团队反复对齐的来回。

对消费市场:当下还是「能跑」的展示阶段,距离「让你玩得停不下来」估计还要半年到一年。