Google Genie 4 + Gemini Is Insane: Let’s Build 5 AI Worlds Live

Peter Yang

AI-assisted summary of Google Genie 4 + Gemini Is Insane: Let’s Build 5 AI Worlds Live — original source. Check the original for full context.

Source published: 2026-10-28

Key ideas

Google Genie 4 represents a major leap in world models by pairing next-frame visual prediction with native integrated audio synthesis, live dynamic prompting, and multi-step agentic planning powered by Gemini. (0:00)
Live interactive demonstrations reveal that while Genie 4 renders complex 3D environments from text prompts within roughly ten seconds, it frequently exhibits visual hallucinations and inconsistent physics rules. (2:25)
World models pose an immediate disruptive opportunity for film and VFX production by slashing rendering costs and iteration cycles, whereas replacing traditional video game engines remains limited by poor gameplay logic and hit registration. (9:32)
Genie 4 is currently assessed at a 'GPT-3.5 stage' of maturity, with full-fledged game and short film generation anticipated once models evolve to GPT-4 or GPT-5 tier fidelity. (12:25)

Detailed notes

## Background Published on YouTube by Peter Yang on October 28, 2026, this video provides an early-access look at Google's Genie 4 world model. According to the page description, Yang explains the underlying mechanics of world models, builds five live simulations—including a San Francisco kaiju attack, a 1980s Tokyo ramen counter, a flying dragon siege, an African safari, and an alien assault on New York—and provides an honest assessment of its implications and shortcomings for gaming and filmmaking.

## TL;DR Peter Yang demonstrates Google's Genie 4, illustrating how world models predict subsequent visual frames and synthesize native audio from user inputs rather than predicting text tokens. Across five live interactive prompts, Yang showcases features like dynamic environment editing and multi-step agentic planning while highlighting current flaws, including visual hallucinations, awkward sound effects, and missing collision mechanics. He concludes that while world models will rapidly disrupt VFX and film pipelines, the technology is currently at a raw 'GPT-3.5 stage' that cannot yet replace game engines, even if the act of building personalized virtual sandboxes may soon become a compelling form of play in its own right.

## Atmosphere & style The video carries an energetic, hands-on exploratory vibe, blending early-adopter enthusiasm with grounded realism. Yang's conversational walkthrough is punctuated by lighthearted amusement as he reacts live to the system's comical quirks, like a dragon greeting him with a human voice or a character calmly filming a kaiju disaster on a smartphone. The demonstration transitions smoothly from playful, curious experimentation into thoughtful, sober industry analysis about the genuine technical gap between generating visuals and designing fun gameplay.

## Bulletin - [00:00](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=0s) Google Genie 4 represents a major leap in world models by pairing next-frame visual prediction with native integrated audio synthesis, live dynamic prompting, and multi-step agentic planning powered by Gemini. - [02:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=145s) Live interactive demonstrations reveal that while Genie 4 renders complex 3D environments from text prompts within roughly ten seconds, it frequently exhibits visual hallucinations and inconsistent physics rules. - [09:32](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=572s) World models pose an immediate disruptive opportunity for film and VFX production by slashing rendering costs and iteration cycles, whereas replacing traditional video game engines remains limited by poor gameplay logic and hit registration. - [12:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=745s) Genie 4 is currently assessed at a 'GPT-3.5 stage' of maturity, with full-fledged game and short film generation anticipated once models evolve to GPT-4 or GPT-5 tier fidelity.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=0s) World models represent the next fundamental frontier in AI after LLMs, predicting subsequent visual frames based on user inputs rather than next text tokens. - [01:07](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=67s) Genie 4 integrates native audio generation, enabling interactive NPC speech and ambient soundscapes alongside real-time visual output. - [01:07](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=67s) Dynamic live prompting allows users to continuously manipulate running 3D simulations through text commands without reloading scenes. - [01:07](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=67s) Gemini coordinates behind the scenes to provide multi-step agentic planning for autonomous step-by-step executions within simulations. - [02:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=145s) First-person scene rendering from text prompts generates playable visual spaces in approximately ten seconds. - [04:00](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=240s) Physical interaction and visual logic remain prone to hallucinations, such as awkwardly stacked food bowls and inability to render characters eating. - [06:13](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=373s) Synthesized audio exhibits early-stage quirks, including fantasy creatures producing human greetings and grunts rather than authentic sound effects. - [07:54](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=474s) Dynamic behavioral logic struggles with context and interaction rules, such as predators ignoring nearby prey during prompted migrations. - [09:32](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=572s) Combat mechanics highlight a fundamental gap between generating high-end graphical output and executing interactive gameplay logic like hit registration and health pools. - [10:54](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=654s) VFX and film workflows stand to gain immediate efficiency improvements by allowing rapid prompt-driven iteration of complex scenes without dedicated render farms. - [10:54](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=654s) World models will not instantly replace game engines because compelling video games demand deterministic feedback loops, progression systems, and physics engines. - [11:46](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=706s) The creative act of prompting and personalizing bespoke virtual worlds may evolve into a form of entertainment comparable to playing traditional games. - [12:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=745s) Genie 4 is currently situated around the developmental level of GPT-3.5, marked by functional foundations mixed with visible bugs and object popping. - [12:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=745s) The arrival of GPT-4 and GPT-5 tier world models is projected to unlock end-to-end generation of coherent interactive games and cinematic short films from text.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=0s) World Models and What's New in Genie 4

Peter Yang introduces Google's Genie 4 as an early-access preview, framing world models as the primary frontier in artificial intelligence following the explosion of large language models. While conventional language models like GPT operate by predicting the next token or word in a sentence, a world model predicts the subsequent visual frame conditioned on user input. When an input key like `W` is pressed to move forward, the model projects what the environment should visually reveal next while retaining memory of previous states to maintain environmental consistency.

Genie 4 introduces three major advancements over prior systems: 1. **Integrated Audio:** The model natively synthesizes accompanying soundscapes, allowing non-player characters to speak interactively while generating environmental audio such as rushing wind, roaring engines, and contextual music. 2. **Live Dynamic Prompting:** Users can edit the running simulation on the fly (for example, typing a command like "shooting star across the sky"), and the engine updates the world in real time to incorporate the change. 3. **Multi-Step Agentic Planning:** Gemini works behind the scenes to coordinate sequential actions, enabling autonomous step-by-step executions such as a robotic arm carrying out a complex laboratory routine.

```mermaid flowchart TD accTitle: Genie 4 World Generation Flow accDescr: Diagram illustrating how Genie 4 creates and updates an interactive world loop using initial prompts, user inputs, and audio generation. subgraph Initialization ["Setup"] A["Text Prompts: Environment & Character"] --> B["First Frame Rendered"] end subgraph ExecutionLoop ["Interactive Simulation Loop"] B --> C["User Input: Keypress / Live Text Prompt"] C --> D["Gemini Multi-Step Planning Engine"] D --> E["Predict Next Visual Frame"] D --> F["Synthesize Native Audio & Voices"] E --> G["Buffer Memory: Retain Prior State"] F --> G G --> C end ```

As Peter Yang summarizes the paradigm shift: “A world model does the same thing, but instead of predicting the next word, it predicts the next frame based on your input.”

---

[02:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=145s) Scene 1: San Francisco Street Walk and Kaiju Attack

To demonstrate the system live, Peter Yang pastes text prompts defining the environment and character into Genie 4. Within roughly ten seconds, the simulation generates an interactive first-person view walking down a San Francisco street. The character holds a cup of coffee in one hand and a smartphone in the other while a classic cable car rattles by on the road.

Yang experiments with live environment alterations: - Directs a friendly golden retriever to appear, which walks up and sniffs the coffee cup. - Adjusts the ambient lighting to cast a warm, late-afternoon "golden hour glow." - Prompts an escalating kaiju assault: a giant beast emerges in the background tearing down city buildings, unleashes an atomic breath beam down the street near the player, and triggers a distant bridge collapse.

Throughout the devastation, the rendered character continues pointing their smartphone at the destruction. Yang remarks with humor on the stubborn realism of human behavior: “I guess this is the iPhone generation. So, even though there's a Kaiju, still want to record everything.”

---

[04:00](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=240s) Scene 2: 1980s Tokyo Ramen Stand

The second live scene moves to a rainy evening at a 1980s Tokyo ramen counter. Upon entering, the non-player character immediately speaks with synthesized voice acting: “Welcome in. Step out of the rain. What can I get started for you?”

Peter Yang prompts the chef to serve a bowl of ramen and show the character eating it. The model responds with voiced hospitality (“Here you go. Enjoy your meal.”), but exhibits visual hallucinations by stacking two bowls awkwardly on top of each other and failing to depict the physical action of eating. Yang continues modifying the atmosphere: - Spawns a cat curled up and sleeping near the counter. - Prompts the chef to pour a carafe of warm sake. - Triggers falling cherry blossom petals drifting past the counter. - Adds an atmospheric radio saxophone solo, which correctly introduces jazz music to the audio stream.

Reflecting on the quirks and potential of the tech, Yang observes: “As you can see, it's not perfect. I don't know why the bowls are stacked on each other, but this is the beginning of something that's really awesome.” Before proceeding, Yang gives a nod to video sponsor Wispr Flow, noting that he dictates his simulation prompts using the voice transcription tool.

---

[06:13](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=373s) Scene 3: Storming a Castle on a Flying Dragon

The third demonstration tests a high-fantasy scene reminiscent of *Game of Thrones*, casting the player as a Daenerys Targaryen-style dragonrider attacking a fortress that resembles King's Landing.

The live simulation generates dynamic aerial mechanics and combat effects: - Directing a "flame burst" causes the mount to breathe fire across the sky. - Generating a "swarm" triggers an incoming arrow storm from the ramparts. - Initiating a "dragon dive" drops the mount toward the defenses amid collision effects.

While the visual spectacle impresses, the synthesized audio reveals comical flaws: the dragon lets out a bizarre, synthesized “Hello” upon initial generation, followed by human-sounding grunts and groans during combat rather than authentic reptilian roars. As Yang remarks, “It doesn't sound like a dragon at all. It sounds like a guy pretending to be a dragon.” To stress-test the model's flexibility, Yang commands a sci-fi spaceship to warp in and battle the medieval beast; the craft briefly materializes in the sky, disappears, and flickers back into the horizon.

---

[07:54](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=474s) Scene 4: African Jeep Safari

Fulfilling a personal travel bucket list item, Yang sets up an African safari environment featuring a stationary Jeep surrounded by wildlife. The initial landscape populates with gazelles, elephants, and lions roaming the open savannah.

Yang tests dynamic animal behavior via text prompts: - Commands a herd of elephants to march across the plains. - Instructs a lion to track and catch an antelope. While the targeted antelope suddenly blinks into existence from off-screen, the lion follows with lackadaisical urgency. - Prompts a massive migration across the road directly in front of the vehicle. A large herd suddenly floods the screen, yet the lion casually strolls through the stampede without engaging, prompting Yang to joke that the predator must have already eaten and lost its appetite.

---

[09:32](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=572s) Scene 5: Alien Attack on New York & Industry Impact

For the final live trial, Yang prompts a gritty, first-person shooter perspective set during an extraterrestrial invasion of New York City, requesting an energetic rock battle soundtrack. The simulation loads alien invaders patrolling urban rubble while delivering menacing synthesized lines: - “Intruders will be purged.” - “You cannot escape the purge.” - “Crush the resistance. Target acquired.” - “The coordinate matrix is locked. Your world is forfeit.”

Yang attempts tactical gameplay commands—targeting aliens, calling in a mothership strike, triggering a helicopter crash into the foreground, and commanding the weapon to fire and eliminate enemies. However, the simulation struggles with core game logic: projectiles fail to register hits, alien entities do not take damage or despawn, and the interactive loops lack meaningful gameplay rules.

Assessing the tech realistically, Yang emphasizes the gulf between visual generation and true interactivity: “Right now, it's really good at generating graphics, but not quite good enough at generating gameplay.”

He breaks down the long-term ramifications across creative industries: - **Film and Special Effects:** World models represent an immediate breakthrough for VFX workflows. Where an intricate dragon siege historically demanded dedicated visual effects studios, extensive render farms, and months of labor, directors can now prompt, iterate, and generate multiple rough or usable takes in single passes, substantially lowering production costs. - **Gaming:** Yang pushes back against claims that generative world models will instantly replace game engines. Because compelling video games require tight interactive feedback, hit registration, physics rules, and deterministic mechanical loops, world models remain far from replacing standard game development suites. As Yang stresses: “Building a good game is much more than just a graphics.”

[11:46](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=706s) The Gap Between Scene Generation and Fun Games

While Google Genie 4 can generate dynamic 3D environments, Peter emphasizes that there is still a massive difference between generating interactive graphics and building an engaging video game. To make a game truly enjoyable, developers must design robust core loops, thoughtful progression systems, and refined gameplay mechanics that keep players invested. Peter candidly observes that the technology is far from delivering a fully realized, polished game experience right out of the box: “So at least right now, this stuff is not close to generating a fun game.”

However, Peter highlights an intriguing counterpoint regarding player psychology and creativity. As world models become faster, more capable, and capable of generating environments on the fly, the creative act itself may transform into the primary form of entertainment. In the same way that everyday users now leverage tools like Claude and ChatGPT to build custom personal apps tailored to their exact needs, future players may shift toward spinning up their own bespoke virtual experiences. Peter suggests that the process of prompting and constructing a customized world could easily rival traditional gaming: “As these world models get better and faster at generating scenes, you can argue that making your own game could become just as fun, if not more fun, than playing someone else's game, right?”

```mermaid flowchart TD accTitle: Evolution from LLM Personal Apps to World Model Personal Games accDescr: Diagram showing how personal app creation via LLMs mirrors the future of personal game generation via world models. subgraph LLMEra["Text & App Era (Current)"] direction TB A["Claude / ChatGPT"] --> B["Prompt-driven development"] B --> C["Personalized software & tools"] end

subgraph WorldModelEra["World Model Era (Emerging)"] direction TB D["Genie 4 & World Models"] --> E["Prompt-driven 3D scene generation"] E --> F["Personalized games & interactive media"] end

LLMEra -.->|"natural progression"| WorldModelEra ```

[12:25](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=745s) The "GPT-3.5 Stage" of World Models and What Comes Next

Offering an honest appraisal of the current state of Genie 4, Peter compares its developmental maturity to the early days of modern language models: “But if I'm being honest, I think this tech is probably at the I don't know GPT 3.5 stage.” The technology is visibly raw, exhibiting noticeable bugs and visual artifacts, such as assets and objects suddenly popping out of nowhere without clear physical continuity. Despite those rough edges, Peter finds the present baseline stunning, especially considering that the model can construct fully navigable, imaginative 3D spaces straight from text descriptions.

Looking toward the horizon, Peter projects that the real paradigm shift will arrive as world models mature to the equivalent of GPT-4 or GPT-5. At that inflection point, fidelity, physics consistency, and narrative logic will align closely enough to allow creators to generate complete short films and playable games entirely through prompt engineering. He concludes the walkthrough by encouraging viewers to keep an eye on this rapidly accelerating frontier, inviting them to like and subscribe for more hands-on tutorials as world models evolve.

- **Current maturity level**: Roughly comparable to the GPT-3.5 phase of large language models—functional and eye-opening, but unpolished. - **Current limitations**: Frequent visual bugs, rough object physics, and assets popping into scenes unexpectedly. - **Future milestone**: Advancing to the GPT-4 and GPT-5 tier of capability, unlocking text-prompted generation of full-length interactive games and cinematic short movies.

## Notable quotes > [11:46](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=706s) As these world models get better and faster at generating scenes, you can argue that making your own game could become just as fun, if not more fun, than playing someone else's game, right? — Peter Yang

> [09:32](https://www.youtube.com/watch?v=kxCrhwHvA_4&t=572s) Building a good game is much more than just a graphics. — Peter Yang

## Mentioned Google Genie 4, Gemini, GPT, Wispr Flow, Game of Thrones, Claude, ChatGPT, Peter Yang

## Video description Copy my best AI workflows to save time and automate busywork: https://www.behindthecraft.com 🤝 Subscribe to my practical AI newsletter read by 150K+ professionals: https://creatoreconomy.so

I think world models are the next big thing in AI after LLMs, and I got early access to Google's Genie 4. In this video, I explain how a world model works and what's new in Genie 4, then build 5 worlds live: a San Francisco street until a kaiju attacks, a talking ramen chef in 1980s Tokyo, a dragon storming a castle, a Jeep safari, and an alien attack on New York. I end with my honest take on what it means for movies and gaming, and where it still falls short.

What I cover: (00:00) Why world models are the next big thing after LLMs (01:07) How Genie 4 builds a world, and what's new (02:25) Walk a San Francisco street until a kaiju attacks (04:00) Order ramen from a talking chef in 1980s Tokyo (06:13) Storm a castle on a flying dragon (07:54) Go on a Jeep safari with lions and antelope (09:32) Defend New York from an alien attack (10:54) What Genie 4 means for movies and gaming

This video is sponsored by Wispr Flow. Get one month free: https://wisprflow.ai/?promo_code=PETERWISPRFLOW&months=1&utm_source=peteryang&utm_medium=sponsorship&utm_campaign=btc-q2-2026&dub_id=KbiIIkNXB4UHiAdU

Where to find me: X: https://x.com/petergyang LinkedIn: https://www.linkedin.com/in/petergyang/

Subscribe to this channel - more tutorials coming soon!

[Source](https://www.youtube.com/watch?v=kxCrhwHvA_4)

中文简报

谷歌 Genie 4 联手 Gemini 表现惊人：现场构建 5 个 AI 世界 Peter Yang 演示了谷歌的 Genie 4，展示了世界模型如何根据用户输入预测后续视觉帧并合成原生音频，而非预测文本标记。通过五个现场交互式提示，Yang 展示了动态环境编辑和多步智能体规划等功能，同时也指出了当前的缺陷，包括视觉幻觉、别扭的音效以及碰撞机制的缺失。他总结称，尽管世界模型将迅速颠覆视觉特效与电影制作流程，但该技术目前仍处于原始的“GPT-3.5 阶段”，尚无法取代游戏引擎，不过构建个性化虚拟沙盒这一过程本身可能很快就会成为一种极具吸引力的新玩法。谷歌 Genie 4 代表了世界模型的一大飞跃，它将下一帧视觉预测与原生集成音频合成、实时动态提示以及由 Gemini 驱动的多步智能体规划相结合。 (0:00) 现场交互演示显示，尽管 Genie 4 能在约十秒内根据文本提示渲染出复杂的 3D 环境，但它经常出现视觉幻觉和物理规则不一致的情况。 (2:25) 世界模型通过大幅削减渲染成本和迭代周期，对电影和视效制作构成了立竿见影的颠覆性机遇，但由于游戏玩法逻辑和碰撞检测表现欠佳，其取代传统视频游戏引擎仍面临局限。 (9:32) Genie 4 目前被评估为处于“GPT-3.5 阶段”的成熟度，预计在模型演进至 GPT-4 或 GPT-5 级别的保真度后，才能实现成熟的游戏和短片生成。 (12:25)

Reading summary

Peter Yang presents Google's Genie 4 preview as an early milestone in world models, shifting AI from predicting the next text token to projecting interactive, sequential visual frames and synthesized audio based on user actions. Core Mechanics of Interactive World Models (0:00) Traditional large language models predict the next word in a sequence, but world models like Google's Genie 4 predict subsequent visual frames conditioned on user input while retaining past states in a memory buffer to maintain continuity. When a user enters a directional command, the model renders the updated vantage point and projects how the environment responds. Genie 4 incorporates three notable architectural enhancements over previous iterations: native integrated audio synthesis that generates ambient soundscapes, contextual music, and non-player character speech; dynamic live prompting that modifies running simulations on the fly; and multi-step agentic planning coordinated by Gemini to manage sequential tasks like robotic routines. Live Environmental Generation and Simulation Quirks (2:25) Across live demonstration environments—including a San Francisco street, a 1980s Tokyo ramen counter, a medieval dragon siege, an African safari, and an alien invasion—Genie 4 rendered navigable 3D settings from text descriptions in roughly ten seconds. Users could alter active simulations by adjusting lighting to a golden hour glow, introducing interactive animals, or initiating dynamic events such as falling cherry blossoms and bridge collapses. However, these live trials exposed frequent visual and physical hallucinations. The simulation awkwardly stacked ramen bowls while failing to depict the physical action of eating, caused animals and spaceships to pop in and out of view, and yielded anomalous synthesized audio, such as a dragon greeting the player with a spoken 'Hello' and human-sounding grunts instead of authentic roars. Visual Synthesis Versus Deterministic Gameplay (9:32) Yang emphasizes a fundamental divide between rendering graphics and establishing playable game mechanics. In a first-person shooter trial set in New York, the model generated atmospheric visual wreckage and menacing synthesized alien dialogue, but completely lacked game logic: projectile inputs failed to register hits, enemies did not take damage or despawn, and core physics rules remained absent. Because of this limitation, Yang argues that world models cannot currently replace traditional game development suites, which require deterministic interaction loops, progression systems, and refined mechanical rules. In contrast, the technology offers immediate utility for visual effects and filmmaking workflows, enabling directors to iterate on complex sequences—such as dragon sieges—in rapid single passes without relying on lengthy studio render pipelines. The GPT-3.5 Stage and Future Trajectory (11:46) Yang assesses Genie 4's present maturity as equivalent to the 'GPT-3.5 stage' of language models: highly capable and indicative of a major paradigm shift, but raw and prone to continuity artifacts. Despite these rough edges, the creative act of prompting and shaping bespoke worlds on the fly may emerge as an entertainment medium in its own right, mirroring how users currently employ tools like Claude and ChatGPT to assemble custom applications. Yang projects that as world models progress toward the equivalent of GPT-4 and GPT-5 milestones, improvements in physical consistency, narrative continuity, and interactive rules will eventually allow creators to generate complete playable video games and cinematic short films directly from natural language prompts. Genie 4 establishes that real-time, text-driven 3D world generation is functional, but unresolved limitations in physical rules, hit registration, and asset continuity currently confine world models to visual prototyping rather than fully realized game engines.

精炼阅读摘要

Peter Yang 深入解析并实机演示了 Google 的世界模型 Genie 4，指出世界模型正在成为继大语言模型之后的 AI 新前沿，其核心范式已从单纯的文本标记预测转向基于用户交互输入的逐帧画面与音频推演。Yang 认为，尽管 Genie 4 在实时视觉合成与动态场景修改上表现惊艳，但该技术目前擅长的是图形渲染而非游戏机制，在物理确定性与核心玩法逻辑上仍处于早期阶段。世界模型运行逻辑与 Genie 4 核心升级 (0:00) Yang 指出，传统大语言模型（如 GPT）的核心逻辑是根据前序上下文预测下一个文本标记（Token），而世界模型则是根据玩家的操作输入（例如按下“W”键移动）实时预测下一帧连续图像，并借助状态记忆缓冲区维持所处环境的物理与视觉连贯性。与过往的系统相比，Genie 4 引入了三项关键技术突破：首先是原生音频集成，模型可直接合成环境声效、情景音乐以及具备交互对话能力的非玩家角色（NPC）配音；其次是实时动态提示词输入，用户能够在模拟运行过程中直接键入指令（例如“流星划过天空”），引擎将即时更新世界状态；最后是引入 Gemini 驱动的多步骤 Agent 规划，使系统能够自主编排连续任务动作，例如模拟机械臂在实验室中执行复杂的步骤化操作。跨题材实机测试：灵活生成与物理幻觉并存 (2:25) 在实机操作中，Genie 4 能够在约 10 秒内由文本提示词构建出可交互的三维空间。测试涵盖了多类复杂题材：在旧金山街头模拟中，用户能够实时添加迎面走来的金毛寻回犬、改变傍晚光照，并动态触发怪兽袭击与大桥坍塌，而第一人称角色甚至能保持手持智能手机记录灾难的逼真行为细节；在 1980 年代东京拉面店中，系统成功合成了老板的迎客语音、倒热清酒动作以及电台的萨克斯爵士乐背景声。然而，实机交互也暴露了显著的内容幻觉与物理因果断裂。在拉面店中，模型在响应上菜指令时错误地将两只碗重叠在一起，且无法展示进食动作；在飞龙攻城演示中，巨龙生成后发出了不合情理的人类语音问候，战斗受击时也仅发出人类哼唱声而非巨兽咆哮；在非洲野生动物场景中，狮子对近在咫尺的大规模迁徙毫无捕食反应，被追逐的羚羊则是突然凭空闪现。这些现象表明，当前系统在维持复杂的生态逻辑与基础物理交互常识上依然存在明显缺陷。视觉渲染与玩法逻辑的断层及产业冲击 (9:32) 通过纽约外星入侵的第一人称射击测试，Yang 阐述了当前技术在游戏应用上的根本瓶颈：系统能够迅速呈现巡逻的外星生物、呼叫母舰空袭和直升机坠毁画面，并合成带有威慑感的反派台词，但射击机制完全无法形成玩法反馈，发射的弹药缺乏命中判定，敌方单位也不会受损或消失。Yang 强调，“生成好看的画面”与“构建好玩的游戏”之间存在巨大差距，一款优秀的游戏依赖于严密的底层互动反馈、碰撞检测、物理法则与数值闭环，世界模型在现阶段根本无法取代标准游戏引擎。在产业应用前景方面，Yang 认为影视与视觉特效（VFX）制作将成为该技术的最直接受益领域。以往由专业特效公司投入庞大渲染农场和数月工期才能完成的复杂战斗镜头，导演如今可以通过自然语言提示在单次操作中快速生成多组动态构图与镜头草稿，从而大幅压缩前期概念探索与制作成本。 “GPT-3.5 阶段”定位与个性化创作变革 (11:46) 综合其实机体验，Yang 将当前的 Genie 4 定位为相当于大型语言模型发展史中的“GPT-3.5 阶段”。虽然仅凭文本描述就能生成完全可导航的三维想象空间令人震撼，但系统仍伴随着明显的视觉伪影、模型穿模和物体突兀闪现等瑕疵。Yang 预测，只有当世界模型演进至相当于 GPT-4 乃至 GPT-5 的成熟期时，画质精度、物理定律一致性与叙事因果逻辑才能真正统一，届时创作者将能够直接通过提示词生成完整的可玩游戏与电影短片。此外，Yang 提出了对未来玩家行为演变的预判：正如当今普通用户借助 Claude 或 ChatGPT 开发符合自身需求的个性化应用一样，当世界模型的生成速度与环境渲染能力达到全新临界点后，玩家通过提示词实时构建定制化虚拟世界的创作过程本身，很可能会超越单纯游玩他人开发的作品，转变为一种全新的主流娱乐方式。 Yang 总结指出，Genie 4 确立了世界模型在交互式场景合成领域的技术范式，但受限于底层物理逻辑缺失与交互判定失效，它目前更适于作为影视特效原型与创意构想工具；该技术能否真正演化为全功能的通用游戏引擎，仍取决于未来版本能否在因果逻辑与精确可控性上取得根本突破。

Source quotes

[11:46] Peter Yang: As these world models get better and faster at generating scenes, you can argue that making your own game could become just as fun, if not more fun, than playing someone else's game, right? [9:32] Peter Yang: Building a good game is much more than just a graphics.

Original transcript

[0:00] I think world models are the next big [0:03] thing in AI after LMS and I was lucky [0:06] enough to test Google's Genie 4 early. [0:09] Here's a quick clip of the worlds that [0:10] it can create complete with sound and [0:13] interactions. [0:16] [music] [0:35] Awesome. So, in this video, I'll first [0:37] explain how a world model works and [0:40] what's new with Genie4. Then we're gonna [0:42] create five worlds together live, [0:45] including a walk through San Francisco [0:47] until a kaiju shows up to attack, [0:50] ordering food in a Tokyo ramen stand, [0:53] storming a castle on a flying dragon, [0:56] chasing lions on a safari, and defending [0:59] New York from an alien attack. And at [1:02] the end of this video, I'll talk about [1:04] what I think this means for movies and [1:06] gaming. But first, what even is a world [1:10] model? So you probably heard that LMS [1:12] like GBT work by predicting the next [1:14] word. A world model does the same thing, [1:17] but instead of predicting the next word, [1:19] it predicts the next frame based on your [1:21] input. So for example, if you press W [1:24] here to move forward, then it'll [1:26] generate what it thinks you should see [1:28] next. So here how Genie 4 builds a world [1:32] in roughly five steps. So first your [1:36] prompt becomes the first frame. So here [1:37] it is. knight rides a black dragon past [1:40] the castle. Then you press a key like W [1:43] to move forward and then it predicts the [1:45] next frame for you and it remembers what [1:48] I already showed. So it repeats this on [1:50] a loop to create a complete scene. Okay, [1:53] so what's new with Genie 4? Well, [1:55] there's three things. First of all, [1:56] sound. It generates audio with the [1:59] world, so characters can talk back to [2:01] you and you can hear the engines, the [2:03] wind, and music. Second, you can make [2:06] live changes. So you can prompt the [2:08] scene and type something like shooting [2:10] star across the sky and the world will [2:13] adjust to account for that. And third, [2:15] Gemini 4 works behind the scenes to plan [2:19] multi-step events like this robot arm [2:22] working through a lab task step by step. [2:25] All right, but enough chatter. Let's [2:27] actually create some worlds together [2:28] live right now. Starting with a walk [2:32] down a San Francisco street and then a [2:34] monster attacks. So here I've included [2:37] the prompts for the environment and the [2:39] character. So let's go ahead and copy [2:41] this into Genie4 for the environment and [2:44] character right here. And let's hit [2:47] start generating the world. All right, [2:50] there we go. That generated in only [2:51] about 10 seconds. You can see I'm [2:53] holding my coffee and my phone and the [2:56] cable car just passed by. Let's see what [2:58] can we add here. Let's add a friendly [3:01] golden retriever. [3:05] All right, it's sniffing the cup. Let's [3:07] add a golden hour glow. [3:12] All right, so far things are pretty [3:13] peaceful here. But now, let's add our [3:15] kaiju. [3:17] Add a kaiju in the background, [3:19] destroying some buildings, and then do [3:21] an atomic breath down the street near [3:24] our person. [3:27] All right, here we go. It's going to get [3:28] a lot more chaotic. Here is our kaiju. [3:32] Let's do an atomic breath. [3:36] Yep. And starting to attack. [3:39] Um, let's do a bridge collapse. [3:43] And you know, this is like super fun, [3:44] right? It's almost like making our own [3:46] little Godzilla game here. And we can [3:48] create our own scene. [3:50] So, that is C1. Okay. And for some [3:52] reason, our tourist here is still [3:54] holding [screaming] the phone. I guess [3:55] this is the iPhone generation. So, even [3:56] though there's a Kaiju, still want to [3:58] record everything. But that's C1. Now [4:01] let's move on to scene two which is [4:03] ordering ramen at a 1980s Tokyo stand. [4:08] So once again I am going to copy the [4:10] environment and character into Genie 4. [4:13] So let's copy the environment. [4:15] Let's copy the character and let's hit [4:20] generate and let's see what it comes up [4:22] with. All right, here we go. [4:27] >> Welcome in. Step out of the rain. What [4:29] can I get started for you? [4:31] >> It's really awesome that the chef can [4:33] actually talk to us. What should we ask [4:35] for? Why don't we ask the chef, hey, can [4:39] you serve me a bowl of ramen with the [4:42] chef and also show me actually eating [4:44] the ramen? [4:50] >> Here you go. Enjoy your meal. [4:54] >> Okay. [4:57] Okay. [laughter] It served me two bowls [4:59] of ramen. I took the bowl, but it's not [5:01] showing me actually eating a ramen. [5:03] That's interesting. Why don't we do [5:04] something else? Let's add the cat. [5:10] Oh, there's the cat over there. It's [5:12] sleeping. Let's have the chef pour us [5:15] some warm sake. [5:17] All right, there we go. [5:20] And let's see if we can actually add [5:22] cherry blossom drift. That sounds [5:23] interesting. [snorts] [5:27] Okay, now it's a lot more beautiful. [5:29] Let's see. Let's add a radio sack solo. [5:31] Let's see if you can hear some music or [5:32] not. Ah, now I can hear music. [5:37] There you go. So, this is our ramen [5:39] scene in Tokyo. As you can see, it's not [5:41] perfect. I don't know why the bowls are [5:42] stacked on each other, but this is the [5:44] beginning of something that's [music] [5:45] really awesome. We can all create scenes [5:47] from our favorite trips and memories. [5:50] All right, so before we move on to the [5:51] next scene, I want to give a quick shout [5:53] out to our sponsor. You see me voice [5:56] dictating all the prompts to Genie 4 [5:58] using my favorite AI voice dictation [6:00] app, which is Whisper Flow. It is super [6:03] accurate and fast at transcribing my [6:05] voice. And you can check out in the link [6:07] in the video description to get one [6:10] month's free with my exclusive link. All [6:12] right, let's move on to our third scene, [6:16] which is storming a castle on a dragon [6:20] Game of Thrones style. All right, so [6:22] this is going to be very exciting. So [6:23] once again, let's go ahead and copy the [6:25] environment [6:27] and the character into Genie 4 and have [6:30] it generate a scene. All right, here we [6:32] go. Looks like we're Daenerys Targaryen [6:34] riding our dragon. [6:36] >> Hello. [6:39] There was a weird voice. And let's see [6:41] what can we do. Let's do a flame burst. [6:47] All right. Uh let's do a swarm [6:53] and then let's see. Oh my god, there's [6:55] an aerostorm coming. And let's do a [6:56] dragon dive. [7:02] All right, let's see what else can we do [7:04] here. Why don't we do an arrow clash? [7:09] [groaning] [7:11] >> [screaming] [7:14] >> I don't know why he keeps making Ra [7:16] sound. Doesn't sound like a dragon at [7:17] all. It sounds like a guy pretending to [7:18] be a dragon, but uh actually this this [7:21] castle looks suspiciously like King's [7:23] Landing a little bit. Let's see what [7:25] else can we do. I want to do something [7:27] crazy. Let's have a spaceship warping [7:29] and fight our dragon. [7:35] [screaming] [7:36] All right, there you go. There's a [7:37] spaceship. Uh it seems to have [7:39] disappeared. [7:40] But you can see here that um it's a [7:43] beautiful scene. I wish a little bit [7:44] higher resolution, but I think it looks [7:46] great. And yeah, that's kind of our [7:50] dragon attacking castle scene. And [7:52] there's a spaceship again. [7:54] All right. Now, now let's move on to our [7:57] next scene, [7:59] which is a jeep safari. You know, I've [8:01] always wanted to go to Africa, and maybe [8:04] one day I'll go when my kids are older. [8:05] But in the meantime, we can try to get [8:07] Genie4 to get us to go on a Jeep safari. [8:10] All right, so once again, I pasted the [8:12] prompts for the environment character [8:13] in. And let's wait for the world to [8:16] generate. All right, so here we go. Our [8:19] Jeep safari in Africa. Uh, doesn't seem [8:22] like our Jeep actually drives, but looks [8:25] like there's a bunch of elephants and [8:27] lions and gazels. Why don't we make the [8:29] elephants march? [8:33] >> [music] [8:36] >> All right. And why don't we actually [8:38] have the lions chase after some of the [8:41] antelopes. [music] So, let's say this. [8:43] Have the lion chase after and capture [8:45] one of the antelopes. [8:48] [music] [8:52] Okay, there's a lion. Um, I don't know [8:54] where the antelope is. It's over here. [8:57] Okay, the antelope came out of nowhere. [8:59] Oh, there's one running over here. [9:02] All right, let's see what else can we do [9:03] that's crazy. Let's just say have a [9:07] massive migration of antelope right in [9:11] front of the jeep. [9:17] All right, there we go. It came out of [9:18] nowhere and the lion is casually [9:21] strolling after the antelope. I guess it [9:23] already ate the antelope, so it's not [9:24] hungry anymore. But yeah, this is our [9:27] African safari scene. Not quite the same [9:29] as being Africa, but you know, it's fun [9:31] to play with. Okay, now let's move on to [9:34] our very last scene, which is aliens [9:36] attacking New York. [9:39] And this time, I'm going to ask for a [9:41] rock battle soundtrack and make it like [9:43] a firstperson shooter video game art. [9:46] All right, so it should be fun. Let's [9:47] try this. Copy the environment [9:50] and let's copy the character. Hit start [9:53] world. All right, here we are. So, we [9:57] have our aliens walking here. Let's see [9:59] if we can do anything. [10:00] >> Intruders will be purged. [10:04] >> So, we can actually hit target acquired. [10:05] Shoot. [10:06] >> You cannot escape the purge. [10:08] >> Crush them. [10:09] >> Voices. Let's do motherhip strike. [10:15] >> All right, let's see what else can we [10:17] do. [10:18] >> Crush the resistance. Target acquired. [10:20] All right, there's a weird voices [10:22] background as to have a helicopter crash [10:25] into the foreground. [10:32] That was the helicopter. [10:34] And uh yeah, that's basically our scene. [10:37] All right, so I said have my gun shoot [10:38] and delete the aliens. It didn't quite [10:41] >> The coordinate matrix is locked. Your [10:43] world is forfeit. [10:45] >> All right, it's not quite that [10:47] impressive. I mean, I think it generates [10:49] the graphics and stuff, but the gameplay [10:52] is not quite there, as you can see. [10:53] Right. All right, guys. So, I want to [10:55] show you guys the real deal. I don't [10:56] want to pretend that this tech is better [10:59] than it is. Right now, it's really good [11:01] at generating graphics, but not quite [11:03] good enough at generating gameplay, [11:05] which brings me to my two subjects. So, [11:08] how is this stuff going to change [11:09] movies? First of all, obviously, I think [11:11] the biggest use is for special effects. [11:14] You know, a dragon battle like this [11:16] typically used to take a visual effects [11:18] studio, a render farm, and months of [11:20] work to make. With a world model, a [11:23] director can type in a prompt and maybe [11:25] just get it to generate a few takes in [11:28] one shot, right? So, this could [11:29] dramatically reduce the price of big [11:31] effect scenes like this. And for gaming, [11:34] you know, many people are saying that [11:36] this stuff is going to replace gaming [11:37] entirely. But as you saw from our [11:39] first-person shooter demo, it's not good [11:42] enough at gameplay. And building a good [11:44] game is much more than just a graphics. [11:46] You need to have a core loop. You need [11:48] to have progression systems. And you [11:50] need to have other mechanics to make [11:52] your game actually fun. So at least [11:54] right now, this stuff is not close to [11:56] generating a fun game. But here's the [11:58] flip side. As these world models get [12:00] better and faster at generating scenes, [12:03] you can argue that making your own game [12:06] could become just as fun, if not more [12:08] fun, than playing someone else's game, [12:11] right? So, just like how everyone is [12:13] using AI to build personal apps, for [12:15] example, in clock or chatbt, we might [12:18] soon be approaching a time where people [12:20] are going to build all their personal [12:22] games using these world models. Awesome. [12:25] So, that is my quick video on Genie4. [12:27] But if I'm being honest, I think this [12:29] tech is probably at the I don't know GPT [12:32] 3.5 stage. It's still early. There's [12:34] still a lot of issues and bugs. Stuff is [12:36] popping out of nowhere. Although, it is [12:38] already creating pretty impressive 3D [12:40] scenes. Anything that we can imagine. [12:42] And you can imagine how this tech could [12:44] become a lot better once it gets to the [12:46] GPT4, GPT5 stage and actually allow us [12:50] to prompt to create short movies and [12:52] games. In the meantime, I'll be on the [12:54] lookout. And if you enjoyed this fun [12:57] demo and tutorial, please like and [12:59] subscribe for more videos like this. [13:02] I'll see you around.