Luma AI Chief Sees Video Models Driving Robots as Text Data Depletes

文本数据趋于枯竭，Luma AI负责人视视频模型为机器人驱动引擎 生成式视频将通过直接对物理现实建模来克服人类文本数据的枯竭，从而为机器人认知赋能并取代传统媒体制作。 Amit Jain 阐述了 Luma AI 的愿景：超越流于表层的“机械式转换”视频生成工具，演进为具备真正物理与空间推理能力的统一多模态模型。他将 LLM 面临的文本数据即将枯竭的危机与视频数据的充沛性进行了对比，指出生成式视频可充当具身机器人所需的基础世界模拟引擎。Jain 还谈及制作成本的断崖式下跌——单个视觉特效镜头的成本从 10 万美元骤降至 10 美元的算力费用，并指出社会信任将逐步转向围绕内容溯源而非未经验证的视觉证据来校准。展望未来，Luma 设想以实时、个性化的媒体生成取代传统的静态视频检索流，使其规模扩大到每天数十亿小时。 Luma AI 创始人 Amit Jain 指出，真正的 AGI 必须跨越视频、音频等多感官模态进行学习；生成式视频不仅能颠覆好莱坞每年 3000 亿至 3500 亿美元的实体制作成本，还将在“机器人大脑假说”下为具身智能提供不可或缺的物理世界模拟与轨迹规划能力。 (2:26) 目前的视频生成工具仍是缺乏上下文记忆的“笨拙翻译模型”，每次生成如同“老虎机”般随机且无法保持角色与场景连贯；行业真正的突破需要将高层语言因果逻辑与视频底层物理信号深度融合。 (8:59) 大型语言模型当前的放缓根源在于人类文本数据的枯竭，而视频数据所蕴含的物理规律与空间信息量比文本高出数十亿倍，为模型训练提供了近乎无限的数据前沿。 (16:07) 当生成单个镜头成本从 10 万美元降至约 10 美元时，自动化深度伪造检测终因视觉瑕疵的消除而失效；社会公众将建立“未经证实的视频不再是客观事实”的新认知习惯，并转向依赖密码学渠道验证身份真实性。 (19:34)

Amit Jain, CEO of Luma Labs, argues that artificial general intelligence requires moving beyond text-only models toward native multimodal world simulation trained across text, audio, images, and video. In his view, this paradigm shift will collapse visual production costs, provide embodied robotics with internal world models, and require society to re-evaluate visual media as proof of reality. Overcoming Text Limits and Architectural Amnesia (8:59) Jain contends that the recent slowdown in large language model improvements is driven by the exhaustion of unique human-written text rather than compute limitations. In contrast, the physical world offers an abundant volume of video data, which Luma ingests through web collection and studio partnerships with Hollywood archives and advertising agencies. However, Jain cautions that existing visual tools—including Luma's current models—remain fundamentally "dumb translation models" that map text tokens directly to pixel arrays without true contextual reasoning. Because current visual generators lack persistent state tracking, sequential five- to eight-second clips suffer from severe narrative amnesia, resetting unpredictably like a casino "slot machine." Existing systems cannot reliably remember character continuity, emotional states, or consequences across scenes. To move past these limitations, Luma is developing native mathematical architectures intended to unify text, audio, and visual dimensions into a coherent multimodal reasoning system rather than relying on fragmented translation pipelines. Physical World Simulation for Embodied Robotics (2:26) Beyond digital content generation, Jain identifies the "Robot Brain Hypothesis" as a vital application for generative video. Modern robotics frameworks predominantly rely on brute-force teleoperation action datasets and lack an internal model of the physical environment. By contrast, human cognition routinely runs predictive mental simulations of spatial trajectories before taking action, such as imagining the path between rooms. Jain argues that video generation models that grasp physical laws, causality, lighting, and spatial constraints serve as the missing internal world models necessary for robots to plan and navigate physical reality. He forecasts that video generation models will require approximately one more year of scaling to achieve full physical accuracy and robust visual chain-of-thought reasoning. The Economics of Production and Compute Density (27:09) Humanity's information consumption has decisively shifted toward visual media, with Hollywood alone spending between $300 billion and $350 billion annually on physical production constrained by actors, sets, lighting, and logistical overhead. Generative video transfers this physical production into software, lowering shot costs from $100,000 Hollywood visual effects pipelines down to approximately $10 in compute. However, video compute remains uniquely resource-intensive: generating a five-second clip requires 300,000 to 400,000 token equivalents, comparable to the compute footprint of an entire novel. Looking ahead, Jain envisions transforming video distribution from pre-recorded database retrieval platforms into personalized, real-time generation delivering 6 billion hours of video globally each day. At current Ray 2 inference rates, generating that daily volume would cost an unsustainable $280 billion per day, necessitating significant architectural and hardware optimization. Epistemic Calibration Beyond Flawed Technical Detection (19:34) Addressing societal concerns over deepfakes and misinformation, Jain rejects upstream legal bans and automated detection algorithms as durable solutions. Because generative models are systematically optimized to eliminate synthetic artifacts to satisfy the visual Turing test, automated detectors will inevitably fail as architectures mature. He argues that blanket governmental prohibitions are unenforceable globally and risk infringing upon free speech. Drawing historical parallels to the adoption of the printing press and photo editing tools, Jain asserts that society must undergo an epistemic adaptation: audiences must abandon the assumption that seeing something on video guarantees it occurred. Verification will increasingly rely on cryptographically authenticated distribution channels rather than raw visual evaluation. Regarding legal conflicts over training data, Jain frames the dilemma as a fundamental choice between adhering to 1920s copyright frameworks or prioritizing the progression of artificial intelligence. While generative video models offer a path toward unified physical world simulation and low-cost production, realization of Jain's long-term roadmap faces major engineering hurdles. Transitioning to real-time generative feeds depends on solving severe inference compute costs, while society and legal frameworks must simultaneously adapt to an information environment where visual evidence can no longer be accepted at face value.

Luma AI 联合创始人 Amit Jain 认为，大语言模型正面临人类文本数据枯竭的瓶颈，而原生多模态视频生成模型通过学习物理规律与空间动态，不仅将颠覆影视与商业营销的生产成本，还将作为构建具身智能“机器人大脑”的基础世界模型。 超越文本：多模态智能与机器人大脑假说 (0:04) Jain 指出，当前人工智能领域将庞大资源倾注于以文本为主的大语言模型，但人类大脑并非仅依靠语言推理，而是整合视觉、听觉等感官流来理解物理环境。视频与图像承载了物理行为、因果关系、运动动态与环境约束，而现有文本模型无法捕捉这部分认知。Luma 的技术路线是从文本、图像、视频和音频四种模态中同步训练，研发如 Ray 2 等模型，并从各模态隔离管线演进为原生统一智能。除了面向消费者的视频生成与动态视觉问答外，Jain 提出了“机器人大脑假说”（Robot Brain Hypothesis）：现存具身机器人系统多依赖遥操作行为数据集，缺乏内在世界模型；具备因果与空间规律理解能力的视频生成模型，能充当机器人在物理世界中规划、构想轨迹所需的内部模拟引擎。Jain 预计，视频生成模型大约还需要一年的规模扩展，才能实现完全的物理精确度与稳健的视觉思维链推理。 视频生产的经济重构与实时生成愿景 (2:26) 从经济规模来看，好莱坞仅过去一年的实体视频制作支出即达 3000 亿至 3500 亿美元，多个主流数字平台依托视频汇聚了超过 30 亿月活跃用户。Jain 将视频变革比作文字从活字印刷走向字处理软件的数字化历程。传统视频受限于演员、场地、灯光等实体“原子”与高昂的资金门槛，生成式视频则将物理现实搬入笔记本电脑中。在单位成本上，传统好莱坞特效镜头制作成本动辄数万至 10 万美元，而生成式计算将单镜头成本压缩至约 10 美元。针对视频生成工具每月约 200 美元的定价争议，Jain 解释其背后是高密度的算力消耗：生成一段 5 秒的视频需要处理 30 万到 40 万个等效 token，相当于一本长篇小说的计算量。在消费端，用户目前在手机端平均每日观看视频 3.5 小时，像 TikTok 等现有平台实质上是在预录数据库中进行静态检索。Luma 的长远目标是将静态检索替换为实时个性化生成，实现全球人均每日生成 1 小时定制视频（日均生成总量约 60 亿小时）。按照 Ray 2 模型目前的推理成本计算，这一体量每日将耗费 2800 亿美元，因此团队正专注于算法效率提升与硬件加速。 克服失忆架构与数据枯竭：从转译走向统一推理 (8:59) Jain 认为当前第一代视觉生成工具（包括 Luma 现有模型）本质上是“愚蠢的转译模型”，仅机械地将文本 token 映射为像素阵列，缺乏真正的上下文推理与持久记忆。这类工具在生成 5 至 8 秒的连续镜头时会出现严重的“叙事失忆”，无法追踪角色的死活状态、情绪连续性或色调搭配，使用体验类似不可预测的“老虎机”。其在苹果开发 Vision Pro 空间摄影与 3D Persona 头像的经历表明，真正的多模态模型必须发明全新的数学底层，将高维的高帧率视频、静态图像、音频和文本统合，直接对光线传输、物理动态与空间运动展开原生推理。在数据端，大语言模型的性能瓶颈源自人类原创文本数据的枯竭，被迫转向合成数据；视频领域则蕴含数十亿倍于文本的信息密度。Luma 除了抓取网络数据与内部私有采集外，还与好莱坞主流制片厂及全球大型广告机构合作，调取其沉淀数十年的历史高清影像档案，因而未受数据墙制约。 深度伪造、认知校准与过时版权法之争 (16:07) 针对伴随超逼真视频而来的深度伪造、虚假信息传播以及数据版权摩擦，Jain 表达了鲜明立场。面对数据训练相关的法律诉讼，Jain 认为社会正面临在 20 世纪 20 年代过时版权法与人工智能技术进步之间的二元抉择，并主张技术的规模化发展应当优先于陈旧的法律限制。在虚假内容治理上，Jain 强调通过源头禁令或政府封禁生成技术在技术上不可行且有损言论空间，单纯依赖自动化检测算法也注定会失效，因为生成模型的演进目标正是彻底消除机器伪造痕迹（例如与相机 CMOS 传感器不符的噪声模式）。正如印刷机催生大众媒体后公众学会不轻信文字、Photoshop 普及后公众学会质疑图像真实性一样，视频领域正在经历认知校准（Epistemic Calibration）。当单镜头成本降至 10 美元且视频可随意编辑生成时，公众将建立起“眼见不再为实”的新共识，未来个人与公众人物的真实性确权将主要依赖可加密验证的发布渠道，而非视频画面本身。 Jain 强调，视频是抢占大众注意力的终极媒介，生成式 AI 将使中小企业摆脱昂贵预算限制、完成高水准叙事。然而，日均生成 60 亿小时个性化视频的高昂推理成本、跨镜头长期物理记忆的架构突破，以及全球法律与公众认知在新媒介环境下的再适应，仍是该技术路径必须跨越的现实挑战。

## Background 该视频由 Ai4 频道于 2025 年 8 月 28 日在 YouTube 平台发布。内容记录了 Luma AI 创始人兼 CEO Amit Jain 与 Fireworks AI 开发者关系主管 Aishwarya Srinivasan 的炉边对谈，探讨生成式 AI 如何超越文本与图像，迈向能够理解物理世界运作法则的多模态世界模型。 ## TL;DR Amit Jain 阐述了 Luma AI 的核心愿景，即超越当下机械的“文本到像素翻译”，构建兼具因果推理与空间感知的统一多模态“机器人大脑”。他指出视频数据的丰富度远超纯文本，面对版权与 Deepfakes 挑战，扩展智能与重塑公众认知标尺才是关键，而非源头封杀。尽管视频生成受制于极高的 Token 密度与算力成本，Luma 仍在通过统一底层数学架构攻克难题，并展望了未来全球每日实时生成 60 亿小时视频、颠覆传统流媒体与企业叙事营销的宏大前景。 ## Atmosphere & style 整场对谈兼具物理学视角的硬核逻辑与直言不讳的极客幽默，Amit Jain 坦率剖析当前模型的笨拙局限，并运用“拉老虎机”、“原始人卖轮子”等生动比喻化解复杂命题。两位机器学习从业者的对话节奏明快、共鸣强烈，从空间智能前沿层层推演至现实商业经济学，给人一种身处多模态范式革命前夜的智识激荡感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=4s) Luma AI 的愿景不只是文本生成视频，而是融合音频、视频、图像与文本，构建能够学习物理世界运动规则与因果关系的通用多模态智能体。 - [03:11](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=191s) 视频制作正经历从昂贵物理供应链（原子世界）向数字化模拟的范式转移，视听信号的高带宽特性决定了视频已成为互联网最高效的信息与知识载体。 - [05:38](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=338s) 视频生成模型不仅平民化了影视创作，还提出了“机器人大脑假说”——视频学习到的物理规律与空间因果表征将直接充当通用具身机器人的视觉大脑。 - [09:47](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=587s) 当前多数视频模型本质上只是“文本到像素的翻译工具”，缺乏上下文记忆与因果连贯性；Luma 致力于赋予模型深层逻辑推演与多模态世界理解能力。 - [15:03](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=903s) 与文本 LLM 逐渐面临数据枯竭不同，现实物理世界的视频数据量是文本的数十亿倍，结合与好莱坞制片厂及广告公司的档案合作，视频生成拥有充足的数据供给。 - [19:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1148s) 在扩展智能与死守传统版权法之间，扩展智能更具社会价值；面对 Deepfakes，源头封堵并不现实，关键在于推动公众建立“眼见不一定为实”的社会心理预期与认知标尺。 - [23:16](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1396s) 从源头立法严管视频生成极易侵犯言论自由，且在全球地缘政治与硬件普及背景下难以奏效，社会自发的心理适应是最有效的防御机制。 - [25:39](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1539s) Luma 正在研发全新的统一多模态底层数学架构，打破文本、图像、音频和视频的分立模型体系，实现原生跨感官的视觉与综合空间推理。 - [28:33](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1713s) 虽然 5 秒视频包含 30-40 万 Token 计算量使单次生成成本高于文本 LLM，但相较传统影视拍摄的高昂实体开销，AI 视频生成的边际成本已极度低廉。 - [29:41](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1781s) Luma 的终极愿景是将内容消费从“预录制检索推荐（如 TikTok）”彻底颠覆为“全球每天 60 亿小时全实时生成视频”，这要求大幅降低算力与推理成本。 - [31:06](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1866s) 任何企业的第一要务都是讲好自己的品牌故事，生成式视频技术打破了顶级制作的高门槛，使中小企业能够以极低成本占据受众的视觉注意力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=4s) 多模态智能的愿景：从文本 LLM 走向物理世界理解 对谈伊始，主持人 Aishwarya Srinivasan 谈及自己作为机器学习从业者的背景，并引出对当前视频生成领域的诸多好奇与未解之问。Luma AI 创始人兼 CEO Amit Jain 拥有物理学背景，他创办 Luma 的初心并非仅停留在“文本生成视频”这一单点功能上，而是希望构建能够全面消化人类数字化足迹（音频、视频、图像和文本）的通用智能体。 Amit 指出，目前整个 AI 领域的核心注意力依然高度集中在纯文本大语言模型（LLM）上，文本模型在逻辑推理与代码编写等任务中确实取得了巨大飞跃。然而，“人类大脑并非仅仅通过语言进行推理，而是在所有感官模态之间协同运作”。Luma 正在探索的下一步，是通过视频和图像去学习物理世界的运行法则：物体如何运动、事物如何相互作用（因果关系）、构图美学以及声音如何在空间中传播。视频模型只是 Luma 交付的第一个阶段性产品形态，其发布的 Ray 2 模型已经具备跨视频、图像和文本的联合生成能力，未来的目标则是将这些跨模态感知与认知彻底融合成单一的统一智能模型。 ```mermaid flowchart TD accTitle: Luma AI 多模态智能学习与演进路径 subgraph S1["人类数字足迹输入"] A["文本 (Text)"] B["图像 (Images)"] C["音频 (Audio)"] D["视频 (Video)"] end subgraph S2["核心世界认知学习"] E["因果规律与物理世界建模"] F["空间结构与视觉构图"] G["动态演变与声学感知"] end subgraph S3["落地形态与产品"] H["Ray 2 跨模态生成模型"] I["统一多模态智能体 / 机器人大脑"] end A & B & C & D --> E & F & G E & F & G --> H --> I ``` - **人类感官全模态学习**：涵盖音频、视频、图像及文本，跳出单纯以语言符号为核心的学习范式。 - **物理与因果理解**：通过动态视频数据直接提炼物理运动规则、空间因果关系及视觉构成。 - **产品演进路线**：现阶段的 Ray 2 已实现多模态联合生成，后续将朝向高度统一的世界智能体演进。 --- ### [03:11](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=191s) 视频作为互联网核心媒介：从“原子世界”向数字模拟的万亿级迁移 针对视频生成是否仅是“炫酷 Demo”或“创作者玩具”的实用性质疑，Amit 从宏观经济与人类信息交互带宽的底层逻辑给出了剖析。仅好莱坞一地，每年的制作预算就高达 3000 亿至 3500 亿美元；若算上全球范围内的商业广告、企业培训和社交媒体内容，视频产业的整体盘子极其庞大。人类对视频内容的渴望近乎无休止，在阅读与观看之间，绝大多数人天然偏好视频，其根源在于人类的眼睛和耳朵相比于处理纯文字是“带宽极高得多的输入通道”，阅读需要消耗巨大的认知能量。 从互联网信息载体的演进来看，2000 年代和 2010 年代主要由纯文本主导，而如今全球已有数个突破 30 亿月活跃用户的超级平台完全由视频驱动，甚至像 Excel 教学这样原本枯燥的技能如今也在 TikTok 视频中被广泛传播。然而，当前视频制作的本质依然是一个完全依赖“原子世界（Atoms）”的繁重物理过程：需要搭设布景、调校灯光、搭建真实环境、雇佣演员及导演调度。正如活字印刷术演进到文字处理器，再到如今大语言模型瞬间为单一用户生成定制文本一样，视频制作也必然会经历从物理供应链向纯数字化软件生成的彻底转变。 - **庞大的内容制作开销**：好莱坞年度制作支出高达 3000 亿至 3500 亿美元，且全流程受限于昂贵的物理实体物流。 - **生物感知的高带宽特性**：视听信号的信息吞吐带宽远高于文本符号，决定了用户消费内容的底层偏好。 - **互联网信息载体范式转移**：三大月活超 30 亿的超级平台完全以视频为核心资产，视频已成为互联网最高效的知识与娱乐载体。 --- ### [05:38](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=338s) 视频模型的落地演进：创作平民化、视觉问答与“机器人大脑假说” Amit 阐述了将物理世界数字化模拟的终极价值——“将整个世界装进你的笔记本电脑，以便你可以随心所欲地操纵它”。在传统影视制作中，由于产能供给受限，创作者必须获得少数掌权者的准入许可并依赖庞大预算与摄制团队。生成式视频模型打破了这一准入门槛，让任何人都能在个人设备上完成电影制作、故事叙述或知识传播。视频生成的图灵测试远比单维度的文本要困难得多，它要求模型在时间序列上准确维系微表情、物理因果与情感连贯性；随着模型正在迈入类似大语言模型的“Scaling Laws（缩放定律）”加速爬坡期，Amit 预计大约再过一年左右，视频模型就能在物理规律与思维链推进上实现高度准确。 除了影视娱乐与广告营销的重塑，Amit 还指出了视频世界的另外两大核心应用场景： 1. **即时视觉问答（Visual Q&A）**：很多向 ChatGPT 提出的现实问题本质上更适合通过即时生成的动态视觉演示来解答，而非苍白的文字。 2. **“机器人大脑假说（Robot Brain Hypothesis）”**：人类在行动前会在大脑中进行高度空间化的视觉规划（例如闭眼想象从大门走到卧室的路径，除非患有心盲症 Aphantasia）。而当今的大多数机器人仅仅依赖遥操作（Teleoperation）采集动作数据进行暴力拟合，大脑中缺乏对物理世界的内部先验模型。视频生成模型所学习到的物理规律和因果表征，未来将直接构成通用机器人的“视觉空间大脑”。 - **消除创作门槛与审批壁垒**：摆脱摄制组与实体器材限制，在单机上实现自由可控的高质量视觉叙事。 - **技术成熟时间表**：随着缩放定律生效，模型正处于能力拐点，预计一年左右即可达到完全合乎物理真实与长线逻辑的水平。 - **机器人大脑核心先验**：解决遥操作动作克隆缺乏世界常识的弊端，通过视频预训练为具身智能注入内部世界模型。 --- ### [08:59](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=539s) 具身世界模型与视频生成图灵测试 Aishwarya 对 Amit 提出的“机器人大脑假说”表示强烈共鸣，并提到这一观点与当天上午李飞飞（Fei-Fei Li）在大会主题演讲中所强调的空间智能理论高度契合。视频生成不仅仅是画面渲染，更被学术界与工业界视为考验人工智能整体认知水平的“全新图灵测试”。 Aishwarya 回顾了生成式 AI 的演进历史：正如早期的 2D 图像生成模型一度跨不过“画手和多余手指”的死穴一样，早期的视频生成模型在处理复杂动态运动时也经常出现违背生理学常识的荒谬画面（例如在生成体操视频时，运动员的腿部会在空中产生 360 度的诡异翻折）。但随着模型架构与训练数据的跨越式发展，视频 AI 正在以惊人的速度跨越这一瓶颈曲线，逐步建立起对真实物理与生物运动规律的精准理解。 ### [09:47](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=587s) 重新定义视频生成：从“翻译模型”迈向真正的“智能” 谈及 Luma Labs 的核心优势与独特之处，创始人 Amit Jain 指出，Luma 与市面上绝大多数视频生成公司最大的不同在于：**Luma 并不把自己定位为一家单纯的“视频生成公司”**。他直言，无论是图像、视频还是文本生成，孤立地看待某一种模态本身就是一种局限。 为了更形象地说明这一观点，Jain 极其坦率地评价道：“当前的视频生成模型——包括我们自己的——都极其蠢笨（dumb as like...）。” 它们与拥有上下文记忆能力的语言模型（LLM）截然不同。在与 LLM 交互时，模型能够保持多轮对话的上下文逻辑，甚至进行自主推演并提出建议；而目前的图像和视频生成模型**仅仅是“文本到像素的翻译模型”**。Jain 强调，将文本翻译成像素根本不是真正的智能，这就好比当年 Google Translate（谷歌翻译）刚推出时，虽然对联合国翻译员等特定岗位造成了冲击，但没有任何人会认为那是通用人工智能（AGI）或真正的 AI。 Jain 补充道，当前的视频模型每次生成内容都像是在演《海底总动员》里失忆的“多莉”（Dory），每次接收 Prompt 都要“从零开始”。例如在电影制作中，如果上一幕某个角色已经死亡，后续场景中如果该角色再次出现，模型必须能够判断这是闪回（flashback）还是变成了鬼魂；如果两个角色刚经历过一场大吵，下一幕绝不应该像多年老友般热情打招呼。这种色彩连贯性、因果动作延续性的缺失，导致当今创作人使用视频生成工具时“就像在拉老虎机（slot machine）”。Luma 的使命正是赋予模型这种深层的理解力与逻辑因果律，通过整合文本中蕴含的人类大脑高级逻辑与视频中丰富的低级信号，打破这种机械的翻译模式。 ```mermaid flowchart TD accTitle: 从传统翻译模型到真正的多模态智能 subgraph Traditional["传统视频/图像模型 (翻译模型)"] A["文本提示词"] --> B["逐像素翻译/渲染"] B --> C["缺乏逻辑与连贯性的输出<br/>(像拉老虎机)"] End subgraph Intelligence["Luma 智能感知模型 (高阶智能)"] D["多模态输入<br/>(文本逻辑 + 视频信号)"] --> E["上下文记忆与因果推演<br/>(角色状态/色彩连贯/情节逻辑)"] E --> F["高连贯性内容输出<br/>(符合创意逻辑与叙事)"] End ``` - **创始人背景与创业原点**：Jain 曾在 Apple 的 Vision Pro 团队工作了四年半，参与开发了 3D 虚拟人（Human Persona，即在跨洲视频通话中能进行眼球对视的 3D Avatar）以及 3D 摄影等世界仿真技术。2020 至 2021 年间，随着大语言模型 Scaling Law 的成功，他灵光一闪，决定将这种扩展范式引入视频与图像等领域，从而创立了 Luma。 - **关于智能的本质**：Jain 强调文本之所以重要，是因为它直接源自人类大脑的高级智力输出，具备天然的逻辑与因果性；而视频则是摄像机记录的底层物理信号，Luma 的目标就是将两者无缝融合。 --- ### [15:03](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=903s) 数据来源、深度伪造（Deepfakes）与版权范式的重塑 主持人 Aishwarya Srinivasan 提出了当前 AI 视频生成领域最受关注的风险与挑战：海量数据的收集渠道、版权归属、Deepfakes（深度伪造）以及虚假信息的传播问题。针对 Deepfakes，Srinivasan 分享了一个生活中的幽默插曲：她的妈妈曾在网上看到一段两只熊靠近自家门前监控摄像头的视频，特意打来电话紧张地询问“湾区现在都有熊跑进院子里了吗？”，而她只能苦笑解释这只是网上疯传的 AI 生成视频。 针对数据供给问题，Jain 透露 Luma 在获取海量高质量视频方面完全没有瓶颈。与大语言模型（LLM）目前面临的“文本数据耗尽、模型提升放缓”不同，现实世界中的视频数据量是文本的数十亿倍。虽然人类自 2005 年左右才开始大规模录制高清视频，但当今世界产生的视频数据极其庞大。Luma 除了自身的数据积累外，还与全球顶尖的好莱坞制片厂以及大型广告公司建立了深度合作，直接获取了它们数十年来积累的庞大视频档案库。因此，相比于其他大模型公司被迫转向合成数据（Synthetic Data）或推理强化，视频模型领域拥有几乎无尽的物理世界真实数据可供训练。 在面对版权争议与 Deepfakes 治理的严肃话题时，Jain 抛出了一个非常犀利且直接的个人观：“**我们必须做出选择：要么死守 20 世纪 20 年代的传统版权法，要么拥抱人工智能。**” 他指出，包含美国、日本在内的多个国家司法体系正在逐步达成共识——即便目前的 AI 尚未达到全知全能的通用人工智能（AGI），它给整个社会带来的巨大效益也远远超出了限制它的理由，因此社会不应过分扼杀技术创新，未来的法律与商业机制自然会找到对原作者进行合理补偿的新方案。 - **视频与文本的数据量对比**：视频数据蕴含的物理世界信息量比纯文本高出数亿倍，目前完全不存在数据枯竭问题。 - **合作生态**：Luma 已经与好莱坞顶级影视制作公司及国际广告巨头建立合作，合法调取其数十年未公开的影视素材库进行深度训练。 - **治理与监管立场**：Jain 认为法律框架最终会向 AI 的巨大社会价值让步，通过创新的收益分配与补偿机制解决版权问题，而非通过死板过时的法规限制技术演进。 ### [19:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1148s) 扩展智能与深度伪造：社会心理的重塑 在探讨 Generative AI 对创作者版权的影响时，Amit Jain 明确表达了自己的立场。他认为，虽然 Luma AI 会持续配合法律法规的发展，积极推进版权与许可问题的解决，但**在“扩展智能（Scaling Intelligence）”与“保护版权（Copyright）”之间，扩展智能才是对人类社会更有价值的正确选择**。 随后，话题转到了更为敏感的深度伪造（Deepfakes）问题。Amit Jain 提出了一个深刻且颠覆传统认知的观点：解决 Deepfakes 的根本途径，不是技术封堵，而是**社会心理与认知标尺（Calibration）的重塑**。 ```mermaid flowchart TD accTitle: 媒介信任度的演进历史 subgraph STAGE1["印刷时代 (高信任)"] A["印刷机出现"] --> B["制作与传播极其昂贵"] B --> C["公众心理：'报纸上写的肯定是真的'"] end subgraph STAGE2["互联网时代 (文本信任崩塌)"] D["互联网普及"] --> E["文本生产与传播成本趋近于零"] E --> F["公众心理：'4chan/Reddit 上写的未必是真的'"] end subgraph STAGE3["视频生成时代 (正在发生的转变)"] G["AI 视频生成技术"] --> H["镜头成本从 $100,000 降至 $10"] H --> I["公众心理：'眼见不再为实，先问是不是假的'"] end STAGE1 --> STAGE2 STAGE2 --> STAGE3 ``` Amit Jain 回顾了信息传播的历史： 1. **印刷机时代**：在印刷和出版极其昂贵且困难的时期，人们看到报纸上的字就会坚信不疑。 2. **互联网时代**：随着网络普及，文本创作与分发变得极度廉价，人们的信任阈值随之改变。现在没有人会仅仅因为某段文字出现在 Reddit、4chan 甚至《纽约时报》上就全盘接收，而是会习惯性追问“这是真的吗？”。 3. **视频与图像时代**：目前图片领域已经完成了这一心理转变（人们看到离奇图片的第一反应是“这是不是 Photoshop 过的？”）。而视频正处于这一转变的关键临界点。 在过去，制作能够逼真伪造某人或某地的高质量视频，需要像迪士尼（Disney）或卢卡斯影业（Lucasfilm）那样拥有庞大预算和顶尖 VFX 团队。但随着 AI 技术的发展，单个镜头的制作成本正在从 **10 万美元骤降至 10 美元**。当视频可以被大规模生产、编辑和修改时，大众将迅速调整心理预期——**“我亲眼在视频里看到的，不代表它就是真的”**。 - **图灵测试与检测难题**：对于法律和特定场景，构建检测 AI 痕迹的框架依然有必要。但 Amit 强调，试图单靠算法检测 AI 内容非常困难。正如 LLM 文本难以检测一样，当 AI 的目标是“通过图灵测试”时，本身就意味着人类和检测器无法区分其真伪。目前生成视频中的伪影（Artifacts）只是因为模型不够完善，随着模型迭代，这些视觉瑕疵将彻底消失。因此最终的解法只有一条：“**不要盲目相信你所看到的一切（Don't trust what you see）**”。 --- ### [23:16](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1396s) 言论自由与治理边界：为什么 Deepfakes 无法从源头封杀？ 主持人 Aishwarya Srinivasan 对 Amit 的观点作出了积极回应。她表示自己有两个极具启发的新认知：一是**视频数据的体量远超文本数据**（打破了以往认为历史文献和书籍积累导致文本数据更多的刻板印象）；二是**对 Deepfakes 的治理思路不应局限于“源头管制”，而应关注“接收端的认知转变”**。 对于“通过立法或行政手段严格管控 AI 视频生成”的呼声，Amit Jain 进行了直接驳斥，并指出了这种管控在现实中的不可行性与潜在风险： * **侵犯言论自由（Suppression of Free Speech）**：强行管控个人使用 AI 生成视频，本质上与禁止人们在 Twitter 上打字或在 Facebook 评论区发言无异。过度干预个人创作工具的使用，极易演变为对言论自由的压制。 * **物理隐蔽性与技术普及**：在现实生活中，人们可能佩戴着带有摄像功能的智能眼镜（Smart Glasses）记录现场，这种记录与生成行为根本无法被彻底禁止。 * **全球地缘政治的不对称性**：即使美国制定极其严苛的独裁式法规（Dictatorial Rule），禁止任何人修改或生成视频，世界上其他国家和地区也绝不可能达成这种一致。在全球化互联网面前，单边管制完全失效。 Amit Jain 总结道，当一项技术的生产与消费成本降低到如此地步时，任何尝试从源头强行堵截的手段都是徒劳的，社会群体自发的心理适应才是最真实、最有效的防御机制。 --- ### [25:39](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1539s) 统一多模态数学架构：重构 AI 的视觉推理能力 谈到技术底层，Aishwarya Srinivasan 提问：既然目前多数视频生成模型本质上还只是“翻译”机制，Luma AI 是否正在探索全新的底层架构？ Amit Jain 证实了这一猜想，并揭示了 Luma AI 团队数年来一直在攻克的核心方向：**发明能够将文本、音频、视频和图像完美融合的全新数学模型（Invent the math）**。 | 现状（分立模型） | 未来（统一多模态模型） | |:--- |:--- | | **技术特征** | 文本、图像、视频由不同模型单独处理 | 单一模型直接建模文本、音频、视觉等多维信号 | | **推理能力** | 仅局限于文本推理（Text Reasoning） | 具备原生视频与图像推理能力（Visual Reasoning） | | **模拟对象** | 机械式的跨模态翻译转换 | 模拟人类大脑的多感官协同与综合推理 | Amit 指出，目前文本、图像和视频模型相互独立，并非因为从业者愚钝，而是因为这些信号在物理和数学特性上极其不同，建模难度极大。但人类大脑在思考时，绝不会将自己局限于“文本推理”。人类大脑时刻在进行着复杂的视觉、空间和听觉推理，只是我们自身往往感知不到这一过程的发生。 Luma AI 已经在此领域秘密研发了数年，目前已迎来突破性进展。他们对新架构展现出的成果与推理能力充满信心。Amit 强调，**要想匹配甚至超越人类大脑的能力，就必须构建出能够像人脑一样跨感官进行统一推理的综合模型**。 --- ### [27:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1628s) 视频生成的经济学与商业应用展望 针对未来的商业落地，Aishwarya 提问：目前 LLM 工具要么免费要么极度便宜，但视频生成模型因算力昂贵，动辄需要每月 200 美元甚至更高的订阅费用。未来 5 年内，视频生成技术的成本曲线会如何演变？初创企业和商业领袖该如何布局？ Amit Jain 给出了一个非常出人意料的对比视角：**“一切都是相对的。”** - **极度廉价的生成成本**：在传统的影视影视工业中，$200 甚至买不到一个高端的相机三脚架（Tripod）。 - **传统拍摄的真实天价**：如果一个剧组要去爱尔兰拍摄实景，他们需要支付昂贵的机票、申请政府许许可、花大价钱封锁街道、雇佣大量演员，还要面临不可控的天气风险。相比之下，每月 $200 的 AI 视频生成工具对于企业和影视制作方而言，**成本实际上低到了可以忽略不计的程度**。 随着底层算力效率的提升和模型的演进，视频生成的绝对成本还将进一步下降，这将彻底颠覆传统的商业视频创作模式，让小型团队和初创公司具备过去只有大制片厂才拥有的叙事与视觉呈现能力。 ### [28:33](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1713s) 视频生成的 Token 密度与算力成本经济学 在探讨生成式视频的成本时，Amit Jain 首先对比了传统影视制作与 AI 计算的差距。在传统制作中，一旦拍摄延期，整个剧组都必须滞留在酒店，那是一种“昂贵到疯狂”的物理开销；相比之下，视频生成的算力成本实际上要便宜数个数量级。然而，若将视频生成与大语言模型（LLM）相比，视频的单次生成确实更贵。Amit 解释了背后的核心技术逻辑：一段仅 5 秒钟的视频，底层包含超过 30 万到 40 万个 Token（按大致类比，相当于 30 万到 40 万个单词的信息量）。 相比之下，用户在与 ChatGPT 等文本 LLM 对话时，模型单次输出通常只有寥寥数语，因为“根本没有人愿意去阅读几十万字的长篇大论”。视频生成消耗的算力实际上是在以数十万单词的体量在进行密集计算，因此成本自然与此相当。为了让这项技术真正普及，Luma AI 正在投入巨大精力优化模型效率，使其运行更快、成本大幅降低。 - **物理制作 vs. 算力生成**：相较于传统摄制组每日食宿、场地等天价开销，AI 算力成本低了数个数量级。 - **Token 规模差异**：一段 5 秒的短视频即包含 300,000 至 400,000 个 Token，相当于瞬间生成一部数十万字的长篇巨著。 - **用户消费心理**：文本交互受限于用户的阅读耐心，而视频则以极高密度的信息流直接呈现给视觉。 ```mermaid flowchart TD accTitle: 文本 LLM 与视频生成模型的单次输出信息规模对比 subgraph LLM["文本大语言模型 (LLM)"] L1["用户输入 Prompt"] --> L2["生成数十至数百个词"] L2 --> L3["受限于人类阅读意愿与速度"] end subgraph VideoGen["视频生成模型 (Video Gen)"] V1["用户输入 Prompt / 镜头控制"] --> V2["生成 5 秒视频"] V2 --> V3["消耗 300,000 - 400,000+ Tokens"] V3 --> V4["极高信息密度流媒体呈现"] end ``` --- ### [29:41](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1781s) 颠覆推荐流媒体：每天 60 亿小时实时生成视频的愿景 Amit 提出了一个极其宏大的愿景：将全球的视频消费模式从“检索推荐”彻底转变为“实时生成”。当前，全球用户平均每天仅在手机上刷视频的时间就长达 3.5 小时（这还不包含电视等大屏时间），年轻群体的时长更为惊人。无论是新闻、娱乐、学习还是教育，视频已经成为人类获取信息的主要载体。但现有的平台（如 TikTok）本质上仍是一个双边撮合市场，其核心护城河是基于预先录制好的视频库进行搜索和推荐匹配。 Luma AI 的目标则是打破这种预录播模式，实现全实时的视频生成。Amit 设想未来每人每天至少消费 1 小时的实时生成视频。在排除婴儿以及部分无法使用其服务的地区后，全球每天将需要生成约 60 亿小时的视频。若按照当前 Dream Machine（Ray 2 模型）的推理成本计算，每天的算力开销将高达 2,800 亿美元，这在商业上显然是不可持续的。但这正是 Luma 必须攻克的技术难题，因为“这并不是我们在凭空创造需求，而是人类已经在这样消费内容，我们要做的是提供终极的解决方案”。 - **用户时长基线**：全球移动端用户日均视频观看时长达 3.5 小时，视频已全方位接管信息消费场景。 - **传统平台局限**：TikTok 的核心本质是“从已有的预录制视频库中进行搜索与推荐算法分发”。 - **终极生成目标**：全球每天实时生成 60 亿小时视频（6 Billion Hours/Day）。 - **降本挑战**：按当前 Ray 2 计算成本，该规模对应每天约 2,800 亿美元的算力支出，模型轻量化与推理加速势在必行。 --- ### [31:06](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1866s) 给企业级客户的建议：视频是叙事与营销的终极载体 在被问及企业客户及早期探索者该如何建立认知时，Amit 直言不讳地指出：“任何企业的第一要务都是讲好自己的故事，顺便提一句，这其实就叫营销。”他幽默地打了个比方：哪怕是旧石器时代第一个发明轮子的原始人，在造出轮子的第一秒，也会立刻跑去岩洞上画壁画宣告“我发明了轮子，快来找我买”。企业必须向客户证明自身的存在，而过去受限于预算，绝大多数公司根本请不起顶级广告代理商或承担高昂的影视级制作成本。 生成式 AI 正在彻底颠覆这一权力结构与成本门槛。Amit 强调，无论是销售实体产品、提供服务，还是传递某种品牌理念，“观众的眼睛在哪里、注意力在哪里，视频就在哪里”。生成式视频技术让任何规模的企业都能以可负担的成本，创造出具备顶级视觉冲击力的内容，企业管理者当下最应思考的，正是如何借助 Luma 等 AI 工具将自己的品牌故事推向全世界。 > “在当今世界，每个销售产品、制造物品或有信息需要传递给世界的公司，都必须意识到：眼球停留在视频上，注意力停留在视频上，而生成式视频则是你能真正负担得起的实现方式。” ## Notable quotes > [00:04](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=4s) 人类大脑并非仅仅通过语言进行推理，而是在所有感官模态之间协同运作。 — Amit Jain > [05:38](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=338s) 将整个世界装进你的笔记本电脑，以便你可以随心所欲地操纵它。 — Amit Jain > [09:47](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=587s) 当前的视频生成模型——包括我们自己的——都极其蠢笨。 — Amit Jain > [15:03](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=903s) 我们必须做出选择：要么死守 20 世纪 20 年代的传统版权法，要么拥抱人工智能。 — Amit Jain > [19:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1148s) 在‘扩展智能’与‘保护版权’之间，扩展智能才是对人类社会更有价值的正确选择。 — Amit Jain > [19:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1148s) 我亲眼在视频里看到的，不代表它就是真的。 — Amit Jain > [19:08](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1148s) 不要盲目相信你所看到的一切。 — Amit Jain > [25:39](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1539s) 要想匹配甚至超越人类大脑的能力，就必须构建出能够像人脑一样跨感官进行统一推理的综合模型。 — Amit Jain > [31:06](https://www.youtube.com/watch?v=YXbc-3bkqGY&t=1866s) 在当今世界，每个销售产品、制造物品或有信息需要传递给世界的公司，都必须意识到：眼球停留在视频上，注意力停留在视频上，而生成式视频则是你能真正负担得起的实现方式。 — Amit Jain ## Mentioned Amit Jain, Aishwarya Srinivasan, Luma AI, Fireworks AI, Ray 2, Dream Machine, Apple Vision Pro, ChatGPT, 李飞飞 (Fei-Fei Li), TikTok, Disney, Lucasfilm, Google Translate ## Video description Amit Jain, Founder & CEO, Luma AI Aishwarya Srinivasan, Head of Developer Relations - Growth, Fireworks AI Abstracts What if generative AI could do more than create? What if it could truly understand the world it depicts? As models move beyond text and images, the next frontier is multimodal intelligence: systems that combine vision, audio, and language to reason across the senses. This talk will explore how emerging AI models are beginning to grasp not just how the world looks, but how it behaves, and why that shift unlocks new possibilities in creativity and storytelling. [Source](https://www.youtube.com/watch?v=YXbc-3bkqGY)

[0:04] Amit Jain: 人类大脑并非仅仅通过语言进行推理，而是在所有感官模态之间协同运作。 [5:38] Amit Jain: 将整个世界装进你的笔记本电脑，以便你可以随心所欲地操纵它。 [9:47] Amit Jain: 当前的视频生成模型——包括我们自己的——都极其蠢笨。 [15:03] Amit Jain: 我们必须做出选择：要么死守 20 世纪 20 年代的传统版权法，要么拥抱人工智能。 [19:08] Amit Jain: 在‘扩展智能’与‘保护版权’之间，扩展智能才是对人类社会更有价值的正确选择。 [19:08] Amit Jain: 我亲眼在视频里看到的，不代表它就是真的。 [19:08] Amit Jain: 不要盲目相信你所看到的一切。 [25:39] Amit Jain: 要想匹配甚至超越人类大脑的能力，就必须构建出能够像人脑一样跨感官进行统一推理的综合模型。 [31:06] Amit Jain: 在当今世界，每个销售产品、制造物品或有信息需要传递给世界的公司，都必须意识到：眼球停留在视频上，注意力停留在视频上，而生成式视频则是你能真正负担得起的实现方式。

[2:26] Amit Jain: Robot Brain Hypothesis [8:59] Amit Jain: dumb translation models [8:59] Amit Jain: like Dory [8:59] Amit Jain: slot machine [16:07] Amit Jain: ungodly amount of video data [19:34] Amit Jain: Don't trust what you see. [31:34] Amit Jain: he went and painted on the caves: 'I got the wheel. Come buy from me.' [31:34] Amit Jain: video is where the eyes are, video is where the attention is.

[0:04] So Amit, thank you so much for joining [0:06] us for this chat and I'm actually very [0:08] glad that I got this opportunity to [0:11] speak with you because I definitely have [0:12] a lot of questions. U I come from the [0:15] machine learning space. I was talking to [0:16] you and there's obviously a lot of [0:18] unanswered questions that I have in my [0:20] mind especially from a video space and [0:23] I'm glad I'm sitting with you today to [0:25] chat a lot about that. So let's let's [0:27] start talking about Luma AI like [0:30] your background is in physics. You were [0:32] mentioning to me that you come from a [0:34] physics uh background and you started [0:36] this company to build videogen using [0:39] text and we just saw that it's very easy [0:42] to prompt that model and like get get [0:44] your videos generated, right? What was [0:46] your motivation behind it and what does [0:47] Luma do? [0:48] >> Yeah. Um thank you for having me. Um so [0:51] yeah basically the the premise of the [0:54] company or what we're trying to do is to [0:56] build um intelligence that is um that [0:59] learns from all of the digital footprint [1:02] that humanity has generated audio video [1:04] images and text. Currently [1:07] most of the world is primarily focused [1:09] on building LLMs or language models [1:12] right and I think uh we have seen great [1:14] gains from um reasoning uh uh encoding [1:17] performance and and mostly text based um [1:20] textbased uh approaches. [1:23] What comes next after that is what we [1:26] are working on which is can we learn [1:28] from [1:31] video? Can we learn physical [1:32] understanding? Can we learn like you [1:34] know how things move? Can we learn how [1:35] things behave? Can we learn causality? [1:37] Can we learn uh like you know from [1:39] images how things look? How composition [1:41] should be? Uh like you know uh what the [1:43] physical rules of the world are from [1:44] images and videos. Can we learn from [1:46] audio how things sound? Right? Because [1:50] the human brain doesn't just reason in [1:53] language. The human brain actually [1:54] operates across all of those modalities. [1:57] So the work we're doing at Luma is to [1:59] build is to basically build models that [2:02] are able to train from audio, video, [2:05] language and images all together. And [2:09] video models are like you know the first [2:11] product that we have released. Uh but [2:12] the current model that is in the market [2:13] it's called ray 2. It's able to generate [2:16] video images and text today and and in [2:19] the future uh like you know the next [2:20] models that we release uh we're trying [2:22] to unify this intelligence into like you [2:24] know one single thing. That's actually [2:26] very interesting and [2:29] I have a followup to that. [2:30] >> Yeah. [2:31] >> What you're speaking about is sounding [2:33] like it's something to do with [2:34] simulation, right? And a lot of times [2:37] any of the model providers launch their [2:39] video model, it sounds very cool. Like [2:41] the demos are very cool. They're very [2:43] hyperrealistic videos that come out. But [2:45] I'm always questioning myself, it looks [2:47] cool, but what is the point? Where do we [2:49] use this? Is it mostly for like say [2:51] creators or YouTubers to add to their [2:54] videos? like what's the point of this? [2:56] What's the real business use case or [2:58] like a real world application for this? [2:59] So, one of the things you mentioned [3:01] sounds like simulation, right? Could we [3:03] simulate certain things and see how it [3:05] would work in a physical world? What [3:08] other use cases are you seeing [3:09] predominantly with the customers that [3:10] you're working with? [3:11] >> Yeah, so video is actually unbelievably [3:14] massive use case. Um, Hollywood alone [3:18] spent about 300 to$350 billion last year [3:21] on production, right? Um if you start to [3:25] add up all the money that is actually [3:27] spent making videos and the reasoning is [3:30] actually really simple uh and it [3:31] shouldn't be shouldn't come as a [3:32] surprise. People cannot get enough [3:34] videos to watch given a choice between [3:37] reading and watching something. Um you [3:40] know most people would prefer to watch [3:42] and and the reasoning is also not that [3:44] like oh you know the civilization is [3:45] going to Uh the reasoning is also [3:47] because vision and and our ears are much [3:50] higher bandwidth input signals than text [3:53] is right reading text takes immense [3:55] amount of energy compared to just [3:56] watching and and uh um you know hearing [3:59] about it. So given a choice, most people [4:02] are choosing to use video for [4:06] everything. Entertainment, education, [4:09] learning about like you know how to do [4:10] silly things in your life, how to do [4:12] like you know like how popular are Excel [4:16] uh Tik Toks now, right? Like you know [4:18] where they teach you how to use Excel. [4:19] Um [4:21] video is now the medium of information [4:24] on the internet. It used to be text. It [4:27] used to be text all of the knots and uh [4:29] tents but now very decidedly you know [4:33] there's like what three large networks [4:35] with more than three billion users per [4:36] month uh that that you know are [4:39] supported by just video. Um so video is [4:42] a very very large business. The things [4:45] you watch interspersed by all the ads, [4:48] they're all videos, right? And currently [4:52] the act of production of video is [4:54] entirely physical, right? You it's it's [4:58] a logistical process. You you do it all [4:59] in atoms. You you need the people, you [5:01] need the lighting, you need the [5:02] environment, you need the sets, you need [5:03] like, you know, the actors and the the [5:05] people to tell the actors what to do. [5:07] All of that. It's a very very physical [5:10] process for something that is this [5:12] popular, right? when text was like that [5:15] you know we had the printing press and [5:17] and then then it became entirely digital [5:19] uh using word processors and that led to [5:21] you know frankly the internet and then [5:24] now and and now we're at a point where [5:26] like you know most of the text that is [5:27] produced in the world most of the words [5:28] that are produced in the world are made [5:30] by LLMs for one person right like you [5:33] know that that are scrolled away as soon [5:34] as you're done you know reading it or [5:36] you don't even read it [5:38] video is not going to be any different [5:40] than that right so you ask for the use [5:44] case. The most primary uh like you know [5:45] on then use case is really [5:48] bringing the world into your laptop so [5:51] you can manipulate it however the hell [5:52] way you want. Uh whether that is for [5:54] like you know making a movie, whether [5:56] that is for making a short video and [5:58] telling a story, whether that is for for [6:00] uh uh communicating something like you [6:02] know that you want to teach the world [6:04] whatever it is now you don't need a [6:07] crew. Now you don't need a large [6:09] production budget. Now you don't need [6:10] someone's permission because you know [6:12] whenever there's a restriction in supply [6:14] and and it's controlled by a few people [6:15] you need those people's permission to do [6:17] something. Um video is too precious and [6:20] too important and too uh uh pervasive to [6:24] be that way. So the most important use [6:26] case for simulation is hey can it just [6:29] be on my laptop and can I just like you [6:31] know uh manipulate it. So that's the [6:33] that's the first one. Uh being able to [6:35] generate entertaining convincing video. [6:39] It's very hard. It's harder than being [6:40] able to generate entertaining, [6:42] convincing text. Um because text has a [6:44] singular dimension, right? Like you [6:45] know, of course, it is still hard. Like [6:47] you know, passing the text during text [6:49] test took a long time. Video is much [6:51] harder. People's expressions, causality, [6:54] emotions, all these kind of things. It [6:56] takes a lot of energy and work to you [6:59] know get there. And our models, as you [7:00] were saying, are not fully there yet. So [7:04] physics and these kind of things are [7:05] getting really better. and we are now [7:07] starting to hit the scaling uh uh ramp. [7:09] So uh uh they're starting to you know [7:11] look and feel very very good but uh they [7:14] will take another year or so before like [7:15] you know they're fully physically [7:17] accurate and uh are able to follow like [7:19] a chain of thought or line of thought [7:21] very very properly. So that's the [7:23] generation side of use case and and like [7:25] you know if you think about the [7:26] industries that it affects it's it's [7:27] entertainment media advertising huge uh [7:31] like overlaps in like you know the [7:33] people that work in those areas how they [7:35] do things and how they're going to use [7:36] these models. Then comes the second part [7:39] of it like you know for video which is [7:40] um [7:42] just being able to use it as a question [7:45] and answer tool right currently most [7:48] people are asking video model uh chat [7:50] chat GBT and and and LLMs answer to [7:52] their questions a lot of those questions [7:54] are better answered visually being able [7:56] to generate it on demand I think that is [7:57] huge and that is going to be very big [8:00] and finally as these models get very [8:03] good they start to become very good for [8:06] planning and thinking [8:08] in the physical world. So that that [8:10] that's what we call the robot robot [8:11] brain hypothesis uh at Luma which is [8:13] that like [8:15] humans plan very visually in their heads [8:18] like if I told you right now to you know [8:19] imagine walking from your uh front door [8:22] to your bedroom right in your house you [8:24] can do that pretty much effortlessly [8:26] until unless you have affantasia right [8:28] like you know [8:30] robots today don't do that they there is [8:33] no way for them to do that they're [8:35] trained on action data uh that that is [8:38] collected through teley operations and [8:40] they have no model of the world in their [8:42] heads. [8:44] It's a very brute force way of making [8:46] something that we want to be super [8:47] general. So [8:50] we want to make these things very good. [8:52] We want to make them very intelligent [8:54] because we think this is going to be [8:56] ultimately the robot brain. So those are [8:58] the use cases. [8:59] >> I love it. It very uh accurately aligns [9:02] with the things that we heard from Faili [9:04] this morning as well during the keynote [9:06] and you very well articulated the [9:08] problem of why video models are so hard. [9:12] uh people who might be new to looking at [9:14] video generated models. It was at a [9:17] point considered as a touring test and [9:18] still is considered as a touring test [9:20] for AI because once you start generating [9:23] models uh generating like videos using [9:26] these models, you can see that the [9:28] activities you know like for images that [9:29] was like producing fingers like it could [9:31] just never get fingers right because it [9:34] was just very now it does right it it [9:36] was a curve right. So um to a point it [9:39] was like if you're trying to produce [9:40] gymnastic videos using the video gen [9:42] models it can take your leg all the way [9:44] around. So that's one of the hardest [9:47] problem that people are still finding. [9:50] What do you think is like makes luma [9:53] unique there? Is it your engineering [9:55] experience? Is it your physics [9:57] experience? Or is it access to really [10:00] good and diverse data that helps you [10:02] produce that high quality of a model? [10:05] What do you think is your differentiator [10:06] and how how did your experience from the [10:08] past help you build Luma? [10:11] >> Right? Those are two separate questions [10:13] to be honest with you. I think the most [10:15] important difference between Luma and [10:17] pretty much every other video generation [10:19] company is that Luma is not a video [10:21] generation company. Um we don't believe [10:24] in in just video generation or image [10:26] generation or text generation. I think [10:28] that is the problem right what you want [10:31] to do. So okay let me let me actually [10:34] articulate the problem very very uh uh [10:36] tactically right like you know what what [10:38] it manifest as current video generation [10:40] models including ours are dumb as [10:43] like [10:45] it's it's very difficult to call them [10:46] intelligence unlike NLM where like you [10:48] know you start a conversation it will [10:50] remember what you're talking about right [10:51] like you know turn by turn by turn um [10:54] it's possible that like you know you you [10:56] talk a lot and then it has to like you [10:57] know purge the context just just because [10:59] context lens are limited but modular [11:02] that in every way, shape or form it is [11:04] it is very much understandable that it [11:06] has it has a context. [11:08] If you're talking to an LLM about like [11:10] you know certain things and and and um [11:13] where like let's say you know oh let's [11:15] let's talk about a movie script or let's [11:16] let's talk about uh like you know this [11:18] analysis problem it will also do some [11:20] thinking on its own part doesn't matter [11:22] if you believe it's thinking or not [11:24] doesn't really matter but like you know [11:25] the output looks like it uh and and it [11:27] will propose other things to you that [11:28] sort of stuff video models image models [11:30] don't do that current image and video [11:32] models produced by every company on the [11:34] planet are translation models they're [11:37] just taking text and translating them [11:39] into pixels. That's not what [11:41] intelligence looks like, right? um you [11:44] don't consider a a language translation [11:46] model like you know goo Google's uh uh [11:49] uh Google translate that was not [11:50] considered AGI that was not considered [11:52] even AI right nobody was losing their [11:55] mind and losing their jobs to that right [11:58] maybe some people did in the UN uh like [12:00] you know translation field but that was [12:02] not the main thing current image and [12:04] video generation models are basically [12:06] very dumb translation models they need [12:09] to become intelligent so when a [12:11] filmmaker is working with them. Every [12:14] single scene you prompt for, every [12:16] single 5second, 8second shot you prompt [12:18] for, it doesn't start like Dory from [12:20] scratch that like, you know, ah, what [12:22] are we talking about? Uh, right. [12:24] Everything from colors to characters, [12:26] like for instance, like you know, if in [12:28] your story a character died, right? And [12:31] then you're generating a next scene and [12:32] they're appearing, right? Like, you [12:34] know, it's like, okay, why are they [12:36] here? Are they supposed to be a ghost? [12:38] Right? uh you should you should be able [12:39] to say, "Oh, this is a a like, you know, [12:41] flashback scene or something like that." [12:43] But models need to be intelligent. They [12:44] need to retain the color coherency. They [12:46] need to be able to like, you know, uh [12:48] continue an action that was done from [12:49] this side. Like, you know, if if two [12:51] characters had a big fight in a previous [12:52] scene and then there's a following [12:54] scene, they shouldn't greet them each [12:55] other that like they're lifelong [12:56] friends. None of that exists today. [13:00] That's why using video generation [13:02] models, image generation models feels [13:03] like, you know, you're using a slot [13:05] machine. Speak of Vegas. Um [13:09] so Luma is different because we are [13:12] trying to build that level of [13:13] intelligence. [13:15] Um until that exists [13:20] most people especially in creative [13:22] fields don't actually have access to [13:24] artificial intelligence. All they have [13:26] are these translation models. So that is [13:29] what we need to build. That is what we [13:30] are doing. And and the work to do that [13:33] looks very different from making a video [13:34] model or image model or text model. That [13:36] looks like we need to combine all these [13:38] the logic and causality that comes from [13:40] text right like you know why is text so [13:42] important because text is already coming [13:44] from an intelligence the human brain [13:46] right video is a low-level signal it is [13:48] coming from recording from a camera so [13:50] if we want to build intelligence for [13:53] creatives if you want to build [13:54] intelligence to be able to like you know [13:56] produce video to be able to make this [13:58] very powerful medium actually automated [14:00] we need to do that work and that's what [14:01] we are doing and and how I come came to [14:04] this so I was working at Apple um on [14:06] vision pro for about uh for four and a [14:08] half years. We we built many things on [14:10] that uh on that device including the [14:12] human persona that those are 3D avatars [14:15] that you know you you can it looks like [14:17] you're sitting across from them and you [14:18] can look in look at them in the eye but [14:20] the other person could be on another [14:22] continent actually um did that work uh [14:25] did a lot of work on like you know 3D [14:27] photography and these kind of things. So [14:28] I I had been doing a lot of work on on [14:30] world simulation in in restricted or [14:33] narrow capacities. But then when [14:35] language model scaling started to show [14:37] promise in 2020 and 2021 [14:40] um the natural leap was like okay well [14:42] if if this also works for other things [14:44] like video or images then that will be [14:46] huge. So that was the genesis of Luma. [14:50] >> I love it. I see how you've gone from [14:52] like building this for one of the [14:54] biggest companies who's doing like uh in [14:56] computer vision and virtual virtual [14:58] reality to building your own company [15:00] Luma [15:03] let's talk about some of the concerns [15:05] that comes with video generation in [15:07] general right whether it be how do you [15:11] collect the data where do you source the [15:13] data from how do you attribute or like [15:16] are there any copyright issues with the [15:18] data that you're collecting or whether [15:21] it be deep fakes. Hey, like I go onto [15:25] the internet, I'm seeing somebody's [15:27] video and I don't know if that person is [15:29] real or not. And we have seen a lot of [15:32] situations where people's avatars have [15:34] been published on the internet without [15:36] their consent. Uh there are a lot of [15:38] misinformation that goes around [15:40] especially like sometimes my mom calls [15:42] me and she's like hey I saw this video [15:44] of like two bears um coming in like you [15:47] know close to your um home home video [15:50] camera uh I saw that and like is is it [15:52] real like does that happen in Bay Area [15:54] and I'm like no that's a AI video which [15:56] is just like circulating on the [15:57] internet. So what are your thoughts [15:59] about about these concerns? Whether it [16:01] be like video coming from sources, how [16:04] do you source them ethically? How do you [16:07] have a continuous source of them rather [16:09] than just the internet? And second part [16:12] is around deep fakes like how do we [16:14] control them? Is there a way to govern [16:16] it? Do we need to do that? How do we go [16:18] about it? [16:20] >> Right? So [16:22] sourcing there is ungodly amount of [16:25] video data in the world. It's just so [16:28] much easier to produce video. It's just [16:30] so much easier to actually like, you [16:32] know, gather incredible amount of video [16:35] showing everything physical, [16:37] non-physical that happens in the world, [16:39] right? Um the internet has a loss and [16:42] like you know then then we do our own [16:44] efforts quite a bit of those. Then we [16:46] have a lot of partners right like you [16:47] know who who have large troves of video [16:49] from previous products from like you [16:52] know things that they make themselves. [16:54] We work with some of the largest [16:55] Hollywood studios in the world today. We [16:57] work with some of the largest [16:58] advertising agencies in the world today [16:59] and they have decades and decades and [17:01] decades of of you know archives of [17:04] video. Um so yeah like actually you know [17:07] getting the data is really not a big [17:09] problem. Um in text it is and actually [17:12] you're seeing that like you know we are [17:14] kind of running out of text information. [17:16] Uh and so so newer language models are [17:20] sort of slowing down in their gains, [17:23] right? Um and that is not happening [17:25] because we we we we don't know how to [17:26] scale the model sizes because we don't [17:28] have computer scale model sizes, but [17:30] that's happening because there's no more [17:31] data to train them on. So there's no [17:33] nothing more new to learn. Of course [17:36] there is and then we know that from the [17:37] physical world that or real world that [17:39] like oh there are many things these [17:40] models are not good at but there's no [17:41] data to use that. So a lot of companies [17:43] are going to synthetic data, going to [17:44] reasoning, going to like you know these [17:46] alternative approaches rather than like [17:47] you know just getting the models on on [17:49] the data that you have. For us that's [17:51] great because you know we actually like [17:53] video is is billions of times more data [17:56] than text right um because video is just [17:59] so easy to produce even though humans [18:00] started producing videos at scale only [18:02] in like 2005 right before that we had [18:05] television and movies and those kind of [18:06] things but not really that many only in [18:08] around the early ons did did we start to [18:10] get really high quality video and now [18:14] there's just so much of it so that's one [18:15] part of it but you asked about like you [18:17] know uh copyright deep fakes those kind [18:19] of things [18:20] I think there's like a general stance [18:22] and this is not Luma's stance this is [18:23] mine particularly which is that like we [18:26] have a choice to make either we want to [18:28] stick to the 1920s copyright laws or we [18:33] want artificial intelligence it's [18:36] basically that way right um [18:40] that's kind of where it's going to come [18:41] out to be and as you're seeing now like [18:43] you know in the legal system in the [18:44] United States in the legal system in [18:46] Japan in legal systems in many countries [18:47] actually they're coming at this of like [18:49] okay you know artificial intelligence is [18:51] such a ridiculous benefit to the society [18:53] even if it doesn't end up being general [18:55] intelligence that like you know does [18:56] everything even in its current state [18:58] it's such a benefit to the society that [19:01] it doesn't make sense to stifle it right [19:04] so [19:06] people will find ways of compensating [19:08] the creators uh people will find ways of [19:10] of like you know correctly licensing [19:12] doing all these kind of things and and [19:14] and we'll we'll be right alongside there [19:16] um and then we currently already do that [19:18] a lot and and and we'll continue to do [19:19] that as as like you know the law [19:21] progresses on that area. But honestly [19:23] like you know between scaling [19:24] intelligence and and copyright I think [19:26] scaling intelligence is what is the [19:28] right choice in the world to do um on [19:31] deep fakes. So uh I think that is a very [19:35] interesting one that is the one I think [19:36] a lot about [19:39] and I think there's a simple solution. [19:43] Let's talk about text, right? Uh I think [19:45] I mentioned that just a second ago, [19:46] which is like [19:48] when text production was difficult and [19:51] we had the printing press, anybody who [19:53] made if you saw text printed on a [19:56] newspaper, you believed it. You're like, [19:59] okay, well, you know, if it says so in [20:00] the paper, then it is true, must be [20:02] true, right? But then when the internet [20:05] came about, the society had to move on [20:08] from that idea because producing text [20:10] and distributing text became so easy and [20:12] so cheap that like you know now if you [20:14] see something written on on 4chan or [20:17] Reddit or even on New York Times a lot [20:19] of times, right? You don't really [20:21] believe it because it's like is it [20:23] really true? I don't know. Just because [20:25] it's written and distributed doesn't [20:27] mean it's true. [20:29] Video is going through that phase now. [20:32] um [20:34] right now or and let's say you know even [20:36] today like majority of the world [20:37] producing video that looks really real [20:41] and convincing of of someone or some [20:44] place whatever have you is expensive and [20:47] if you want to really fake it you need [20:49] the resources of a Disney or a Lucasfilm [20:51] right you need that level of like like [20:53] VFX and and budgets and all these kind [20:55] of things but with the work we are doing [20:58] it is going to become so cheap it is [21:00] already is like you know uh from [21:02] $100,000 a shot you get to like you know [21:04] $10 a shot. So so when you are able to [21:07] massroduce it, mass edit it, mass modify [21:10] it very soon and this is not like I'm [21:13] saying oh society should do this. No, [21:15] this is going to happen very soon. [21:17] People will change uh like you know [21:19] their their calibration. If I saw it on [21:22] video doesn't mean it's true, right? Uh [21:24] right now we're not there but I think [21:26] it's going to get there very very [21:27] quickly. Images we're already there, [21:28] right? Uh oh, is this Photoshop? is this [21:30] doctor like this is the first question [21:31] people ask right when something looks [21:34] not so believable to them um video is [21:37] going to go the same way so what do we [21:40] do about deep fakes right like [21:41] especially like someone's identity [21:43] someone's likeness things like that just [21:46] like I could put something on the [21:47] internet right now and it says like [21:49] written by Ashwaria what can you do [21:52] right people not going to believe it's [21:53] written by because it's not coming from [21:54] your channels right it's just not people [21:57] the first things they're going to go is [21:59] ask you hey did you write this. I saw [22:01] this thing on Reddit. Did you write [22:02] this? Right? Video is going to be the [22:04] same way. They're like, "Hey, uh, I see, [22:07] you know, Trump doing this or or or [22:09] like, you know, doing this. Is this is [22:11] this actually her?" Right? We just [22:13] haven't crossed that window yet. Uh, it [22:15] will happen very soon. But, uh, um, [22:18] yeah, for legal and and other purposes, [22:20] I think there's necessity for making [22:22] frameworks which can detect these [22:23] things. [22:26] But like how well can you detect test [22:28] text written by LLMs? [22:30] Can you right? Uh um it's very [22:33] difficult. It's very very difficult [22:34] because on one side you're trying to say [22:36] hey we're going to pass the touring [22:37] test. The touring test literally means [22:39] that humans can't tell what was [22:42] generated. [22:44] So how you going to do that right? Um [22:48] and very soon like you know currently [22:50] videos generated by uh our models have [22:53] subtle things in them that you can tease [22:57] apart right they don't look like they're [22:59] coming from a camos sensor of of a [23:01] camera but very soon that will also go [23:03] away right like you know because they're [23:05] just artifacts of the models not being [23:06] good enough yet so I think the ultimate [23:10] solution is that don't trust what you [23:12] see [23:14] >> I got really intriguing things from this [23:16] chart that like two very intriguing [23:18] things. One, I did not know that we have [23:20] way more video data compared to text. I [23:22] think that was news to me because [23:24] historically we just think that all the [23:26] archives of the newspaper, books, [23:28] everything used to be in text. And by [23:31] common logic, one might think that [23:33] there's way more text than videos [23:35] because the instruments to even record [23:37] these videos are very recent and it's [23:38] expensive to record, host, etc. But [23:41] that's very interesting to me. Second [23:43] thing that I found very interesting is [23:45] that a lot of times when I chat with [23:47] people about deep deep fakes, their [23:49] natural response is we should govern it. [23:52] Right? That's the natural response. And [23:54] I find your [23:56] method very interesting is that one [23:59] thing is to govern or uh control it from [24:02] the source. The other is how how does it [24:05] go to the receivers and how should we as [24:07] people understand how we need to change [24:09] our thought process with AI tools being [24:12] used and all of those output being shown [24:14] to us every single day in every single [24:17] platform that we use. So I think that's [24:19] a very very interesting uh perspective [24:21] that you shared with me. [24:22] >> You can govern it like uh you can try [24:25] but that is very quickly starts to come [24:28] up uh uh very close to like you know [24:30] suppression of free speech. How do you [24:32] govern it? Like can you govern someone [24:34] typing something on Twitter? No, you [24:36] can't. Can you govern something like you [24:38] know sharing something on the internet [24:39] in Facebook comments? No, you can't. If [24:41] you try to do that, that is suppression [24:43] of free speech. Um similarly like you [24:45] know you can't govern someone being able [24:47] to make a video. You already cannot [24:49] govern someone being able to make a [24:50] video, right? Some of you might be able [24:52] to record this, right? Like you might [24:53] have met glasses that are like you know [24:55] capturing this and and I can't do [24:57] about it. Um you just can't do that. uh [25:02] you would want to do that like you know [25:04] we can we can have a huge conversation [25:05] about responsible AI and all these kind [25:07] of things but at the end of the day when [25:09] something is meant for mass consumption [25:11] and production and and the costs are [25:13] this low I mean what are you going to do [25:15] like you can say in the United States [25:16] we're going to like you know institute a [25:18] a dictatorial rule nobody can alter [25:20] videos all right um the other countries [25:23] in the world who will never you know uh [25:25] uh agree to that um so you you can't [25:28] govern it I think the right solutions [25:30] are are or what I just said like you [25:32] know people will just very quickly adapt [25:35] they're like oh yeah what I see is not [25:37] really true [25:39] >> so the nerd in me obviously wants to ask [25:41] you this you mentioned that currently [25:42] the videog models are mostly translation [25:46] so [25:48] are you saying that luma is working on a [25:50] completely different architecture of how [25:51] video gen should look like [25:54] >> how is that like what's your I'm not [25:56] asking for your secret sauce like what's [25:58] your thought process behind it [25:59] >> yeah so I As I was saying, um, we need [26:02] to invent the math that is necessary to [26:04] combine text, audio, video, images [26:06] together. Right now, they're different [26:08] models, not because like, you know, [26:09] people are stupid, but because [26:11] technically they're really different [26:12] signals to model. So, the architectures [26:15] of these models and and and how they're [26:17] trained is very different from each [26:20] other. I mean and if you want to combine [26:22] it if you want to like you know unify it [26:24] into one single place [26:26] then it takes it takes like you need to [26:29] build different kind of models. Uh so [26:31] that is the work that we are doing and [26:32] we have been doing for for a couple of [26:34] years now and like you know it's coming [26:35] to head now where like you know we're [26:36] very confident in in uh the results that [26:39] it is producing and and and the [26:42] capabilities that you get out of it the [26:44] kind of reasoning you can get out of it [26:46] from like you know having video [26:47] reasoning and image reasoning and these [26:48] kind of things like right now people [26:49] only think about text reasoning but why [26:51] should that be only in text? Human brain [26:53] doesn't limit itself to thinking about [26:54] only in text. Human brain does all kinds [26:56] of reasoning right like you just don't [26:58] know it's happening in your brain right [27:00] like you know you you you just don't [27:01] know about that so if you want to match [27:03] the capabilities of of human brain and [27:06] surpass that's what you need to do [27:08] >> so from 5 years from now where do you [27:10] think we'll see all of these like image [27:13] outputs being made especially because of [27:15] the cost like for example right now any [27:18] of the LLM based tools are available [27:21] either as a premium [27:23] >> or at a price which is extreme extremely [27:25] low. But when it comes to video models, [27:28] either they are available if you are [27:30] paying like $200 a month or just like [27:32] buying the tool or licensing things [27:34] around it is really really expensive and [27:36] fairly expensive because of the compute [27:38] that goes behind it. Right? So where do [27:40] you see from 5 years from now how things [27:42] would move from a video generation [27:44] perspective? Where do you where do you [27:47] see people especially like businesses [27:48] using these uh toolkits more and in the [27:52] audience if we have early professionals [27:54] or um early entrepreneurs or business [27:57] leaders who are thinking about if my [27:59] business requires a part of video [28:00] generation what would that early thought [28:03] process be for them like who should be [28:05] looking at this right now [28:07] >> everything is relative if you look at [28:09] the actual cost of video production [28:11] today [28:12] $200 is like cost of a tripod stand for [28:16] a camera, right? Actually, a good tripod [28:19] stands are much more expensive. Um, if [28:21] you think about like the physical [28:23] production process, like, you know, you [28:24] want to shoot something in Ireland, [28:25] well, let's go fly to Ireland. Let's get [28:27] like, you know, permission from the [28:28] government. Let's pay them to like, you [28:29] know, close off the area where I'm doing [28:31] this shoot. Let's get all the actors [28:32] there. Let's make sure the weather is [28:33] right. If it is not tomorrow, everyone [28:35] is staying in the hotels. It's an [28:37] insanely expensive process. So if you [28:40] compare that to [28:42] the cost of compute on video generation, [28:44] it is absurd, right? Like you know it's [28:46] orders of magnitude cheaper. But I get [28:48] the point you're making like compared to [28:50] LLM's like you know video generation is [28:52] more expensive. But you know why it is [28:53] more expensive? Because each video a [28:55] 5-second video is is upwards of 300,000 [28:57] tokens. If you if you take the rough [28:59] translation of like you know one token [29:00] is one word. It's not but like it's it's [29:02] about the vicinity of that. That's about [29:04] you're generating in for for 5-second [29:06] video you're generating about 300,000 [29:08] 400,000 words right so when you're using [29:11] tajiv when you're using an LLM you [29:13] you're really the responses are what a [29:15] few few words right uh because you're [29:18] not going to read if it's anything more [29:19] than that coming back to nobody likes to [29:21] read anything um you're not going to [29:23] read the thing right like you know if [29:24] it's 200,000 words so you're actually [29:28] consuming compute at the scale of 200 [29:30] 300,000 words so yeah the cost is [29:32] commensurate to that And of course we're [29:35] working a lot on like you know uh making [29:37] them more efficient, making them faster, [29:39] um making them a lot cheaper because I [29:41] think my goal is actually um very [29:44] simple. Currently people are spending 3 [29:47] and 1/2 hours every day watching video [29:49] right on their phone. This is the [29:52] average on phone. This does not include [29:53] time spent watching it on television, [29:55] right? Which I don't know what what what [29:57] that number is. Um this is the average. [29:59] So that means like you know younger [30:01] people are actually watching a lot more. [30:02] All the people are watching a bit less [30:04] um they're using video for everything. [30:07] Education, entertainment, learning, [30:09] whatever have you, right? News, but most [30:12] of it currently it's created by one [30:15] person and then searched and served. [30:16] What is Tik Tok? Tik Tok is a two-sided [30:18] marketplace where the algorithm, the [30:20] main secret source is a search [30:22] algorithm, right? It it it serves the [30:24] recommendations to you. It searches from [30:25] a pre-existing repository of [30:27] pre-recorded videos. We're going to make [30:29] it live generation to be live and we're [30:32] going to be able to generate at least an [30:34] hour of video per person per day. That's [30:37] 6 billion hours of video per day, right? [30:39] I'm counting infants and some people in [30:42] China as not being able to use our [30:43] services. So, uh uh like you know 6 [30:45] billion hours of video per day is what [30:46] we want to generate at current costs of [30:49] like you know Ray 2 uh that comes out to [30:51] be about $280 billion a day the cost. So [30:54] of course that's not going to work out. [30:55] So we have a lot of work to do in front [30:56] of us but it's going to happen and and [30:59] uh not because like you know we will [31:00] make it happen but because this is what [31:02] people are doing already and we are [31:04] going to solve that problem. [31:06] So coming back to the question of like [31:08] if there are people here from [31:09] enterprises uh probably the last [31:11] question for us would be like if there [31:13] are people here what should be that [31:14] early thought process be for people like [31:16] especially from an enterprise [31:17] perspective uh [31:18] >> I think if you're a company any kind of [31:21] company you need to tell your story [31:24] right that's called marketing by the way [31:25] like just so you know um you have to [31:28] tell your customers you exist I'm pretty [31:30] certain that like you know the the first [31:32] early men who invented the wheel right [31:34] like you know right after he made the [31:35] wheel. He went and painted on the caves. [31:37] I got the wheel. Come buy for me. Right. [31:40] You have to tell your story. You have to [31:42] do marketing currently. Like, you know, [31:44] you are limited by what you can afford. [31:46] But can you get the best agencies? Can [31:47] you get like, you know, these production [31:48] budgets for videos, things like that? [31:50] Not anymore. I think if you're a [31:51] company, you need to think about uh uh [31:53] uh like, you know, how can you actually [31:55] get out in the world and and how can [31:57] companies like Luma and and our work [31:59] actually uh help you do that? Um, every [32:02] company that is selling anything, making [32:05] anything, or has a message to to get out [32:07] to the world, video is where the eyes [32:09] are, video is where the attention is, [32:11] and generative video is how you can [32:14] afford to do it.