## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=PI49GKzzRcs&t=4s) Introduction and the Need for a New AI Framework
Ryan Wang, Founding Partner at Outpost Capital based in the Bay Area, introduces Spatial AI as an underrated "hidden gem" in the artificial intelligence landscape that could be the next major frontier. He highlights his firm's focus on early-stage frontier technologies—including spatial computing, robotics, AI, and blockchain—and critiques the typical pitches made by AI investors. Wang notes that nearly all venture capitalists claim to invest in AI today, making the statement almost meaningless without specifying *where* in AI they focus.
Wang displays a standard industry chart showing the AI stack—divided into an infrastructure layer, middle-layer developer tools/DevOps, and application layers split between consumer apps and enterprise verticals. He expresses frustration with this conventional framework, arguing that it is far too crowded across all layers and reduces the ecosystem to a single-threaded narrative. Pointing out that the conference grew from 2,000 attendees to 8,000 this year, he emphasizes that entrepreneurs and investors should seek "blue ocean" opportunities rather than entering red ocean markets where 20 to 30 competitors build identical solutions.
- **Outpost Capital Focus Areas:** Early-stage investments in spatial computing, robotics, AI, and blockchain.
- **Industry Growth Example:** The Ai4 conference expanded from 2,000 to 8,000 attendees.
- **Industry Trend Critique:** Last year was dominated by Retrieval-Augmented Generation (RAG); this year is dominated by AI agents.
---
[05:13](https://www.youtube.com/watch?v=PI49GKzzRcs&t=313s) Defining Spatial AI and the First Industry Map
Wang clarifies what Spatial AI actually means, distinguishing it from related terms. While AR/VR and devices from Apple or Meta's Reality Labs focus on spatial computing as a media layer for immersive content, Spatial AI adds an essential layer of artificial intelligence. Furthermore, while Dr. Fei-Fei Li and World Labs coined "spatial intelligence" from an academic research standpoint—focusing on 3D-capable foundation models—Wang defines Spatial AI in simpler terms: **the intersection between Artificial Intelligence and 3D** [06:20, 07:21].
```
+-------------------------------------------------------------+
| SPATIAL AI |
| (Intersection of AI and 3D) |
+------------------------------+------------------------------+
|
+-----------------------+-----------------------+
| |
+------v----------------------+ +----------v-------+
| FOUNDATION 3D MODELS | | PHYSICAL AI & |
| - World Labs, Gen-3D | | ROBOTICS |
| - 3D-native multimodal labs | | - Real & virtual |
+--------------+--------------+ +----------+-------+
| |
+-------+-----------------------+---------------+
| |
+------v----------------------+ +------v-----------------------+
| CONSUMER MEDIA & GAMES | | ECOSYSTEM ENABLERS |
| - 3D animation / simulation | | - Nvidia Omniverse / GTC |
| - Game engines (Unity, UE) | | - Synthetic data generation |
+-----------------------------+ +------------------------------+
```
To help map this emerging market, Wang and his partner Michael Goet, a professor from the School of Visual Arts (SVA), developed the world's first Spatial AI Industry Map and interactive database. The taxonomy spans research-focused foundation 3D models (such as World Labs, Meta, Grok, and OpenAI) alongside production-ready generative 3D tools. It incorporates physical AI and robotics, given that navigating real or virtual environments requires identical 3D spatial capabilities. The map also features media and gaming applications built on 3D game engines, as well as ecosystem enablers like Jensen Huang's Nvidia Omniverse for enterprise warehouse/factory simulations and synthetic data generation for autonomous vehicles [12:34, 13:51].
- **Core Definition:** Spatial AI = AI ∩ 3D.
- **Key Industry Collaborator:** Michael Goet (Professor at School of Visual Arts).
- **Ecosystem Categories:** Foundation 3D Models, Generative 3D, Physical AI & Robotics, Media & Gaming, and Enterprise/Synthetic Data Enablers.
---
[14:18](https://www.youtube.com/watch?v=PI49GKzzRcs&t=858s) Why Spatial AI is a Pathway to AGI
In his closing section, Wang shares why he is uniquely passionate about Spatial AI and how it solves fundamental limitations of current AI architectures. Over 90% of current AI systems rely on Large Language Models (LLMs), which operate on one-dimensional text data. Because LLMs process the world through text-based reasoning, they lack true spatial, temporal, or physical understanding.
Wang contrasts this with human intelligence. Human babies learn about space, 3D object interactions, and sensory input (touch, sight, sound, smell) long before acquiring language. Standard AI models invert this natural learning process. By creating 3D-native Spatial AI models that directly take 3D data as input and output 3D structures without converting to 1D text, AI can gain innate spatial reasoning. Wang references Chief AI Scientist Yann LeCun's public assertion that LLMs alone will never achieve Artificial General Intelligence (AGI) due to missing causal and spatial physics reasoning. Wang concludes that 3D-native Spatial AI provides the crucial missing link and a viable alternative pathway toward true AGI.
- **Current AI Limitation:** >90% of current systems rely on 1D language inputs/outputs.
- **Human Learning vs. LLMs:** Humans develop 3D spatial intuition and multi-sensory awareness before learning language.
- **Yann LeCun's Perspective:** LLMs cannot reach AGI because they lack spatial reasoning, physical grounding, and causality.
- **Direct Quote:** "To me, I think a 3D native spatial AI model is the solution and it's a potential pathway to AGI."
The transcript snippet provided contains only a brief concluding remark and applause wrapping up the talk:
[19:33](https://www.youtube.com/watch?v=PI49GKzzRcs&t=1173s) Closing and Applause
Ryan Wang concludes his presentation, thanking the audience for their time and attention. The host then invites the attendees to give a warm round of applause for the speaker, bringing the session on Spatial AI to a close.
## Video description
Ryan Wang, Founding Partner, Outpost Capital
Abstracts
What is Spatial AI? What does it mean for AI to operate with a sense of space, context, and physical presence?
This talk will examine the evolution of Spatial AI as a foundational layer for AR/VR, robotics, autonomous systems, and next-generation interfaces. I will outline the current industry landscape, highlight emerging technologies and companies, and present an investment thesis grounded in technical readiness, market timing, and long-term impact.
Attendees will gain a structured view of where Spatial AI is creating value today, and how to assess opportunities in this rapidly developing frontier.
The session is designed for investors, founders, and strategists looking to navigate and engage with this critical shift in the AI ecosystem.
空间智能（Spatial AI）：AI新前沿的行业全景与投资逻辑
Outpost Capital 创始合伙人 Ryan Wang 指出传统 AI 赛道竞争高度内卷，呼吁转向被低估的蓝海前沿“空间智能”（Spatial AI，即 AI 与 3D 的交集），并发布了涵盖 3D 基础模型、物理 AI 与机器人等方向的首份行业全景图。他认为现有依赖一维文本 LLM 的架构缺乏物理与因果理解，而原生 3D Spatial AI 模型有望补全物理常识，成为通向 AGI 的关键路径。
Outpost Capital 创始合伙人 Ryan Wang 指出当前传统 AI 技术栈高度内卷且叙事单一，呼吁投资人与创业者避开 RAG 及 AI Agent 等拥挤红海，转向被严重低估的蓝海前沿——空间智能（Spatial AI）。 (0:04)
Wang 将 Spatial AI 核心定义为“人工智能与 3D 的交集（AI ∩ 3D）”，并联合发布了首份行业全景图，覆盖 3D 基础模型、物理 AI 与机器人、传媒游戏以及 Nvidia Omniverse 等合成数据与模拟生态。 (5:13)
现有超过 90% 的 AI 系统依赖处理一维文本的语言模型（LLM），颠倒了人类先建立 3D 感官与空间直觉、后掌握语言的自然认知规律，导致模型缺乏真实的物理与因果理解。 (14:18)
结合 Yann LeCun 关于单一文本 LLM 无法通向通用人工智能（AGI）的论点，Wang 认为直接输入与输出 3D 数据的原生 Spatial AI 模型是补全物理常识的关键拼图，也是迈向 AGI 的潜在路径。 (14:18)
Reading summary
Ryan Wang of Outpost Capital argues that Spatial AI—defined as the intersection of artificial intelligence and 3D data—offers a differentiated frontier that addresses fundamental physical reasoning limitations inherent in text-based large language models.
Moving Beyond Saturated Conventional AI Stacks (0:04)
Wang critiques standard venture capital approaches to the AI market, which classify the ecosystem into infrastructure, developer tooling, and enterprise or consumer applications. He argues that this conventional framework has become severely overcrowded, pointing to the Ai4 conference expanding from 2,000 to 8,000 attendees and rapid shifts from Retrieval-Augmented Generation (RAG) to AI agents as evidence of herd behavior. Rather than competing in crowded "red ocean" categories where dozens of startups build identical products, Wang contends early-stage investors must identify "blue ocean" opportunities in emerging technical domains.
Taxonomy of the 3D-Native AI Ecosystem (5:13)
Wang defines Spatial AI strictly as the intersection between artificial intelligence and 3D. He distinguishes this from AR/VR hardware that serves as an immersive media layer, as well as narrower academic definitions of spatial intelligence. Together with School of Visual Arts professor Michael Goet, Wang mapped the emerging sector across several core verticals: foundation 3D models (such as World Labs, Meta, and OpenAI), generative 3D production tools, physical AI and robotics navigating real and virtual environments, and media or gaming applications built on 3D game engines.
The framework also identifies industrial ecosystem enablers, pointing to platforms like Nvidia Omniverse that provide factory simulation and synthetic data generation for autonomous vehicles as critical infrastructure for bridging spatial models with real-world deployments.
Grounded 3D Processing as a Route to AGI (14:18)
Wang identifies current language-centric architectures as a core bottleneck in AI advancement. Over 90% of contemporary AI systems rely on large language models (LLMs) processing one-dimensional text strings, which Wang argues deprives them of spatial, temporal, and physical grounding. He contrasts this with human cognition, noting that infants develop 3D spatial intuition and multi-sensory awareness long before developing language, meaning standard LLMs invert natural cognitive development.
Citing Meta Chief AI Scientist Yann LeCun's position that LLMs cannot reach Artificial General Intelligence (AGI) due to missing causal and physical reasoning, Wang asserts that 3D-native Spatial AI—models that directly ingest and output 3D representations without intermediate text conversion—serves as the necessary technical foundation and a viable pathway to AGI.
Wang positions 3D-native Spatial AI as an essential corrective to the structural constraints of text-based models, though its adoption depends on the ongoing maturation of 3D foundation models, robotics integration, and industrial simulation ecosystems.
精炼阅读摘要
Outpost Capital 创始合伙人 Ryan Wang 提出，空间人工智能（Spatial AI，即人工智能与 3D 的交集）是当前被忽视的新前沿，有望弥补传统一维语言模型的物理认知缺陷，成为通向通用人工智能（AGI）的潜在路径。
传统 AI 架构的红海竞争与蓝海需求 (0:04)
Ryan Wang 指出，当前风险投资领域普遍声称布局人工智能，但若缺乏具体的细分赛道聚焦，此类表态已失去实质意义。目前的行业通用框架将 AI 划分为基础设施层、中间层开发者工具/DevOps 以及分为消费端和企业垂直端的应用层。Wang 认为这一框架在各层级均已过度拥挤，将整个行业生态简化为单一赛道的同质化叙事。例如 Ai4 大会参会人数从 2,000 人激增至 8,000 人，行业热点也迅速从去年的检索增强生成（RAG）转向今年的 AI 智能体（AI agents）。
Wang 强调，在数十家竞争对手制造完全相同解决方案的红海市场中竞争并不可取，创业者和投资机构应当跳出既有框架，积极寻找未被充分开发的“蓝海”机会。
空间 AI 的核心定义与产业生态图谱 (5:13)
在概念界定上，Wang 将空间 AI 明确定义为“人工智能与 3D 的交集”（AI ∩ 3D）。他将该概念与关联领域进行了区分：Apple 和 Meta Reality Labs 的空间计算（AR/VR）侧重于沉浸式内容的媒体交互层；李飞飞博士及 World Labs 提出的“空间智能”则偏向学术研究范畴的 3D 基础模型；而空间 AI 的本质是为 3D 赋予人工智能能力。
为厘清行业全貌，Wang 与纽约视觉艺术学院（SVA）教授 Michael Goet 共同制定了全球首个空间 AI 产业图谱及交互数据库。该生态涵盖五个主要方向：以 World Labs、Meta、Grok、OpenAI 为代表的研究型 3D 基础模型与生成式 3D 工具；因虚拟与实体导航具备相同 3D 需求而涵盖的物理 AI 与机器人；基于 Unity 和虚幻引擎等构建的消费级媒体与游戏；以及提供企业级工厂/仓储模拟的 Nvidia Omniverse 和自动驾驶合成数据生成等生态赋能平台。
突破一维文本局限与迈向 AGI 的路径 (14:18)
Wang 认为空间 AI 具备解决当前底层模型缺陷的关键潜力。目前超过 90% 的主流系统依赖大语言模型（LLM），本质上是在一维文本数据上运行，依赖基于文字的推理方式，缺乏对空间、时间和物理规律的真实感知。
相比之下，人类婴儿在习得语言之前，便已通过触觉、视觉、听觉和嗅觉等感知机制建立起对 3D 空间与物体交互的本能理解，而现行大模型颠倒了这一自然认知顺序。结合 Meta 首席 AI 科学家 Yann LeCun 关于“LLM 因缺乏因果机制与空间物理推理而无法实现 AGI”的论断，Wang 提出，直接输入并输出 3D 数据、无需转化为一维文本的原生 3D 空间 AI 模型，能够赋予机器内生的空间推理能力，是通往 AGI 的关键拼图与可行替代路径。
Wang 将原生 3D 空间 AI 视作弥合大语言模型认知缺陷并通向 AGI 的核心方向，但这一观点属于前沿技术路线与投资视角的探索性主张，该技术在实际场景中的演进速度及行业接受度仍有待后续观察。
Detailed notes — zh
## Background
本视频由 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台。演讲由 Outpost Capital 创始合伙人 Ryan Wang 主讲，深入探讨了空间 AI（Spatial AI）的定义演进、行业全景图谱，以及其作为 AR/VR、机器人和次时代交互界面底层基础设施的投资逻辑与市场机遇。
## TL;DR
演讲直击当前 AI 创投界在基础设施、中间件与应用三层架构上的严重内卷与跟风困局，提出将“AI 与 3D 空间的交集”——空间 AI 作为跳出红海的全新蓝海赛道。Ryan 系统性梳理了基础 3D 模型、生成式 3D、物理 AI/机器人、工业数字孪生及合成数据等核心板块的产业格局。他尖锐地指出当前 90% 建立在 1D 线性文本上的大语言模型无法真正理解三维物理世界，唯有构建“3D 进、3D 出”的原生空间基础模型，弥补物理常识与空间几何推理的断层，才是通往通用人工智能（AGI）的终极突破路径。
## Atmosphere & style
整场演讲风格鲜明犀利且极具前瞻洞察力，讲者以饱满的热情和笃定的投资人视角，瞬间打破了行业大会中常见的套话与炒作泡沫。叙事层层递进，从剖析同质化竞争的痛点，到厘清学术与产业概念，再到上升至人类认知哲学与 AGI 底层逻辑，逻辑缜密而发人深省。现场氛围既带有揭示“潜伏赛道”的探索兴奋感，又传递出一种在技术周期狂热中保持清醒的理性思辨气息。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=PI49GKzzRcs&t=4s) 空间 AI（Spatial AI）目前处于被市场低估且少有人深入关注的状态，有望成为驱动下一波科技浪潮的核心蓝海赛道。
- [00:48](https://www.youtube.com/watch?v=PI49GKzzRcs&t=48s) 传统 AI 风险投资已陷入同质化的红海竞争，行业叙事呈现单一线性的跟风趋势（如从去年的 RAG 全面转向今年的 AI Agent），缺乏长期差异化价值。
- [05:10](https://www.youtube.com/watch?v=PI49GKzzRcs&t=310s) 空间 AI 被凝练定义为“人工智能与 3D 维度的交集”，区别于侧重沉浸式硬件交互的“空间计算”和偏学术研究的“空间智能”。
- [07:33](https://www.youtube.com/watch?v=PI49GKzzRcs&t=453s) 空间 AI 涵盖四大核心支柱：基础 3D 模型与实验室、生成式 3D 生产工具、物理 AI 与具身智能机器人，以及下一代沉浸式交互接口（XR/BCI）。
- [11:25](https://www.youtube.com/watch?v=PI49GKzzRcs&t=685s) 在消费级应用端，媒体娱乐制作与依托游戏引擎的“AI for Gaming”因其天然的原生 3D 环境，成为空间 AI 最先落地的核心场景。
- [12:30](https://www.youtube.com/watch?v=PI49GKzzRcs&t=750s) 在企业级生态端，工业数字孪生（如 NVIDIA 仿真平台）与高质量 3D 合成数据生成构成了驱动机器人与自动驾驶训练的关键支撑。
- [14:18](https://www.youtube.com/watch?v=PI49GKzzRcs&t=858s) 主流大语言模型（LLM）本质是 1D 线性文本系统，无法真正建立对立体空间、因果几何与物理常识的深层理解，与人类原生先验的多感官 3D 认知路径背道而驰。
- [17:20](https://www.youtube.com/watch?v=PI49GKzzRcs&t=1040s) 构建原生“3D 进、3D 出”的基础空间模型是解决当前模型渐进式创新天花板、真正通往通用人工智能（AGI）的关键路径。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=PI49GKzzRcs&t=4s) 开场与背景：空间 AI 这一被低估的前沿赛道
演讲伊始，Outpost Capital 创始合伙人 Ryan Wang 热情地向全场观众介绍了本次主题——空间 AI（Spatial AI）。在他眼中，这是整个人工智能领域中最令他兴奋的方向之一。然而在当前的创投热潮中，空间 AI 却处于一种极度被低估且少有人关注的“潜伏”状态（under the radar），是一颗名副其实的“隐藏宝石”（hidden gem），极有可能成为驱动下一波科技浪潮的核心力量。Ryan 表示将利用接下来的演讲，深入阐述为什么空间 AI 会成为下一个重磅趋势，以及创业者与投资人应当如何切入并审视这一领域。
Ryan 简要介绍了自己的背景与机构定位。Outpost Capital 总部位于硅谷湾区，长期专注于前沿科技（Frontier Technology）的早期投资，核心关注赛道覆盖空间计算（Spatial Computing）、机器人（Robotics）、人工智能（AI）以及区块链（Blockchain）。正因长期扎根于底层前沿交叉领域，团队对技术周期的演进与范式转移有着敏锐的洞察。
### [00:48](https://www.youtube.com/watch?v=PI49GKzzRcs&t=48s) 传统 AI 投资框架的同质化与“红海”困局
谈及当下的创投环境，Ryan 直言不讳地指出，几乎所有风险投资机构都宣称自己是“AI 投资人”，但这在今天已经毫无实质意义，“因为说实话，现在谁不投 AI 呢？”真正关键的问题在于“你在 AI 领域究竟投什么”。当向绝大多数 AI 投资人提出这个问题时，十有八九会得到一套极其同质化、标准化且令人审美疲劳的三层架构图：底层是基础设施层（Infrastructure），中间是开发者工具与 DevOps 构成的中间件层（Middleware），顶层则是面向各种垂直行业的企业级与消费者端应用层（Application）。
```mermaid
flowchart TD
accTitle: 传统 AI 赛道拥挤现状
subgraph TraditionalAI["传统 AI 产业三层架构（高度内卷的红海）"]
App["应用层：消费级应用 & 企业各垂直应用"]
Mid["中间件层：DevOps / 开发者工具"]
Infra["基础设施层：算力 / 基础模型"]
Infra --> Mid --> App
end
subgraph ParadigmShift["破局路径"]
BlueOcean["空间 AI（AI 与 3D 物理/虚拟空间的交汇点）"]
end
TraditionalAI -.->|"跳出单线叙事 / 寻找新蓝海"| BlueOcean
```
Ryan 强调，这种传统的划分方式和叙事逻辑不仅让创业生态陷入严重的“红海竞争”，而且整个行业的注意力呈现出单一线性的跟风趋势。以本届 Ai4 大会为例，参会人数从过去的 2,000 人暴增至 8,000 人，展厅内无数团队在高度重叠的方向上激烈厮杀。去年全行业都在言必称 RAG（检索增强生成），讨论如何连接内部数据库与大语言模型；而今年几乎每一场演讲都在谈论 AI Agent。对于创业者而言，与几十家同质化对手在拥挤赛道竞争是极度危险的；而前沿科技投资的核心使命，正是要打破这种固化的认知框架，跳出当下的短期风口，去提前布局 2 年、3 年乃至 5 年后的下一片蓝海。
- 行业会议规模剧增：Ai4 参会人数从早期的 2,000 人激增至当届的 8,000 人，行业竞争白热化。
- 叙事单线条更迭：去年行业全面聚焦 RAG 技术，今年则全盘转向 AI Agent。
- 投资逻辑反思：避免在基础设施、中间件和常规应用层与数十家竞品正面肉搏，聚焦前瞻性蓝海。
### [05:10](https://www.youtube.com/watch?v=PI49GKzzRcs&t=310s) 厘清概念：何为空间 AI（Spatial AI）？
为了打破既有框架，Ryan 带领现场观众重新界定“空间 AI”的内涵。当他在现场提问有多少人听过 Spatial AI 时，约有一半观众举手。针对市场上的概念混淆，Ryan 逐一进行了对比与剖析。首先是 Apple、Meta Reality Labs 及 Oculus 等硬件巨头主推的“空间计算”（Spatial Computing），它更偏向于媒体与交互层，旨在打造沉浸式的 VR/AR/XR 数字化体验，但缺少深度的机器智能内核；其次是李飞飞（Fei-Fei Li）创立 World Labs 时所倡导的“空间智能”（Spatial Intelligence），这主要是从学术与前沿科研视角出发，探讨如何让基础模型具备原生 3D 理解与世界建模能力。
在抛开繁琐枯燥的学术术语后，Ryan 给出了极度凝练而清晰的定义：“空间 AI 就是人工智能与 3D 的交集（the intersection between artificial intelligence and 3D）”。只要涉及让 AI 系统理解、生成、交互并穿梭于三维空间（无论是虚拟三维世界还是物理现实世界），都归属于空间 AI 的范畴。这一唯简的定义不仅打破了原有以文本、二维图像为主的技术边界，也为 AI 的演进打开了全维度的发展空间。
| 概念名称 | 核心侧重点 | 典型代表 / 提出背景 | 与空间 AI 的本质差异 |
|:--- |:--- |:--- |:--- |
| **空间计算 (Spatial Computing)** | 媒体层与交互层，注重沉浸式显示与软硬件接口 | Apple Vision Pro、Meta Reality Labs | 侧重新型媒体与交互形式，缺少深度的认知智能底座 |
| **空间智能 (Spatial Intelligence)** | 科研与学术视角，探索大模型理解三维物理世界的机理 | 李飞飞（Fei-Fei Li）、World Labs | 偏向学术研究与基础大模型层面的理论探索 |
| **空间 AI (Spatial AI)** | **AI 与 3D 维度的全面融合与产业应用** | Outpost Capital 产业图谱 | **兼具智能模型与空间属性，涵盖虚拟与物理多维应用** |
### [07:33](https://www.youtube.com/watch?v=PI49GKzzRcs&t=453s) 空间 AI 行业全景图与核心技术板块拆解
为了系统化梳理该赛道，Ryan 与纽约视觉艺术学院（SVA）的 Michael 教授合作，联合推出了全球首个“空间 AI 行业全景图谱”（Spatial AI Industry Map），并现场演示了其交互式数据库网站。该平台收录了全球范围内代表性的空间 AI 创新企业与实验室档案，旨在为全行业的创业者与投资人提供一站式研究与参考索引。
```mermaid
flowchart LR
accTitle: 空间 AI 核心构成板块
subgraph SpatialAI["空间 AI 核心技术范畴"]
direction TB
F3D["基础 3D 模型与前沿实验室<br/>(World Labs, Meta, xAI, OpenAI)"]
G3D["生成式 3D 技术与生产工具<br/>(多模态演进：文本/音视频 ➔ 3D)"]
PAI["具身智能与物理 AI<br/>(机器人、自动驾驶、物理世界导航)"]
XR["沉浸式交互接口<br/>(VR / AR / 脑机接口 BCI)"]
end
```
Ryan 随后对全景图中的核心技术板块进行了详细拆解：
1. **基础 3D 模型与前沿实验室（Foundation 3D Models & AI Labs）**：处于偏基础科研的前沿阵地，以李飞飞的 World Labs 为典型代表，致力于打造能够原生输出 3D 物体与三维世界场景的下一代基础模型；Meta、xAI（Grok）以及 OpenAI 内部同样组建了专门的团队攻坚空间 AI。
2. **生成式 3D（Generative 3D）**：相较于基础科研，这一板块更加偏向工程落地与产业级生产应用。多模态 AI 在跨越了语言、音频、二维图像与视频之后，进入 3D 维度是必然的技术跃迁。
3. **物理 AI 与机器人（Physical AI & Robotics）**：现实物理世界本质上就是一个支持实时交互的三维环境。赋予 AI 理解和导航 3D 空间的能力，正是机器人与物理 AI 的核心底座。无论 AI 是在虚拟仿真空间还是在现实物理世界中作业，其底层的 3D 空间表征与运动规划算法具有高度的共通性。
4. **下一代交互接口**：涵盖 VR、AR 以及脑机接口（BCI）等，构成人类感知与控制三维智能体的重要通路。
### [11:25](https://www.youtube.com/watch?v=PI49GKzzRcs&t=685s) 消费端应用：媒体娱乐与原生 3D 游戏引擎
进入消费端应用领域，媒体与娱乐（Media & Entertainment）是一个规模极其庞大且历史悠久的落地场景。这一领域并非新兴产物，而是经历了数十年的技术沉淀与演进，从传统电子游戏到各类新媒体动画。如今，借助 AI 技术，创作者已经可以直接生成 3D 动画与多样化的物理模拟，彻底改变了影视节目的制作与生产流程。
另一个核心消费级赛道是“游戏 AI（AI for Gaming）”。众所周知，现代绝大多数游戏都是构建在各类游戏引擎之上的。由于游戏引擎天生就是一个 3D 原生环境，因此该领域的 AI 探索与应用自然而然地完全契合空间智能（Spatial AI）的技术范畴。
- **媒体娱乐制作**：利用 AI 自动化生成 3D 动画、角色与动态模拟，重塑传统内容管线。
- **游戏开发与引擎**：依托 Unreal、Unity 等原生 3D 游戏引擎，构建空间交互和智能 NPC 生态。
---
### [12:30](https://www.youtube.com/watch?v=PI49GKzzRcs&t=750s) 生态赋能者：工业数字孪生、大厂布局与合成数据
在基础设施与生态赋能者（Ecosystem Enablers）层面，企业级应用展现出了巨大的商业价值，尤其是面向工业场景的大规模仿真模拟。NVIDIA 创始人黄仁勋（Jensen Huang）便是该方向的强力倡导者。他在 GTC 大会上可以一连演讲三四个小时，甚至连洗手间都不用去，期间展示了大量通过虚拟仿真驱动实体产业的案例——为智慧工厂、大型物流仓储等构建 1:1 的工业级数字孪生环境。在虚拟环境中完成搭建后，企业便可针对机器人及系统展开海量的训练任务、流程优化与前瞻预测。
除了科技巨头的生态建设，企业级数据服务也是核心一环，特别是合成数据（Synthetic Data）的生成与标注。以自动驾驶汽车（AV）的训练为例，当今绝大部分训练数据并非直接采集自真实世界，而是在 3D 模拟环境中通过算法合成生成的。这种基于空间仿真的数据生成能力，构成了整个空间智能生态的关键支撑。
```mermaid
flowchart TD
accTitle: 空间智能工业仿真与合成数据闭环
subgraph 虚拟仿真层["3D 原生仿真环境 (如 NVIDIA Omniverse)"]
A["工业数字孪生<br/>(智能工厂 / 仓储空间)"]
B["高保真 3D 合成数据生成"]
end
subgraph 赋能应用层["智能实体与业务优化"]
C["自动驾驶算法训练<br/>(非现实路采数据)"]
D["机器人操作与路径优化"]
E["前瞻性产能与安全预测"]
end
A --> B
B --> C
A --> D
A --> E
```
- **工业仿真模拟**：利用虚拟数字孪生重构物理工厂与仓储，低成本完成算法压力测试与流程推演。
- **合成数据主导训练**：现代自动驾驶的大多数长尾极端场景（Corner Cases）依赖 3D 模拟合成，而非实车实采。
---
### [14:18](https://www.youtube.com/watch?v=PI49GKzzRcs&t=858s) 穿透炒作看本质：1D 语言模型与人类 3D 感知的根本鸿沟
抛开行业当下的喧嚣与炒作，空间智能之所以具有无可替代的战略地位，源于当前主流大语言模型（LLM）的底层局限。目前全球超过 90% 的主流 AI 系统都建立在 LLM 之上，但语言模型的本质是“以一维（1D）语言文本为输入，再输出一维语言文本”的基础模型。无论 LLM 表现得多么强大，即便为其加上视觉编码插件，它眼中的世界归根结底依然是扁平的一维符号推演，缺乏对真实物理世界与空间几何的深层理解。
相比之下，人类智能的发育过程则截然相反。婴儿在尚未学会说话、行走之前，就已经在直观地与三维空间发生交互——婴儿能够直接感知眼前的立体物体、判断自身在空间中的相对位置。人类拥有视觉、触觉、嗅觉、听觉等多维度、多层级、多感官的感知体验，3D 认知是与生俱来的本能，而语言反而是后天习得的符号系统。现行 AI 基础模型“先学语言、再试图理解空间”的路线，完全颠倒了智能发展的认知顺序。
- **LLM 的维度局限**：基于 1D 线性文本的推理机制，缺乏对空间因果律和物理尺度的直观建模。
- **人类认知演进路径**：原生 3D 空间直觉与多感官交互先行，离散语言符号学习在后。
---
### [17:20](https://www.youtube.com/watch?v=PI49GKzzRcs&t=1040s) 原生 3D 空间基础模型：解锁 AGI 的终极路径
真正通往通用人工智能（AGI）的突破口，在于构建原生（Native）的 3D 空间 AI 基础模型。这种模型直接以原生 3D 数据作为输入并输出 3D 数据，无需将其强行降维翻译为一维文本。此类基础模型在底层即内嵌了空间推理（Spatial Reasoning）、几何坐标与物理世界规律的认知机制，其对世界的理解与学习逻辑将与人类智能高度一致。
近期 GPT-5 等前沿文本模型的进展放缓，暴露出渐进式创新的天花板。图灵奖得主 Yann LeCun 也曾多次公开指出：仅凭大语言模型永远不可能实现 AGI，因为文本模型不懂物理世界、缺乏空间坐标与空间推理能力。唯有原生 3D 空间 AI 模型，才能补全物理常识这一决定性短板，成为迈向 AGI 的核心基石。
演讲最后，Ryan Wang 展示了由他和合伙人 Michael 共同整理的空间智能行业全景生态图（Spatial AI Landscape Map），并开放了早期访问二维码，邀请现场投资者与创业者共同交流探讨，推进该前沿领域的投资与建设。
- **原生 3D 架构**：实现“3D 进、3D 出”的端到端建模，在神经网络内部原生编码物理常识与空间几何。
- **突破 AGI 瓶颈**：响应 Yann LeCun 等学者的批判，将 AI 的核心驱动力从纯文本符号统计转向具身空间物理理解。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=PI49GKzzRcs&t=4s) 空间 AI 是整个人工智能领域中最令我兴奋的方向之一。 — Ryan Wang
> [00:48](https://www.youtube.com/watch?v=PI49GKzzRcs&t=48s) 因为说实话，现在谁不投 AI 呢？真正关键的问题在于你在 AI 领域究竟投什么。 — Ryan Wang
> [00:48](https://www.youtube.com/watch?v=PI49GKzzRcs&t=48s) 前沿科技投资的核心使命，正是要打破这种固化的认知框架，跳出当下的短期风口，去提前布局 2 年、3 年乃至 5 年后的下一片蓝海。 — Ryan Wang
> [05:10](https://www.youtube.com/watch?v=PI49GKzzRcs&t=310s) 空间 AI 就是人工智能与 3D 的交集（the intersection between artificial intelligence and 3D）。 — Ryan Wang
> [07:33](https://www.youtube.com/watch?v=PI49GKzzRcs&t=453s) 多模态 AI 在跨越了语言、音频、二维图像与视频之后，进入 3D 维度是必然的技术跃迁。 — Ryan Wang
> [14:18](https://www.youtube.com/watch?v=PI49GKzzRcs&t=858s) 以一维（1D）语言文本为输入，再输出一维语言文本。 — Ryan Wang
> [14:18](https://www.youtube.com/watch?v=PI49GKzzRcs&t=858s) 人类拥有视觉、触觉、嗅觉、听觉等多维度、多层级、多感官的感知体验，3D 认知是与生俱来的本能，而语言反而是后天习得的符号系统。 — Ryan Wang
> [17:20](https://www.youtube.com/watch?v=PI49GKzzRcs&t=1040s) 3D 进、3D 出 — Ryan Wang
> [17:20](https://www.youtube.com/watch?v=PI49GKzzRcs&t=1040s) 仅凭大语言模型永远不可能实现 AGI，因为文本模型不懂物理世界、缺乏空间坐标与空间推理能力。 — Yann LeCun
## Mentioned
Ryan Wang, Outpost Capital, Ai4, Apple, Meta Reality Labs, Oculus, 李飞飞 (Fei-Fei Li), World Labs, Apple Vision Pro, 纽约视觉艺术学院 (SVA), Michael, xAI (Grok), OpenAI, Unreal Engine, Unity, NVIDIA, 黄仁勋 (Jensen Huang), NVIDIA Omniverse, Yann LeCun, GPT-5
## Video description
Ryan Wang, Founding Partner, Outpost Capital
Abstracts
What is Spatial AI? What does it mean for AI to operate with a sense of space, context, and physical presence?
This talk will examine the evolution of Spatial AI as a foundational layer for AR/VR, robotics, autonomous systems, and next-generation interfaces. I will outline the current industry landscape, highlight emerging technologies and companies, and present an investment thesis grounded in technical readiness, market timing, and long-term impact.
Attendees will gain a structured view of where Spatial AI is creating value today, and how to assess opportunities in this rapidly developing frontier.
The session is designed for investors, founders, and strategists looking to navigate and engage with this critical shift in the AI ecosystem.
[Source](https://www.youtube.com/watch?v=PI49GKzzRcs)
Source quotes
[0:04] Ryan Wang: 空间 AI 是整个人工智能领域中最令我兴奋的方向之一。
[0:48] Ryan Wang: 因为说实话，现在谁不投 AI 呢？真正关键的问题在于你在 AI 领域究竟投什么。
[0:48] Ryan Wang: 前沿科技投资的核心使命，正是要打破这种固化的认知框架，跳出当下的短期风口，去提前布局 2 年、3 年乃至 5 年后的下一片蓝海。
[5:10] Ryan Wang: 空间 AI 就是人工智能与 3D 的交集（the intersection between artificial intelligence and 3D）。
[7:33] Ryan Wang: 多模态 AI 在跨越了语言、音频、二维图像与视频之后，进入 3D 维度是必然的技术跃迁。
[14:18] Ryan Wang: 以一维（1D）语言文本为输入，再输出一维语言文本。
[14:18] Ryan Wang: 人类拥有视觉、触觉、嗅觉、听觉等多维度、多层级、多感官的感知体验，3D 认知是与生俱来的本能，而语言反而是后天习得的符号系统。
[17:20] Ryan Wang: 3D 进、3D 出
[17:20] Yann LeCun: 仅凭大语言模型永远不可能实现 AGI，因为文本模型不懂物理世界、缺乏空间坐标与空间推理能力。
Original transcript
[0:04] So yeah, welcome to this uh session
[0:06] about spatial AI. I think it's uh one of
[0:09] my favorite topic in the AI space. I
[0:12] think it's very undervalued, very under
[0:15] the radar. It's a hidden gem and it
[0:17] could be the next big thing. And I'm
[0:19] going to use the next 20 minutes to tell
[0:21] you why I think it's a next big thing
[0:24] and uh how we should look into it.
[0:29] Just a quick intro on myself. My name is
[0:31] Ryan Wong. I'm the founding partner with
[0:33] a VC firm called Outpost Capital based
[0:36] in the Bay Area. And we focus on early
[0:39] stage frontier technology investments
[0:42] such as uh spatial computing, robotics,
[0:45] AI and uh blockchain.
[0:48] So you see um I'm sure like throughout
[0:51] this conference you met a lot of AI
[0:53] investors, right? And um and I think uh
[0:56] you know nowadays when we talk to
[0:58] different uh venture capitalists they
[1:00] all claim to be a AI investor as well
[1:02] but uh it's a little bit um meaningless
[1:06] to claim yourself to be a AI investor
[1:08] because honestly like nowadays who
[1:11] doesn't invest in AI right so the real
[1:13] question is what in AI do you invest and
[1:17] I think you know uh when you talk to any
[1:19] AI investors this is one of the first
[1:21] question that you should ask as
[1:24] And um nine out of 10 times I think uh
[1:28] the answer that I've been hearing is
[1:31] something like this.
[1:34] Something can be kind of like you know u
[1:37] uh interpreted using a graph like this.
[1:39] I'm sure you see different version of
[1:41] this kind of graph, right? But the the
[1:43] idea is that um you know there's a
[1:46] infrastructure layer right and on top of
[1:48] infrastructure infrastructure you have
[1:50] this kind of middle layer where you can
[1:52] do devops you can do like um developer
[1:55] tools things like that and on the
[1:57] application layer you have the consumer
[2:00] side right all kinds of consumer
[2:01] application and all kinds of enterprise
[2:03] application for different enterprise
[2:05] verticals right so I'm sure we all heard
[2:07] about the story and I think you know
[2:10] nine out of 10 times when you ask any AI
[2:12] investor
[2:13] what do you invest? Their answer falls
[2:16] into something that can be interpreted
[2:18] by this chart, right? And I don't like
[2:21] it. Why I don't like it? because I think
[2:24] uh you know it's um
[2:27] one the narrative is too similar and two
[2:31] um I think you at the keynote yesterday
[2:33] we said hey right now we are at the
[2:35] largest AI conference and this AI
[2:38] conference has grew from uh 2,000 people
[2:40] to 8,000 people this year right so
[2:43] obviously the industry has come a very
[2:45] long way and uh you know so many people
[2:48] working on so many different projects so
[2:51] many different problems right the thing
[2:53] is it's getting so crowded. It's getting
[2:56] so crowded at the infrastructure. It's
[2:59] getting so crowded at the middleware and
[3:01] it's getting so crowded at consumer
[3:04] applications and enterprise
[3:05] applications. So if you are
[3:09] entrepreneur, you don't want to be in a
[3:11] very crowded space because you know you
[3:12] have too many competitors, right? We
[3:15] want to identify the type of opportunity
[3:18] where it's less crowded, where it still
[3:20] feels that it's a it's a blue ocean
[3:23] opportunity. It's not a red ocean
[3:24] opportunity, right? You don't want to
[3:26] have like 20 30 other companies
[3:28] competing with you working on the same
[3:31] thing. So, it's getting too crowded. And
[3:33] I think uh you know this kind of
[3:35] interpretation and this kind of
[3:37] breakdown for the entire AI industry is
[3:40] also one single threaded right. So for
[3:44] for example last year everyone was
[3:46] talking about rack right that's the
[3:48] hottest thing you know how can we apply
[3:51] like a uh internal database and using AI
[3:54] to retrieve information from the
[3:56] internal database but still using the
[3:58] capability of RLM for some kind of you
[4:01] know enterprise use cases right hot
[4:02] thing last year everyone's talking about
[4:04] it and this year I haven't went to any
[4:08] single session at AI4 that haven't even
[4:11] talked about the word agent right every
[4:14] session AI agent right so I think you
[4:16] know it's just getting so crowded and
[4:18] everyone is telling the same story I'm
[4:20] just getting so frustrated with it and
[4:23] uh as a venture capitalist I think uh
[4:25] you know especially focus on frontier
[4:27] technology what is frontier right it's a
[4:29] next big thing right we should uh look a
[4:32] little bit ahead and try to figure out
[4:34] hey you know what's a what's interesting
[4:36] next right not only what is interesting
[4:39] today but uh like uh two years 3 years 5
[4:42] years down the road what is interesting
[4:45] at that time. So I think uh in order to
[4:48] answer that question, we need to kind of
[4:50] break the framework that I previous
[4:53] showed right and uh I know like um a lot
[4:57] of you were taking pictures of that uh
[4:59] slides but we need to break that
[5:01] framework and we need to look into this
[5:04] industry have a fresh new perspective
[5:07] right that's what we need to do so that
[5:10] comes to my topic spatial AI right what
[5:13] is spatial AI Okay.
[5:16] Um I think you know it's a um it's a new
[5:20] frontier but then how many of you have
[5:22] heard about the spatial AI? Show a raise
[5:24] of hand.
[5:26] Okay about half right. So I think you
[5:28] know different people have different
[5:29] understanding about spatial AI. Some of
[5:32] you feel like hey you know this spatial
[5:34] computing right the vision that uh you
[5:37] know Apple meta with a reality lab and
[5:40] Oculus right all these companies are
[5:42] trying to do right going to the VR AR XR
[5:46] side of the world make the whole you
[5:48] know digital media much much more
[5:50] immersive well I think it is kind of
[5:53] spatial computing but not um not quite
[5:56] right because you know we still need the
[5:58] intelligence layer right spatial
[6:00] computing is more a term interpreting
[6:02] around the media layer. It's a type of
[6:05] new media where we can interact with
[6:07] content or information, right? And then
[6:09] a lot of you this morning have obviously
[6:11] come to the Fifi le uh fireside chat
[6:14] which is very inspirational and we all
[6:16] know Fifi with our world labs have been
[6:18] coining a new term called a spatial
[6:20] intelligence right so is a spatial AI
[6:24] something the same or identical with
[6:27] spatial intelligence I think it's you
[6:30] know it's getting closer right but it's
[6:32] not quite there because uh when we talk
[6:34] about spatial intelligence it's really
[6:36] from a research and arch academia side
[6:39] of the view to try to understand hey
[6:42] when we are building all these kind of
[6:44] foundation models right can we uh can we
[6:46] make it a little bit more 3D native can
[6:48] we use this foundation models to solve
[6:50] some of the 3D based uh problems right
[6:53] so that's kind of like spatial
[6:54] intelligence so the question is what is
[6:59] spatial AI and to me like you know how
[7:01] would I define spatial AI
[7:05] and you know I won't bore you is with
[7:09] any kind of like terminologies and
[7:12] paragraph of definition. I love
[7:14] simplicity. So my definition about
[7:17] spatial AI is that simple. It is the
[7:21] intersection between artificial
[7:22] intelligence and 3D. So the intersection
[7:26] between AI and 3D that's spatial AI and
[7:30] that can cover quite a lot of things,
[7:33] right? Remember this one? I was saying
[7:36] that hey we should forget about this
[7:38] right so stop taking pictures
[7:41] uh okay get ready to take pictures so
[7:44] this is a new perspective and this will
[7:47] show you from a spatial perspective how
[7:51] we can look into the AI sector right and
[7:54] uh here you know me and my partner his
[7:56] name is Michael a professor from a um
[7:59] school of visual art we've been working
[8:01] together and we've been um trying to
[8:04] launch this uh world's first industry
[8:06] map for spatial AI, right? So, um we are
[8:10] making this kind of like a interactive
[8:12] website as well and uh down the road we
[8:15] are thinking about you know making this
[8:17] a um a u a database where you guys can
[8:21] just you know refer and uh and uh look
[8:24] up as well. So here's a short demo. So
[8:28] you can go to each of the companies,
[8:29] right? Check out the profile word less
[8:32] by Fifi Lee, right? Uh a lot of other
[8:34] companies, different categories and uh
[8:38] and uh also you know we have some ideas
[8:40] to make it more interactive so that uh
[8:42] eventually this will evolve into a
[8:45] database that uh you know for any of the
[8:47] companies you want to uh you want to
[8:49] learn more about you can use this as a
[8:51] as a as a reference point.
[8:54] So
[8:56] let's let me bring the mouse back.
[9:01] Yes, let's break let's break it down.
[9:04] What are there? So first of all, we have
[9:06] the foundation 3D models, right? What is
[9:09] foundation 3D models? Um I mean wart lab
[9:12] is a great example, right? It's a very
[9:14] researchoriented lab dedicated to build
[9:17] the next generation of foundation models
[9:19] that has, you know, much stronger 3D
[9:21] capabilities, right? a model not only
[9:24] able to output language but also able to
[9:27] output 3D objects and 3D wars as well,
[9:29] right? And uh there are you know uh in
[9:33] addition to the foundation models some
[9:35] of the foundation AI labs like meta
[9:38] right grock open AAI they all have some
[9:40] you know um uh spatial AI team um
[9:44] building internally and also there's a
[9:46] bigger category what I call as a
[9:48] generative 3D which is more um which is
[9:51] more application and production ready
[9:53] right so I think a little bit if you
[9:55] look at to the right side the the the
[9:57] the models and the labs is very research
[10:01] focused but the the um the the icons and
[10:04] companies on the left side is a little
[10:05] bit towards production side right so
[10:08] which is more on the from the
[10:10] multimodality perspective we already we
[10:13] already do a very good job in term of
[10:15] language audio um when we go into 2D
[10:19] right we have um images and videos right
[10:22] but you know 3D is a just a natural next
[10:25] step for um for uh for the multi
[10:28] multiodality Right. So I kind of you
[10:31] know put everything that in into a big
[10:33] bucket and uh also when I have a lot of
[10:37] conversations they will be like hey what
[10:39] about the robotics right what about this
[10:41] whole physical AI space and I think you
[10:44] know that make a lot of sense as well
[10:46] why because our physical world by
[10:50] default is also a 3D interactive world
[10:53] right so by having giving AI the
[10:55] capability to navigate our 3D
[10:58] interactive board. Essentially, we are
[11:01] also creating a spatial AI technology
[11:03] and whether you are navigating in the
[11:06] whether you are asking the AI to
[11:08] navigate a 3D world or a uh a real world
[11:11] or the virtual world is all 3D world.
[11:13] So, a lot of methodologies, a lot of you
[11:15] know core development is uh shared
[11:18] across the table, right? And in addition
[11:20] to robotics, VR, AR, we talked about
[11:22] that before, BCI, right? I think uh you
[11:25] know especially down the road can be a
[11:28] different alternative on how you know AI
[11:30] can navigate in the in the 3D world as
[11:32] well. Um then on the consumer side right
[11:37] now we are going getting into the
[11:39] application side. So on the consumer
[11:40] side media entertainment right is uh
[11:44] quite big use case and uh it's not a new
[11:46] use case it has decades of development
[11:49] as well right from you know uh gaming to
[11:52] different kind of new media animation
[11:54] right now right now we all see some kind
[11:56] of a productions where we can do 3D
[11:59] animation using AI right we can generate
[12:02] those kind of u um AI generated uh
[12:04] animation or different type of
[12:06] simulation so that we can produce shows
[12:08] differently as for right and uh also
[12:11] there's a big category around AI for
[12:13] gaming and uh as we all know most of the
[12:16] game are based on uh some kind of game
[12:18] engines and uh for game engine it's a 3D
[12:22] native environment so um obviously it
[12:26] falls into the spatial AI category as
[12:28] well
[12:30] and uh last but not least there are the
[12:34] ecosystem enablers right what does that
[12:36] mean a lot of enterprise applications
[12:39] like you know different uh enterprise
[12:41] simulations. uh Jinsen Huang if you any
[12:44] of you have went to GTC right he's
[12:46] talking for you know three four hours
[12:49] non-stop not even going to the restroom
[12:51] don't know how he he did that but um
[12:54] he's a big advocate in that space right
[12:57] um you know he's he's using a lot of
[12:59] example on how you can create um uh
[13:02] industry simulations right for your
[13:05] factories warehouse right and after you
[13:09] have created all these simulations then
[13:11] you can do a lot you know training
[13:13] related task. you can use a use a
[13:16] simulation to do some kind of
[13:17] optimization and prediction type of work
[13:20] right and then you know um a lot of use
[13:23] cases from that perspective in in
[13:26] addition to that big tech is always a a
[13:29] huge um um a huge ecosystem provider in
[13:33] the space and also you know uh
[13:35] enterprise data right and also is
[13:38] essentially um a lot of companies uh
[13:40] trying to develop new um data data
[13:43] generation data labeling type of um uh
[13:46] type of services. Um in particular on
[13:49] the synthetic side, right? How do we
[13:51] mostly train autonomous vehicle nowadays
[13:54] for example, right? Most of the training
[13:56] data um is actually not captured in real
[14:00] life, right? The majority of the
[14:01] training data is synthetically generated
[14:04] within a 3D simulated environment. So
[14:08] you know anything in that category is
[14:10] also part of this uh this big ecosystem.
[14:16] So
[14:18] um I think you know uh before I conclude
[14:21] my session, I want to share a little bit
[14:25] on um why am I so passionate about
[14:29] spatial AI and also most importantly um
[14:34] uh when we look beyond the hype right
[14:36] when we look into when we look beyond
[14:39] what is capable by all these different
[14:41] applications and different verticals
[14:43] that we just been through what's the the
[14:46] future that can potentially be unlocked
[14:50] by spatial AI and can only be unlocked
[14:53] by spatial AI.
[14:55] So you you know here's the thing. Um
[15:00] we've been talking about how AI how good
[15:03] AI is and uh you know um how it's useful
[15:07] today right how it's you know increasing
[15:09] productivities and uh drive the economy
[15:12] in all kinds of directions but currently
[15:17] 90% of the AI system owner over 90% of
[15:20] AI system are based on large language
[15:22] models and what is large language model
[15:25] it's a foundation model that takes Take
[15:27] language as input and take language as
[15:31] output and language is onedimensional
[15:35] data. So the way on how large language
[15:39] model works is solely based on text
[15:43] languagebased reasoning right there the
[15:47] no matter how powerful large language
[15:48] model is the vision the if the large
[15:52] language model has eyes right the word
[15:54] that the language model see um in this
[15:57] uh in our world is always 1D right so it
[16:02] has it lacks true understanding about
[16:04] the world but for human intelligence
[16:07] right for us we all know ourselves very
[16:09] well we know that we our intelligence
[16:12] develop very very differently so you
[16:15] know h how does human how does human
[16:18] intelligence develop how do humans learn
[16:20] when when we are a baby right when we um
[16:24] don't even know how to walk how to speak
[16:26] right we all have our interaction with
[16:29] the world already we all understand oh
[16:31] you know we are kind of in a 3D world
[16:34] right and there's some object objects in
[16:36] front of me and I'm in this kind of
[16:38] space right so everything come to our
[16:41] human intelligence 3D come to our human
[16:44] intelligence very very intuitively right
[16:48] so we don't need to really actually
[16:50] study but we we already interact with 3D
[16:53] and we also have our sensory experience
[16:56] right we can we can feel we can touch we
[16:58] can smell we can hear right so it's a
[17:01] multi-dimensional multi-layer
[17:03] multiensory type of experience
[17:05] And language is only learned later,
[17:08] right? So, you know, we learn language
[17:11] later on. And u um and uh it's kind of
[17:15] like the reverse of uh of foundational
[17:17] AI models, which is interesting. But
[17:20] here's a here's a wishful thinking,
[17:22] right? So, uh imagine one day we can get
[17:25] into some kind of spatial AI, some kind
[17:27] of foundation model that can take 3D
[17:30] input natively, right? without
[17:33] translating the 3D input into a
[17:36] one-dimensional space, right? We don't
[17:38] do that kind of conversion. Take 3D data
[17:40] as a input and take 3D data as output,
[17:44] right? So handle a real um spatial AI
[17:47] foundation model that that is capable of
[17:50] doing that and it will by default embed
[17:54] all these spatial reasoning capabilities
[17:58] within the model itself. Right? So the
[18:00] in the understanding the interpretation
[18:03] and the um the learning about the world
[18:06] is very very similar to human being
[18:09] right. So if we can get to that point, I
[18:12] think you know that potentially will
[18:14] unlock a new way for us to achieve AGI
[18:17] right because you know um I'm sure a lot
[18:19] of you have heard about how for example
[18:21] recently GPT5 has been a disappointment
[18:24] right not too much breakthroughs most of
[18:27] the most of the new features are
[18:28] incremental and um young Aquin has
[18:32] publicly said that uh through large
[18:35] language model we will never reach to
[18:36] AGI right because it doesn't understands
[18:39] space it doesn't doesn't understand you
[18:42] know spatial reasoning doesn't
[18:44] understand corate doesn't understand
[18:46] physics right so it has so many so many
[18:49] limitations but to me I think a 3D
[18:52] native spatial AI model is the solution
[18:56] and uh it's a potential pathway to AGI
[18:59] so yeah um this is a this this is a link
[19:02] to the to the map it's as you can see
[19:05] it's still a mess right so it's a um
[19:07] it's early version, but all of you guys,
[19:10] you have early access today by scanning
[19:12] the QR code. Um, and uh, yeah, if you
[19:15] want to learn more, chat more, this is
[19:17] me and my uh, my partner, Michael Goat.
[19:20] He's a um, we've been working
[19:23] collaboratively on this initiative and
[19:25] uh, that's our LinkedIn. So, yeah, feel
[19:27] free to reach out if you have any
[19:29] thoughts, questions, or um, yeah,
[19:32] anything you want to discuss. Thank you
[19:33] guys so much. And hopefully,
[19:35] >> of applause, please, for Mr. Dian Wang.
