我的征尘是星辰大海。。。

The dirt and dust from my pilgrimage forms oceans of stars...

-------当记忆的篇章变得零碎，当追忆的图片变得模糊，我们只能求助于数字存储的永恒的回忆

作者:黄教授

手机视频列表

塞翁失马焉知非福杨立昆离开Meta与具身智能的广阔前景

视频

音频

原始脚本

塞翁失马，杨立昆离朝 Meta 与具身智能的破局机遇。
AGI 终极赛道的博弈从未停歇，当谷歌、OpenAI 扎堆冲刺大语言模型顶峰，Meta 放走图灵奖得主杨立昆的举动，曾被视作巨头军备竞赛中的常规取舍。
如今却随其具身智能研究落地 逐渐显露出深层意味。
于杨立昆而言，脱离巨头战略束缚，深耕空间智能，恰是锚定 AI 务实落地的明智转身。
于 Meta 而言，错失具身智能先发红利，沦为大模型红海竞赛的被动追随者，成了战略取舍中的遗憾之笔。
这场双向选择的背后，藏着 AI 产业从追顶峰到接地气的价值转向。
一，语义智能的天然桎梏，连续向量与离散语言的不可逆偏差，大语言模型的幻觉难题与落地瓶颈。
根源藏在语义表达的底层矛盾中，这一矛盾近乎是技术逻辑与人类语言特性的天然对抗，短期内难寻根治之法。
从技术原理来看，大模型对语义的理解与生成，核心依托高维连续语义向量空间。
所有语言概念都会被转化为空间中的向量坐标，相似语义，如高兴、愉快、快乐，对应相邻坐标。
甚至存在大量介于多个概念之间的中间态向量。
比如某一语义计算结果可能43%贴近高兴，47%偏向愉快，剩余10%融合少量轻松的语义属性。
这种连续性恰恰契合人类语义认知的模糊边界，却与人类语言的表达形式形成根本冲突。
人类依赖离散的文字符号体系传递语义，字典中的每一个词都是孤立的离散节点，不存在覆盖所有中间态语义的精准符号。
而大模型生成文本时，必须通过 Softmax 函数计算概率，从离散的词库中选取概率最高的 token 输出。
这一步连续向量离散符号的转化，本质是一次不可逆的语义妥协。
中间态语义的精准内涵会被强行归拢到某一个具象词汇上，天然产生微小偏差。
而大语言模型的自回归生成机制，会让这一偏差持续放大，就像走路时第一步轻微偏航。
后续每一步都以偏航后的位置为起点，偏差会随文本长度累积，最终脱离原始语义，形成幻觉或逻辑断层。
更关键的是，这种偏差无根治可能。
人类语言的离散性是长期社交约定的结果，无法为适配模型而重构符号体系。
即便多模态模型尝试用图像、语音辅助语义表达，最终仍需落地到离散语言才能被人类理解，中间太语义的精准传递。
始终存在断层。
这种天然桎梏让语义智能的落地不仅依赖模型迭代，更受限于语言本身的特性。
验证纠错需复杂逻辑校验，落地周期长、风险高，也让 Meta 压住的大模型赛道陷入高投入、难领跑的消耗战。
二 空间智能的核心优势，连续动作与物理世界的无偏差式，配与语义智能的困境相反。
杨立昆聚焦的具身智能与空间智能完美契合物理世界的运行逻辑，从根源上规避了转化偏差，成为低门槛、高可靠的落地赛道。
物理空间的本质是连续的，机器人手指的转动角度、肢体的移动轨迹、物体的空间位置都能通过精准参数，如角度 距离、速度、量化。
而大模型赋能的空间智能，核心是连续语义向量、连续动作参数的直接转化，星星无需经过离散符号的妥协环节。
模型从视频中提取的空间动作逻辑，会直接转化为机器人关节的转动角度，机械臂的伸缩距 距离等连续参数，输出结果与物理空间的动作需求完全匹配，不存在语义智能那样的转化偏差。
这种适配性带来两大关键优势，一是验证直观高效，动作精准度无需复杂逻辑校验，人类肉眼即可判定机器人是否精准抓取物体，是否复刻标准运动姿势，视觉观察就能完成验证，偏差能及时定位，及时纠偏。
二是迭代可靠可控，无自回归偏差放大问题。
每一次动作参数的优化都能直接作用于结果，训练效率远超语义模型。
就像生物进化中的低等生物，蜻蜓无需复杂认知，仅靠基础视觉感知与肌肉控制就能完成精准飞行捕食。
郎平本能实现高效追猎，这些动作的核心是空间位置与肢体动作的精准匹配，无需高阶语义思考，却能稳定兑现实用价值。
而具身智能的动作生成逻辑正是对这种生物本能的技术复刻，聚焦精准执行而非深度思考，落地门槛大幅降低。
三、进化视角与商业逻辑。
动作智能是落地优先的务实选择。
从生物进化规律来看，动作及空间智能本就是智能进化的基础底盘。
地球生物的智能进化始于对物理空间的感知与动作控制。
单细胞生物的趋光运动、昆虫的飞行捕食、哺乳动物的肢体协作，都是空间智能的早期形态。
高阶认知智能，如逻辑推理、语言表达，是后期进化的叠加项，而非必选项。
这一规律 映射到 AI 发展中，意味着空间智能无需等待 AGI 的高阶认知突破，仅凭精准动作执行能力就能对接千行百业的刚需场景，其商业落地的可行性远高于语义智能。
杨立昆离开 Meta 后的研究转向，本质是对这一逻辑的精准把握。
 AGI 固然是 AI 的终极目标，但通往顶峰的路径从非唯一。
而空间智能是当下最易走通、最快盈利的康庄大道。
工业场景中，机器人可通过生产视频学习精密装配动作，替代重复高危人力。
体育康复领域，标准动作视频能直接转化为陪练机器人的训练逻辑，成本远低于人类教练。
军事场景中，机器狗格斗机器人无需复杂战术分析，仅凭精准动作执行就能完成侦查、作战任务，落地即能创造收益。
这些场景无需高阶智能，却能依托空间智能的低偏差、高可靠特性。
快速形成技术研发、商业落地、资金回流、再研发的良性循环，为 AI 产业提供持续生命力。
四、Meta 的战略困局与具身智能的错位机遇。
扎克伯格放走杨立昆，本质是 Meta 被主业绑定的战略取舍困局，而非主动放弃具身智能的长期价值。
社交媒体是 Meta 的根基，当下 AI 与社交的绑定核心落在大语言模型、智能客服、内容生成、社交互动优化，每一项都依赖语义智能支撑，这是 Meta 无法割舍的基本盘。
反观具身智能，虽商业潜力可期，但与社交主业直接关联度低，短期难反哺核心业务。
且需单独搭建机器人研发物理场景落地的完整链路，投入大、回报周期长。
在谷歌、OpenAI 全力冲刺大模型第一梯队的压力下，Meta 资源有限，只能优先押注与主业强绑定的语音赛道。
舍弃远水解不了近渴的具身智能，即便看清其价值，也难兼顾双线作战，最终陷入大模型红海的同质化竞争，错失抢占新赛道的核心筹码。
但 Meta 的遗憾恰恰成了行业的机遇。
具身智能的落地逻辑本就更适配当下产业的刚需，尤其契合制造业大国的发展环境。
大模型替代的多是办公都是高端脑力岗位，受众窄，落地场景有限。
而具身智能瞄准的是工业生产、服务行业、民生 场景的基础动作需求，替代的是海量重复性、低门槛人力岗位。
看似是低端智能，却能覆盖千行百业。
工厂装配机器人降本提效，家庭服务机器人便利生活，体育康复设备精准辅助，战场无人装备保障安全。
每一项都能快速落地变现，走薄利多销的规模化路线。
既符合产业升级需求，又能快速兑现商业价值，成为 AI 泡沫退潮后最能扎根现实的核心赛道。
杨立昆的离潮绝非失意退场，反而跳出巨头战略束缚，成了具身智能赛道的领军者。
启用语数机器人完成的动作复刻研究，正是对这一赛道价值的有力印证。
而 Meta 则困在大模型赛道内耗，错失了 AI 产业回归务实价值的关键机遇。
这场塞翁失马的转折，本质是 AI 发展逻辑的理性回归， AGI 顶峰值得追逐。
但能解决现实问题，持续创造价值的智能，才是当下产业最迫切的需求。
具身智能正以低难度、快落地、广场景的优势，成为 AI 走出泡沫、迈向规模化价值时代的关键破局点。

修正脚本

塞翁失马，杨立昆离巢 Meta 与具身智能的破局机遇。
AGI 终极赛道的博弈从未停歇，当谷歌、OpenAI 扎堆冲刺大语言模型顶峰，Meta 放走图灵奖得主杨立昆的举动，曾被视作巨头军备竞赛中的常规取舍。
如今却随其具身智能研究落地，逐渐显露出深层意味。
于杨立昆而言，脱离巨头战略束缚，深耕空间智能，恰是锚定 AI 务实落地的明智转身。
于 Meta 而言，错失具身智能先发红利，沦为大模型红海竞赛的被动追随者，成了战略取舍中的遗憾之笔。
这场双向选择的背后，藏着 AI 产业从追顶峰到接地气的价值转向。
一、语义智能的天然桎梏，连续向量与离散语言的不可逆偏差，大语言模型的幻觉难题与落地瓶颈。
根源藏在语义表达的底层矛盾中，这一矛盾近乎是技术逻辑与人类语言特性的天然对抗，短期内难寻根治之法。
从技术原理来看，大模型对语义的理解与生成，核心依托高维连续语义向量空间。
所有语言概念都会被转化为空间中的向量坐标，相似语义，如高兴、愉快、快乐，对应相邻坐标。
甚至存在大量介于多个概念之间的中间态向量。
比如某一语义计算结果可能43%贴近高兴，47%偏向愉快，剩余10%融合少量轻松的语义属性。
这种连续性恰恰契合人类语义认知的模糊边界，却与人类语言的表达形式形成根本冲突。
人类依赖离散的文字符号体系传递语义，字典中的每一个词都是孤立的离散节点，不存在覆盖所有中间态语义的精准符号。
而大模型生成文本时，必须通过 Softmax 函数计算概率，从离散的词库中选取概率最高的 token 输出。
这一步连续向量离散符号的转化，本质是一次不可逆的语义妥协。
中间态语义的精准内涵会被强行归拢到某一个具象词汇上，天然产生微小偏差。
而大语言模型的自回归生成机制，会让这一偏差持续放大，就像走路时第一步轻微偏航。
后续每一步都以偏航后的位置为起点，偏差会随文本长度累积，最终脱离原始语义，形成幻觉或逻辑断层。
更关键的是，这种偏差无根治可能。
人类语言的离散性是长期社交约定的结果，无法为适配模型而重构符号体系。
即便多模态模型尝试用图像、语音辅助语义表达，最终仍需落地到离散语言才能被人类理解，中间态语义的精准传递始终存在断层。
这种天然桎梏让语义智能的落地不仅依赖模型迭代，更受限于语言本身的特性。
验证纠错需复杂逻辑校验，落地周期长、风险高，也让 Meta 押注的大模型赛道陷入高投入、难领跑的消耗战。
二、空间智能的核心优势，连续动作与物理世界的无偏差适配，与语义智能的困境相反。
杨立昆聚焦的具身智能与空间智能完美契合物理世界的运行逻辑，从根源上规避了转化偏差，成为低门槛、高可靠的落地赛道。
物理空间的本质是连续的，机器人手指的转动角度、肢体的移动轨迹、物体的空间位置都能通过精准参数，如角度、距离、速度量化。
而大模型赋能的空间智能，核心是连续语义向量、连续动作参数的直接转化，本身无需经过离散符号的妥协环节。
模型从视频中提取的空间动作逻辑，会直接转化为机器人关节的转动角度，机械臂的伸缩距离等连续参数，输出结果与物理空间的动作需求完全匹配，不存在语义智能那样的转化偏差。
这种适配性带来两大关键优势，一是验证直观高效，动作精准度无需复杂逻辑校验，人类肉眼即可判定机器人是否精准抓取物体，是否复刻标准运动姿势，视觉观察就能完成验证，偏差能及时定位，及时纠偏。
二是迭代可靠可控，无自回归偏差放大问题。
每一次动作参数的优化都能直接作用于结果，训练效率远超语义模型。
就像生物进化中的低等生物，蜻蜓无需复杂认知，仅靠基础视觉感知与肌肉控制就能完成精准飞行捕食。
蜻蜓本能实现高效追猎，这些动作的核心是空间位置与肢体动作的精准匹配，无需高阶语义思考，却能稳定兑现实用价值。
而具身智能的动作生成逻辑正是对这种生物本能的技术复刻，聚焦精准执行而非深度思考，落地门槛大幅降低。
三、进化视角与商业逻辑。
动作智能是落地优先的务实选择。
从生物进化规律来看，动作及空间智能本就是智能进化的基础底盘。
地球生物的智能进化始于对物理空间的感知与动作控制。
单细胞生物的趋光运动、昆虫的飞行捕食、哺乳动物的肢体协作，都是空间智能的早期形态。
高阶认知智能，如逻辑推理、语言表达，是后期进化的叠加项，而非必选项。
这一规律 映射到 AI 发展中，意味着空间智能无需等待 AGI 的高阶认知突破，仅凭精准动作执行能力就能对接千行百业的刚需场景，其商业落地的可行性远高于语义智能。
杨立昆离开 Meta 后的研究转向，本质是对这一逻辑的精准把握。
 AGI 固然是 AI 的终极目标，但通往顶峰的路径并非唯一。
而空间智能是当下最易走通、最快盈利的康庄大道。
工业场景中，机器人可通过生产视频学习精密装配动作，替代重复高危人力。
体育康复领域，标准动作视频能直接转化为陪练机器人的训练逻辑，成本远低于人类教练。
军事场景中，机器狗格斗机器人无需复杂战术分析，仅凭精准动作执行就能完成侦查、作战任务，落地即能创造收益。
这些场景无需高阶智能，却能依托空间智能的低偏差、高可靠特性，快速形成技术研发、商业落地、资金回流、再研发的良性循环，为 AI 产业提供持续生命力。
四、Meta 的战略困局与具身智能的错位机遇。
扎克伯格放走杨立昆，本质是 Meta 被主业绑定的战略取舍困局，而非主动放弃具身智能的长期价值。
社交媒体是 Meta 的根基，当下 AI 与社交的绑定核心落在大语言模型、智能客服、内容生成、社交互动优化，每一项都依赖语义智能支撑，这是 Meta 无法割舍的基本盘。
反观具身智能，虽商业潜力可期，但与社交主业直接关联度低，短期难反哺核心业务。
且需单独搭建机器人研发物理场景落地的完整链路，投入大、回报周期长。
在谷歌、OpenAI 全力冲刺大模型第一梯队的压力下，Meta 资源有限，只能优先押注与主业强绑定的语义赛道。
舍弃远水解不了近渴的具身智能，即便看清其价值，也难兼顾双线作战，最终陷入大模型红海的同质化竞争，错失抢占新赛道的核心筹码。
但 Meta 的遗憾恰恰成了行业的机遇。
具身智能的落地逻辑本就更适配当下产业的刚需，尤其契合制造业大国的发展环境。
大模型替代的多是办公等高端脑力岗位，受众窄，落地场景有限。
而具身智能瞄准的是工业生产、服务行业、民生场景的基础动作需求，替代的是海量重复性、低门槛人力岗位。
看似是低端智能，却能覆盖千行百业。
工厂装配机器人降本提效，家庭服务机器人便利生活，体育康复设备精准辅助，战场无人装备保障安全。
每一项都能快速落地变现，走薄利多销的规模化路线。
既符合产业升级需求，又能快速兑现商业价值，成为 AI 泡沫退潮后最能扎根现实的核心赛道。
杨立昆的离巢绝非失意退场，反而跳出巨头战略束缚，成了具身智能赛道的领军者。
启用具身机器人完成的动作复刻研究，正是对这一赛道价值的有力印证。
而 Meta 则困在大模型赛道内耗，错失了 AI 产业回归务实价值的关键机遇。
这场塞翁失马的转折，本质是 AI 发展逻辑的理性回归， AGI 顶峰值得追逐。
但能解决现实问题，持续创造价值的智能，才是当下产业最迫切的需求。
具身智能正以低难度、快落地、广场景的优势，成为 AI 走出泡沫、迈向规模化价值时代的关键破局点。

英文翻译

When the old man at the frontier lost his horse, Yann LeCun left Meta, presenting a breakthrough opportunity for embodied intelligence.

The game on the ultimate track of AGI has never ceased. While Google and OpenAI were pushing the limits of large language models, Meta's decision to let go of Turing Award winner Yann LeCun was once seen as a routine trade-off in the arms race among tech giants. Now, as his embodied intelligence research comes to fruition, the deeper implications are gradually emerging.

For Yann LeCun, breaking free from the strategic constraints of a tech giant to focus on spatial intelligence is precisely a wise shift toward anchoring AI in practical applications. For Meta, missing the first-mover advantage in embodied intelligence and becoming a passive follower in the red ocean of large model competition is a regrettable consequence of strategic trade-offs.

Behind this two-way choice lies a shift in the AI industry’s value orientation—from chasing the pinnacle to grounding itself in reality.

**I. The Inherent Constraints of Semantic Intelligence**  
The irreversible deviation between continuous vectors and discrete language, the hallucination problem of large language models, and the bottleneck in real-world deployment all stem from a fundamental contradiction at the level of semantic expression. This contradiction is almost a natural conflict between technical logic and the characteristics of human language, with no easy short-term solution.

From a technical perspective, large models rely on a high-dimensional continuous semantic vector space for understanding and generating language. All linguistic concepts are transformed into vector coordinates in this space, where similar semantics—such as "happy," "joyful," and "delighted"—correspond to adjacent coordinates. There are even numerous intermediate-state vectors that lie between multiple concepts. For example, a certain semantic calculation might be 43% close to "happy," 47% inclined toward "joyful," and the remaining 10% blended with a slight attribute of "lighthearted." This continuity aligns with the fuzzy boundaries of human semantic cognition but fundamentally conflicts with the form of human language expression.

Human beings rely on a discrete system of written symbols to convey meaning. Every word in the dictionary is an isolated, discrete node, with no precise symbol covering all intermediate semantic states. When a large model generates text, it must compute probabilities through a Softmax function, selecting the token with the highest probability from a discrete vocabulary. This conversion from continuous vectors to discrete symbols is essentially an irreversible semantic compromise. The precise meaning of intermediate states is forcibly consolidated into a specific word, inherently introducing small deviations.

The autoregressive generation mechanism of large language models then amplifies this deviation, much like taking a first step slightly off course. Each subsequent step starts from the deviated position, causing the error to accumulate as the text lengthens, eventually drifting away from the original semantics, resulting in hallucinations or logical gaps. More critically, this deviation is fundamentally incurable. The discreteness of human language is a result of long-term social conventions and cannot be restructured to fit the model. Even multimodal models attempt to aid semantic expression with images or speech, but they ultimately need to fall back on discrete language for human comprehension, leaving a persistent gap in precise transmission of intermediate semantics.

This inherent constraint means that the deployment of semantic intelligence depends not only on model iteration but is also limited by the very characteristics of language. Verification and error correction require complex logical checks, leading to long deployment cycles and high risks. Consequently, the large model track that Meta bet on has become a costly, hard-to-win war of attrition.

**II. The Core Advantage of Spatial Intelligence**  
Continuous actions perfectly align with the physical world without deviation, contrasting sharply with the plight of semantic intelligence. The embodied intelligence and spatial intelligence that Yann LeCun focuses on are perfectly suited to the operational logic of the physical world, fundamentally avoiding conversion errors and becoming a low-threshold, high-reliability track for real-world application.

The physical space is inherently continuous: the rotation angle of a robot's finger, the trajectory of a limb's movement, and the spatial position of an object can all be quantified with precise parameters such as angle, distance, and speed. The core of spatial intelligence empowered by large models lies in the direct conversion of continuous semantic vectors into continuous action parameters, without needing to compromise through discrete symbols. The spatial action logic extracted from videos is directly transformed into continuous parameters like joint rotation angles or robotic arm extension distances. The output perfectly matches the action requirements of physical space, with none of the conversion errors seen in semantic intelligence.

This compatibility brings two key advantages. First, verification is intuitive and efficient: action accuracy does not require complex logical checks. The human eye can directly judge whether a robot is precisely grasping an object or replicating a standard movement posture. Visual observation suffices for verification, allowing timely identification and correction of errors. Second, iteration is reliable and controllable: there is no problem of error amplification through autoregression. Each optimization of action parameters directly impacts the outcome, making training efficiency far superior to that of semantic models.

This is like lower organisms in biological evolution. A dragonfly, without complex cognition, can achieve precise flight and predation merely through basic visual perception and muscle control. Its instinct enables efficient hunting. The core of these actions is the precise matching of spatial positions with limb movements, requiring no high-level semantic reasoning yet delivering consistent practical value. The action generation logic of embodied intelligence is a technological replication of such biological instincts, focusing on precise execution rather than deep thought, greatly lowering the barrier to real-world deployment.

**III. Evolutionary Perspective and Business Logic**  
Action intelligence is a pragmatic choice prioritizing deployment. From the laws of biological evolution, action and spatial intelligence are the foundational chassis of intelligent evolution. The intelligent evolution of Earth's organisms began with the perception of physical space and control of actions. The phototactic movement of single-celled organisms, the flight and predation of insects, and the limb coordination of mammals are all early forms of spatial intelligence. Higher-order cognitive intelligence, such as logical reasoning and language expression, is a superimposed later development, not a necessity.

This principle, when mapped to AI development, means that spatial intelligence does not need to wait for the breakthrough of high-level AGI cognition. Simply by precise action execution, it can address the essential needs of countless industries. Its commercial feasibility far exceeds that of semantic intelligence. Yann LeCun's pivot after leaving Meta essentially reflects a precise grasp of this logic. AGI is indeed the ultimate goal of AI, but the path to the summit is not singular. Spatial intelligence is the most accessible and fastest way to profitability right now.

In industrial settings, robots can learn precision assembly movements from production videos, replacing repetitive and hazardous human labor. In sports rehabilitation, standard movement videos can be directly transformed into the training logic of companion robots at a cost far lower than human coaches. In military scenarios, robotic dogs or fighting robots can complete reconnaissance and combat missions with precise action execution alone, without complex tactical analysis. These scenarios do not require high-level intelligence but can leverage the low error and high reliability of spatial intelligence to quickly form a virtuous cycle of R&D, commercial deployment, capital return, and further R&D, providing sustainable vitality for the AI industry.

**IV. Meta's Strategic Dilemma and the Misaligned Opportunity of Embodied Intelligence**  
Zuckerberg's decision to let Yann LeCun go essentially reflects Meta's strategic trade-off dilemma bound by its core business, not an active abandonment of the long-term value of embodied intelligence. Social media is Meta's foundation. The current integration of AI and social interaction centers on large language models, intelligent customer service, content generation, and social interaction optimization—each dependent on semantic intelligence. This is the base Meta cannot afford to sacrifice.

In contrast, although embodied intelligence has promising commercial potential, its direct relevance to the social media core is low, and it would be difficult to feed back into the core business in the short term. Moreover, it requires building a complete chain from robot R&D to physical-world deployment, with high investment and long return cycles. Under pressure from Google and OpenAI sprinting to the top tier of large models, Meta's limited resources forced it to prioritize the semantically driven track strongly tied to its core business. Abandoning embodied intelligence—which could not solve immediate needs—even while recognizing its value, Meta found itself unable to fight on two fronts. Ultimately, it fell into homogeneous competition in the red ocean of large models, missing the key opportunity to seize a new track.

But Meta's loss is precisely the industry's opportunity. The deployment logic of embodied intelligence is inherently better suited to the immediate needs of current industries, especially in manufacturing-driven economies. Large models mostly replace high-end brain-work positions like office jobs, with a narrow audience and limited application scenarios. In contrast, embodied intelligence targets the basic action needs of industrial production, service industries, and everyday life, replacing massive, repetitive, low-skill labor. It may seem like low-end intelligence, but it can cover countless industries. Factory assembly robots reduce costs and improve efficiency; home service robots make life more convenient; sports rehabilitation devices provide precise assistance; battlefield unmanned equipment ensures safety. Each of these can be quickly deployed and monetized, following a path of small margins and large volumes.

This aligns with industrial upgrade needs and can rapidly realize commercial value, making embodied intelligence the core track most capable of taking root in reality as the AI bubble recedes. Yann LeCun's departure is by no means a retreat of frustration. Instead, by escaping the strategic constraints of a giant, he has become a leader on the embodied intelligence track. The action replication research using embodied robots is a strong testament to the value of this track. Meanwhile, Meta remains mired in internal friction on the large model track, missing the key opportunity for the AI industry to return to pragmatic value.

This twist of "the old man losing his horse" essentially reflects a rational return to the development logic of AI. The peak of AGI is worth pursuing, but intelligence that can solve real problems and continuously create value is what the industry urgently needs now. Embodied intelligence, with its low difficulty, fast deployment, and broad scenarios, is becoming the key breakthrough point for AI to emerge from the bubble and enter an era of scaled value.