Core Insights - The CEO of XPeng Motors, He Xiaopeng, shared insights on the AI industry after discussions with nearly thirty AI professionals in the U.S. [1][7] Group 1: U.S. Entrepreneurship and Robotics - The U.S. has a high concentration of startups in AI, biotechnology, and finance, particularly in SaaS and physical AI robotics, with notably high valuations [3][8] - U.S. robotics companies often focus on models, while Chinese companies tend to start from joints and control systems [3][8] - Human-like robots are expected to become a competitive field among major players, while specialized robots will see numerous successful entrants across various sectors [3][8] Group 2: Physical AI - He Xiaopeng emphasized that breakthroughs in physical AI, such as autonomous driving reaching Level 4 capabilities and human-like robots advancing from Level 1 to Level 4, are anticipated within the next three years [4][9] - The development of physical AI may be slower than digital AI but is expected to have a more significant impact on daily life [4][9] Group 3: AI Market Dynamics - He Xiaopeng noted that every technological era experiences phases of bubbles, which is a natural competitive process leading to market order [4][9] - Currently, there is no AI bubble; the market is still in its early stages, with significant opportunities ahead [4][9] Group 4: Arrival of AGI - Today's AI primarily relies on imitation learning and reinforcement learning, which allows for rapid improvement in tasks like driving [5][11] - True AGI requires advancements in capabilities such as multi-modal learning and long-term planning, which are still years away from realization [6][11]

2025年下半年，AI泡沫的讨论从边缘话题演变为市场焦点。美股前10大市值的公司，已有8家是AI相关 科技公司，AI集中度远高于互联网泡沫时期。 资本流动更是呈现极端化特征。一方面，科技巨头开启"军备竞赛"，另一方面，机构投资者纷纷获利了 结。一时间，"泡沫论"甚嚣尘上。 他还以自动驾驶举例，因为可以快速从百万人的驾驶习惯中学习，所以会看到很快可以比大部分人开车 更好更安全，以及出现全新的涌现能力，但是这些并不是真正的创造力。 在何小鹏看来，"真正的AGI还需要很多能力，例如从多模态到世界模型（类似我们的知其所以然，或 者类似自动驾驶世界模型的左右互搏学习的升级版），例如持续学习，例如长时序洞见和规划能力等都 还没有完全到来，可能还需要数年时间以及等待部分底层能力的再次提高。" 何小鹏对于AI泡沫等方面的看法，与当前行业的不少分析类似。 瑞银证券报告显示，中国互联网龙头资本支出占收入比例（10%）显著低于美国同业的27%。在应用 端，垂直领域存在较高的服务能力壁垒，头部公司主动将AI功能整合至自有应用中。 12月17日，小鹏汽车董事长、CEO何小鹏在朋友圈分享了自己在美国出差后的一些感受，谈到对机器 人、AI ...

Core Viewpoint - Xiaomi announced the launch of the Xiaomi MiMo-V2-Flash model at the "Human-Vehicle-Home Ecosystem" partner conference, highlighting advancements in AI technology and the importance of open-source models in achieving AGI [1][11]. Group 1: Xiaomi MiMo-V2-Flash Model - The Xiaomi MiMo-V2-Flash is a compact model with a total of 309 billion parameters and 15 billion active parameters, ranking it among the top two globally in terms of code capability [4]. - The model outperforms DeepSeek V3.2 and K2-Thinking in most benchmark tests, achieving lower costs and higher speeds while reducing parameter counts by 50% to 67% compared to similar top models [6]. - The model is designed to simulate the real world and aims to transition from merely answering questions to completing tasks, incorporating memory, reasoning, planning, decision-making, and execution capabilities [7]. Group 2: Role of Luo Fuli - Luo Fuli, a prominent figure in AI research, emphasized the significance of open-source models in democratizing AGI and ensuring collective human intelligence evolution [7]. - Her background includes a strong academic record with multiple publications in top AI conferences and experience at Alibaba and DeepSeek, showcasing her expertise in AI model development [10]. - Luo's recent transition to Xiaomi was marked by her commitment to building a future where AI integrates with the physical world, reflecting the company's strategic focus on AI applications [11][14]. Group 3: Xiaomi's AI Strategy - Xiaomi is heavily investing in AI, with a focus on deep integration of AI with the physical world, as indicated by the development of various models in the MiMo series [14]. - The company aims to provide a strong platform for top talent, as highlighted by the recruitment of Luo Fuli and the ongoing advancements in AI technology [11][14].

官方体验页面信息显示，MiMo-V2-Flash还支持深度思考和联网搜索功能，既能对话聊天，也能在需要实时数据、最新动态或资料核对的场景里派上用 场。 上证报记者刘暄摄 发布会上，罗福莉谈及下一代智能体系统时认为，下一代智能体系统不是一个"语言模拟器"，而是一个真正能够理解世界、并与之共存的"智能体"。 "AI进化的下一个起点一定要有可以跟真实环境产生交互的物理模型。我们本质上要打造的并不是一个程序，而是一个具备物理一致性、时空连贯性的虚 拟宇宙。这本质上代表着AI的能力有本质的跨越，它不仅仅满足于看画面，而是要理解背后的物理规律；不仅仅局限于处理文本，而是要推演整个世界 的运作逻辑。"她说。 Xiaomi MiMo大模型负责人罗福莉12月17日在2025小米人车家全生态合作伙伴大会上迎来入职后首秀。她在发布会上表示，MiMo-V2-Flash围绕极致推理 效率设计了模型结构，通过3层MTP推理加速并行Token验证，实现了2.0至2.6倍的推理速度提升。 上证报记者刘暄摄 据了解，小米12月16日晚间发布了Xiaomi MiMo-V2-Flash开源MoE模型，该模型总参数量达3090亿，活跃参数为150亿 ...

Core Viewpoint - The chairman of XPeng Motors, He Xiaopeng, believes that there is currently no AI bubble and that the future AI market holds immense opportunities [5][7]. Group 1: AI and Robotics Landscape - The U.S. is experiencing a surge in entrepreneurship in AI, biotechnology, and finance, with a significant focus on SaaS and physical AI robotics, leading to high valuations [2][9]. - In the robotics sector, U.S. companies tend to focus on models, while Chinese companies often start with joints and control systems [9]. - The competition in humanoid robots is expected to be dominated by major players, while specialized robots will see a multitude of competitors across various fields, creating numerous success opportunities [2][9]. Group 2: Physical AI Developments - The development of physical AI is anticipated to yield significant breakthroughs in the next three years, particularly in areas like autonomous driving and humanoid robots, which may rapidly advance from Level 1 to Level 4 capabilities [3][10]. - The pace of physical AI development may be slower than digital AI, but its impact on daily life is expected to be much greater [10]. Group 3: AI Bubble Discussion - Any technological era may experience localized bubbles, but this is part of the natural competitive process from chaos to order [4][10]. - The current stage of AI is viewed as being at the very beginning, with a potential for substantial societal transformation [4][10]. - Valuations in China are perceived as more reasonable compared to the higher valuations in the U.S., which tends to focus more on cutting-edge research rather than market applications [4][10]. Group 4: AGI Development - Current AI capabilities are primarily based on imitation learning and reinforcement learning, lacking true creativity [6][12]. - The realization of Artificial General Intelligence (AGI) will require advancements in multi-modal learning, world modeling, continuous learning, and long-term planning abilities, which are still several years away [6][12].

12月17日，小鹏汽车董事长何小鹏在朋友圈发文谈到了机器人和AI泡沫的话题。他指出，人形机器人 将来会是巨头的竞争，而不同的专用机器人则会有大量不同领域的选手，且会有非常多的成功机会。他 认为，当前没有AI泡沫，未来AI的市场有着巨大机遇。 以下为全文： 关于美国的新创业和机器人 感觉美国在AI，生物，金融方面的创业多。而AI方面的创业，从硅谷看，在SaaS和物理AI机器人领域 特别多，估值也特别高，我这次聊了其中一半都是做机器人创业的朋友，非常集中。中国的机器人公司 很多是从关节和控制入手，而美国的机器人公司很多是从模型入手。我认为人形机器人将来会是巨头的 竞争，而不同的专用机器人则会有大量不同领域的选手，且会有非常多的成功机会。 关于物理AI 维特根斯坦曾经说过，"语言即世界"。语言是知识的总结，信息的传递，但某种程度也束缚了对世界的 想象。很多事情靠语言很难想象和学习，例如小朋友如何学习行走，如何学习游泳，主要都必须依靠对 世界的观察，模仿和强化。所以，大家都在期望从语言大模型，走向多模态甚至世界模型。我认为未来 3年，最有可能有大突破的不一定是在数字世界（可以很明显看到OpenAI虽然核心在说AGI， ...

罗福莉加入小米后的首次公开演讲，来了！ 智东西12月17日北京现场报道，刚刚，小米举办了年度"人车家全生态"合作伙伴大会，现场人头攒动座无虚席，展区不少展台都已被围观人群 挤满，讨论热度颇高。 ·相比生物演化的稳固根基，AI的发展是"空中楼阁"； ·AI进化的下个起点，一定是有一个能跟物理世界交互的模型； 小米创始人兼CEO雷军此次并未来到现场，小米集团合伙人、集团总裁卢伟冰率先登台演讲，今天第三位演讲的是当前备受关注的原 DeepSeek核心成员、被业内称为"天才少女"的罗福莉，她现在的职位是小米MiMo大模型负责人。 罗福莉在演讲过程中虽然稍显紧张，但她抛出的海量金句却句句令人印象深刻： ·AI正以非线性的方式重演人类大脑6亿年的进化史； ·语言是人类思维和物理世界在符号空间的"投影"； ·下一代智能体系统不是一个"语言模拟器"，而是一个真正理解我们世界、并与之共存的"智能体"； ·算力和数据也并非最终的护城河，真正的护城河，是科学的研究文化与方法，是将未知问题结合模型优化转化为可用产品的能力； ·开源的价值本质上是一种分布式的技术加速主义； ·开源是实现AGI的普惠化，是确保所有人类的智慧共同进化的唯一 ...

从市场影响来看，MiMo-V2-Flash的入局可能搅动现有开源AI模型的竞争格局。其官方公布的每百万输入token 0.1美元、输出token 0.3美元的极低 成本，结合高达150 tokens/秒的推理速度，为开发者和企业提供了极具吸引力的选择，或将加速高性能AI技术在更广泛场景的应用和普及，尤其 是对其庞大的"手机 x AIoT"生态系统形成强大赋能。 11个小时前，小米深夜"突袭式"发布并开源了其最新的专家混合架构（MoE）大语言模型MiMo-V2-Flash。该模型总参数量达3090亿，活跃参数为 150亿，采用对开发者友好的MIT开源协议，基础版权重也已经在Hugging Face上发布。 小米MiMo团队负责人罗福莉（Fuli Luo）在社交平台上明确表示："MiMo-V2-Flash已经上线。这只是我们AGI路线图上的第二步。"这一表态凸 显了小米在AI领域的长远规划和技术雄心。 性能媲美DeepSeek-V3.2，并且"极具性价比" MiMo-V2-Flash在多个权威基准测试中展现了强大的实力，其性能表现足以和部分顶尖的开源及闭源模型同台竞技。 根据小米官方公布的数据，在衡量编程能力的S ...

Group 1 - The core viewpoint is that overseas cloud vendors maintain an optimistic outlook on capital expenditures in the AI sector, with Oracle's capital expenditure significantly exceeding Bloomberg's consensus expectations and a substantial upward revision for FY2026 [3][8] - Oracle reported a capital expenditure of $12.03 billion, far surpassing the expected $8.25 billion, and has raised its FY2026 capital expenditure forecast to $50 billion [3][8] - The communication ETF (515880) has performed exceptionally well in the A-share market, with a year-to-date increase of over 110%, making it the top-performing ETF in the market [1][8] Group 2 - The communication ETF (515880) is heavily involved in the overseas computing power supply chain, with over 54% of its components being optical modules and more than 19% being servers, indicating a strong position in the market [8][12] - The AI industry is expected to see infrastructure spending reach $3 trillion to $4 trillion by 2030, with a projected compound annual growth rate (CAGR) of around 40% over the next five years [9][11] - The growth of AI is being accelerated by advancements from companies like Google, which are driving increased demand for computing chips and related technologies, thereby enhancing the overall industry growth [11][12]

作为一个从小就非常敏感，容易大喜大悲的人，林宋琪在成长中逐渐意识到这种对情绪的敏感其实是一种可以称之为“锐感力”的宝贵天赋，在AI时代反而拥有特别的意义。 林宋琪是⼀位连续创业者、AI创业先锋，Tanka AI CEO，曾任TikTok全球商业化战略总监和Facebook集团战略经理。她北京⼤学+⾹港⼤学⾦融经济学本科毕业，却裸辞摩根⼤通，选择了互联⽹和科技领域，在Facebook从销售做到集团战略经理，后加⼊字节任战略总监，之后开始AI领域创业。4年AI圈创业经验 + 8年互联⽹经验，⾹港/⼤陆/新加坡/硅⾕多地⼯作经历，喜欢不停挑战，不停蜕变。凭借对AGI浪潮的敏锐直觉和丰富的跨国实战经验，她持续探索着在⼈⼯智能时代，技术创新与⼈类价值的全新结合点。 This talk was given at a TEDx event using the TED conference format but independently organized by a local community. Learn more at https://www.ted.com/tedx ...