以下文章来源于真故研究室 ，作者龚正 真故研究室 . 真问题，更商业 作者丨 龚正 来源丨 真故研究室 （ID： zhengulab ） 编辑丨 脸叔 图源丨Midjourney 部分网友对 AI 做的视频，认知和接受度仅限于两年前风靡网络的"甄嬛吃汉堡"，作为一种片段式的鬼畜文化来猎奇娱乐。 最近，在 B 站上出现了一个长达近 7 分钟的"纯 AI 综艺"，讲全世界 6 位厨师如何把灭绝了 6500 万年的远古沧龙做成 6 道菜，收获了 700 多万点击。有人压根没看出来这是 AI 做 的，还以为是美国烹饪竞技真人秀《地狱厨房》出了续集。 这部 AI 综艺对行业内最大的启发是：加入了作者主体性的 AI 长视频内容，是否会成为各平台 UP 主和创作达人，一个升级版的内容赛道。在这之前，据说有 50% 的网友对 AI 做的内 容是"排斥的"。 科技综艺，骗过700万网友 沧龙是一种海生爬行动物，有"史前水生蜥蜴"之称，有时候凶残起来连同伴也吃进肚里。 在这部名为《把远古沧龙做成六道菜》的 AI 综艺里，肉质可期的沧龙，被来自世界 6 个国家的主厨，各自做成了一道家乡菜，参与厨艺大赛，争夺黑衫。 图 | 综艺截图 ...

Cuz videos are expensive to me. For people that are doing that like hundreds of times a day, the traditional naive thing that it's like 1% of users create content, 10% leave comments, and 100% view. Maybe a lot more want to create content, but it's just been harder to do.I think that's a very cool. ...

2025年9月23日（周二），OpenAI CEO萨姆·奥尔特曼（Sam Altman）出席在美国得克萨斯州阿比林市举办的星门人工 智能数据中心媒体参观活动。图片来源：KYLE GRILLOT / BLOOMBERG—GETTY IMAGES OpenAI正规划截至本十年末的盈利逆袭之路，但实现增长需经历阵痛期。《华尔街日报》获取的财务 文件显示，该公司预计未来数年将持续面临巨额亏损，仅2028年就将出现约740亿美元的运营亏损，随 后将在2030年前实现爆发式盈利。 今年夏季向投资者披露的文件显示，其激进增长战略依赖前期在计算基础设施、芯片及数据中心的巨额 投入。CEO萨姆·奥尔特曼称，要满足其眼中"对人工智能能力永无止境的需求"，这一支出必不可少。 该公司预计今年销售额将达到130亿美元，同期现金消耗约为90亿美元，现金消耗率约占营收的70%。 但财务状况在出现好转前将进一步恶化。文件显示，受算力成本飙升推动，OpenAI预计到2028年，其 运营亏损将增至当年营收的四分之三左右。据《华尔街日报》报道，竞争对手Anthropic则预计在同年 实现盈亏平衡。 这些数字凸显了两大最具价值人工智能初创企业之间的 ...

神译局是36氪旗下编译团队，关注科技、商业、职场、生活等领域，重点介绍国外的新技术、新观点、新风向。 编者按：左手拉拢AMD造芯，右手把应用装进ChatGPT。OpenAI要做"Windows"的野心藏不住了。文章来自编译。 OpenAI 铺天盖地的公告简直让人应接不暇。仅摘录上个月的部分内容（根本列不完！）： 最后两项公告让上述所有布局的脉络清晰了。简而言之，OpenAI 正在谋求成为"AI 领域的 Windows"。 在近二十年的时间里，智能手机，特别是 iOS，一直是人们讨论平台时的重要参考。不过，必须指出的是，尽管苹果的软硬件一体化战略利润极其丰厚，但 也不可避免地为竞争平台的出现敞开了大门。作为一家硬件公司，挑战在于你必须实实在在地制造设备，这决定了你无法满足所有人的需求；尤其是苹果， 既没有能力也无意愿走向低端市场，这就给安卓创造了机会——不仅建立了一个竞争平台，而且在市场份额上实现了对 iOS 的大幅超越。 这意味着，如果我们想为"平台的绝对主导地位"寻找一个历史类比——而这似乎正日益成为 OpenAI 的目标——我们就必须进一步追溯到 PC 时代和 Windows。 平台的建立 在 Windo ...

A new legal battle is making headlines in the tech world. The personalized video app Cameo is suing OpenAI, the maker of Chat GPT. It's accusing OpenAI of infringing on its trademark name by launching a feature in its AI video app Sora 2 that it's also calling Cameo.If you're not familiar, Cameo allows fans to commission customized videos from real celebrities and other notable people. While the cameo feature on Sora 2 allows users to generate deep fake videos of themselves, friends, and some public figures ...

Core Viewpoint - The article discusses the evolving role of AI, emphasizing that it should serve as a supportive tool for humans rather than replacing them, highlighting the importance of responsible AI development and usage [2]. Group 1: AI Development and Ethics - Tencent's Vice President Guo Kaitian asserts that as AI becomes more powerful, it challenges humanity to ensure it remains benevolent, advocating for AI to be a human's intelligent co-pilot rather than a replacement [2]. - The controversy surrounding OpenAI's Sora 2 tool, which generates content in the style of renowned animator Hayao Miyazaki, has led to backlash from Japanese IP companies, claiming copyright infringement due to unauthorized use of their content for AI training [2]. Group 2: AI Model Performance and Cost - Moonshot AI's Kimi K2 Thinking model, which achieved a 44.9% score in the Human Ultimate Exam, reportedly cost only $4.6 million to train, significantly less than the billions invested by OpenAI in its models [3]. - Baidu's Wenxin model 5.0-Preview has ranked first in China and second globally in text capabilities, outperforming OpenAI's GPT-5-High, showcasing advancements in AI technology within China [3].

Group 1 - The core viewpoint is that the AI sector is experiencing short-term adjustments, with the Sci-Tech Innovation Artificial Intelligence ETF (589010) down by 1.91% and only 2 out of 30 component stocks showing gains [1] - Microsoft has formed a super-intelligent team within its Microsoft AI division, focusing on affordable AI companions, expert-level medical diagnostics, and clean energy management [1] - The trading volume for the AI sector is approximately 23.94 million yuan, indicating moderate market turnover [1] Group 2 - OpenAI is set to launch its first Sora model in February 2024, marking a significant breakthrough in the video domain, akin to the GPT-1 moment [2] - The Sora 2 model, expected in September 2025, will enhance physical simulation, realism, and controllability, offering a unified audiovisual experience [2] - The Gemini API by Google has released a paid preview of Veo 3.1, which includes significant upgrades for richer audio support and improved narrative control [2]

Group 1 - OpenAI has surpassed 1 million global enterprise customers, marking it as the fastest-growing commercial platform in history [1] - Notable enterprise clients include Amgen, Commonwealth Bank of Australia, Booking.com, Cisco, Lowe's, Morgan Stanley, T-Mobile, Target, and Thermo Fisher Scientific, showcasing OpenAI's broad application across various industries [1] - The consumer user base supports the expansion of OpenAI services among enterprise users [1] Group 2 - ChatGPT has over 800 million weekly active users, facilitating smoother deployment and shorter pilot cycles for enterprises [2] - Paid users of ChatGPT for Work have exceeded 7 million, with a 40% growth in two months, and enterprise version users have surged ninefold year-on-year [2] - New features include company knowledge tools and integration with platforms like Slack, SharePoint, Google Drive, and GitHub, utilizing the optimized GPT-5 model [2] Group 3 - 75% of enterprises report positive return on investment from OpenAI technologies, with less than 5% reporting negative results [3] - OpenAI's CEO Sam Altman stated that technology is transitioning from being an auxiliary tool to a core partner for enterprises [3] - A recent $38 billion agreement with Amazon Web Services has been signed to ensure stable supply of computing power [3]

Core Insights - OpenAI has surpassed 1 million enterprise customers, making it the fastest-growing enterprise platform in history [1] - The rapid growth of enterprise customers is attributed to the widespread acceptance of the consumer market, with over 800 million weekly active users of ChatGPT [3] Group 1: Customer Growth - The number of ChatGPT for Work seats has increased by 40% in just two months, exceeding 7 million [3] - ChatGPT Enterprise seats have seen an annual growth rate of 9 times [3] Group 2: New Tools and Features - OpenAI has launched several new tools and integration features to support enterprise customers transitioning from trial to full deployment [4] - The company knowledge feature, based on the optimized GPT-5 version, allows for collaborative reasoning across tools like Slack, SharePoint, and Google Drive [4] - The Codex model for code generation has seen a tenfold increase in usage since August, with companies like Cisco reducing code review time by 50% [4] Group 3: AgentKit and Multi-Modal Models - The AgentKit tool has lowered the barriers for building and deploying enterprise agents, reducing the time to turn ideas into applications from months to days [4] - The multi-modal model upgrade has expanded business process applications, enabling teams to handle text, image, video, and audio data within the same system [4] Group 4: Business Integration - Companies such as Canva, Figma, Zillow, and Spotify have integrated applications directly with ChatGPT [5] - Retailers like Shopify, Etsy, Walmart, and PayPal are leveraging ChatGPT's agent-based commerce protocol to create new shopping experiences [5]

Core Insights - OpenAI's Sora 2 has shifted its core positioning from a traditional video generation tool to a "world simulator" that understands and simulates the laws of the physical world and causal relationships [1][26] - The introduction of the Cameo feature allows users to integrate themselves and friends into generated videos, creating a socially driven generative network [1][12] - OpenAI has announced that Sora 2 is now more accessible, as it no longer requires an invitation code for use, marking a significant shift in its technical approach [1][26] Technical Foundations - Sora 2 utilizes Diffusion Transformer (Dit) technology, which enables the model to reconstruct a complete video from noise rather than generating it frame by frame [3][4] - The model employs "space-time patches" to simultaneously handle spatial and temporal information, allowing for a more coherent understanding of video continuity [4][5] - This approach results in improved object permanence and logical action sequences, as the model can maintain consistency in character appearance and behavior throughout the video [5][6] Emergence of Intelligence - Sora 2's model begins to exhibit agent-like characteristics, where it not only generates actions but also evaluates their logical validity based on physical principles [7][8] - The model's ability to simulate realistic outcomes, such as a basketball bouncing back if a shot misses, signifies a shift from mere visual generation to causal simulation [9][11] - The emergence of these intelligent traits occurs naturally as the model scales, similar to advancements seen in the GPT series [10][11] Product Dynamics - The Cameo feature has transformed user interaction, encouraging participation in video creation rather than passive consumption [12][16] - User engagement metrics indicate that once users gain access, they are highly active, with a significant percentage returning to create more content [15][16] - This social aspect of content creation fosters a community-driven environment, where users are motivated to include friends in their generated videos [16][20] Future Vision - OpenAI envisions Sora as a micro-reality platform rather than just a video generation tool, aiming to create a parallel space that includes users and their interactions [21][25] - The concept of a digital clone is introduced, where the model could eventually simulate a user's actions, preferences, and relationships in a virtual environment [21][25] - The long-term goal is to establish a new reality structure where AI can understand and participate in human decision-making processes [30][31]