深度机智（北京）科技有限公司创始人陈凯：用人类“第一视角”重构具身智能“大脑”

Core Viewpoint - The development of embodied intelligence in China is currently rated very low, with expectations for improvement by 2025 being minimal, around 1 to 0 out of 10 [1][2]. Group 1: Company Overview - Deep Intelligence (Beijing) Technology Co., founded by Chen Kai, aims to enhance the physical intelligence of foundational models using human "first-person" data [2][3]. - The company was established in May 2025, with a team that has a high proportion of PhD holders, focusing on a unique technical path that does not rely on expensive motion capture equipment [3][4]. Group 2: Technical Approach - The company collects "first-person" video data from real-world scenarios to build a general embodied intelligence model, which has faced skepticism from investors initially [2][4]. - Chen Kai believes that human "first-person" data contains deep laws of the physical world that cannot be fully described in words or rules, and this data needs to be compressed into large models for better understanding [4][6]. Group 3: Market Validation - The shift in Tesla's approach to reduce reliance on remote operation data in favor of "first-person" video learning has validated the company's technical direction [4][5]. - The emergence of Generalist AI and Physical Intelligence has further confirmed the importance of real-world data in enhancing model generality, aligning with the company's hypotheses [5][6]. Group 4: Data Collection and Goals - The company aims to reach a data collection scale of "one million hours" by mid-2026, which is expected to significantly improve the understanding of physical intelligence and validate the Scaling Law [7][8]. - Currently, the company collects over 1,000 hours of data daily, but achieving the target requires extensive data cleaning and processing [7]. Group 5: Industry Perspective - The gap between China and the U.S. in embodied intelligence is reportedly widening, primarily due to a lack of convergence in technical paths among many companies [10]. - However, there is optimism for 2026 as the industry is expected to accelerate, with increased investment and a clearer consensus on data collection methods [10][11]. Group 6: Future Outlook - Key themes for the future of embodied intelligence include acceleration, scaling of data and models, and a sense of hope for overcoming initial skepticism in the industry [11]. - The company believes that the cost advantages of collecting "first-person" data in China could lead to a competitive edge in the global market [10].