Core Insights - The article discusses the transition of AI from model capability competition to execution capability in the physical world, highlighting the challenges and opportunities in this domain [2][3]. Company Summaries - Zhi Bian Liang is focused on developing embodied intelligence foundational models and general-purpose robots, emphasizing the need for a physical model that operates in the real world, distinct from virtual models [4]. - Yuan Rong Qi Xing has been involved in autonomous driving, witnessing the industry's evolution from high-precision mapping to end-to-end models, and has successfully deployed 200,000 vehicles with their driving assistance systems, with a projection of reaching one million vehicles next year [5]. Challenges in AI Implementation - The transition from simulation to real-world application presents significant challenges, including the need for extensive pre-training based on real-world data, which is not easily replicated in simulated environments [6][7]. - The physical world introduces complexities that are not present in simulations, such as the need for precise manipulation and the impact of minor errors on outcomes [9][10]. Importance of Data and Training - The collection of vast amounts of real-world data is crucial for effective pre-training, and the integration of language models can enhance learning efficiency [7][18]. - The current data generation from 200,000 vehicles is substantial, necessitating careful selection and quality control to optimize model performance [18]. Future of Commercialization - The commercialization of embodied intelligence is expected to gain momentum by 2026, with predictions of significant advancements in practical applications and return on investment [21][22]. - The industry is currently in a phase similar to early autonomous driving, with many companies still in the demo stage, but there is optimism about achieving scalable commercial applications soon [19][20]. Role of Language Models - Language models are seen as essential for providing supervisory information during training, aiding in the rapid learning of complex tasks [12][13]. - However, there is debate about the necessity of language in physical AI, with some arguing that while it enhances understanding, it may not be critical for all applications [15][26]. Technical Considerations - The development of physical AI models requires overcoming significant engineering challenges, including the need for real-time feedback and the limitations of current computational resources [25][26]. - The scaling laws in AI suggest that with sufficient data and resources, it is feasible to train models that can operate effectively in the physical world within a reasonable timeframe [24][26].
三个人，聊了很多AI真相