从推荐算法优化到AI4S、Pico和大模型，杨震原长文揭秘字节跳动的技术探索

Group 1 - The core viewpoint of the article highlights ByteDance's commitment to fostering academic excellence through the ByteDance Scholarship, which has increased its award amount and expanded its reach to more universities [1] - The scholarship attracted over 500 applicants from 66 universities in China and Singapore, with 20 students awarded in various fields including AI and robotics [1] - The award amount has been upgraded from 100,000 yuan to 200,000 yuan, which includes 100,000 yuan in cash and 100,000 yuan for academic resources [1] Group 2 - ByteDance's Vice President of Technology, Yang Zhenyuan, shared insights on the company's technological explorations and encouraged students to tackle significant technical challenges [1][2] - The company has been involved in large-scale machine learning and recommendation systems since 2014, aiming to build a recommendation system capable of handling trillions of features [7][10] - ByteDance has made significant advancements in scientific computing, particularly in solving the Schrödinger equation to simulate various phenomena, indicating a focus on AI's potential in real-world applications [13][15] Group 3 - In 2021, ByteDance acquired Pico to explore XR technology, aiming to enhance user experience through hardware innovations [27] - The company is focusing on improving clarity in XR devices, targeting a pixel density (PPD) of nearly 4000, significantly higher than existing products [29][32] - ByteDance is also developing a dedicated chip for XR devices to address processing challenges, achieving a system latency of around 12 milliseconds [34][35] Group 4 - The emergence of large models, particularly after the launch of ChatGPT, has prompted ByteDance to invest in AI technologies, leading to the development of popular AI applications like Doubao [39][40] - The company has established a robust training infrastructure, achieving a floating-point utilization rate exceeding 55%, which is significantly higher than mainstream frameworks [39] - ByteDance is exploring the future of large models and their potential impact on various industries, emphasizing the need for continuous learning and interaction capabilities in AI [43][44]