起底“豆包手机”：核心技术探索早已开源，GUI Agent布局近两年，“全球首款真正的AI手机”

Core Insights - The "Doubao Phone" has gained significant popularity, with the first batch of 30,000 units selling out quickly and prices doubling in the second-hand market. This phone features a technical preview of the Doubao Phone Assistant, which automates complex tasks across applications [1][36]. Group 1: Technology and Development - The Doubao Phone Assistant is built on ByteDance's nearly two-year investment in the "system-level GUI Agent" space, showcasing its ability to automate tasks like leave requests and travel bookings [1][3]. - The core technology behind the Doubao Phone Assistant is the UI-TARS model, which has been optimized for mobile use and outperforms its open-source predecessor [3][17]. - The UI-TARS model has undergone significant evolution, with the latest version, UI-TARS-1.5, introducing reinforcement learning mechanisms that enhance reasoning capabilities before executing actions [17][30]. Group 2: Performance Metrics - UI-TARS-1.5 has achieved state-of-the-art (SOTA) results in various benchmarks, outperforming competitors like OpenAI's CUA and Claude 3.7 in tasks related to computer and browser use [18][19]. - In gaming scenarios, UI-TARS-1.5 demonstrated superior performance compared to other models, achieving perfect scores in several games [23][32]. Group 3: User Experience and Feedback - Users have praised the Doubao Phone for its ability to handle tasks autonomously, with one entrepreneur describing it as the "world's first true AI smartphone" [46][47]. - The assistant's capability to interact with various applications seamlessly, even in a non-English interface, has been highlighted as a significant advancement in mobile technology [55][56]. Group 4: Privacy and Security Concerns - The Doubao Phone Assistant has faced scrutiny regarding its use of system-level permissions, but the company clarified that user consent is required for such operations, similar to existing voice assistants [36][41]. - Users have expressed confidence in the assistant's design, noting its isolation and local processing capabilities, which mitigate potential privacy risks [42][43].