“一切都那么难”,上市后智谱创立发起人披露失误过程与发展目标

Core Insights - Beijing Zhiyu Huazhang Technology Co., Ltd. (02513.HK) has officially listed in Hong Kong on January 8, 2023, and is set to launch its next-generation model GLM-5 soon [2] Group 1: Company Development - The company launched its self-developed large model algorithm architecture GLM in 2020, attempting to train a base model with 10 billion parameters, which was trialed by several enterprises including Meituan [2] - From 2021 to 2022, the development of large models faced challenges, with skepticism about the feasibility of making machines think like humans, leading the team to decide to train a larger model with 130 billion parameters [3] - In mid-2022, the GLM-130B model was created, and the MaaS platform was launched, resulting in the first batch of real API users [3] Group 2: Market Trends and Challenges - In 2023, the company recognized that AI would disrupt search and browsers, leading to a new AI assistant for everyone, which would fundamentally reshape computing logic [3] - The emergence of DeepSeek served as a wake-up call for the team, highlighting the challenges ahead, including a nationwide price war and the need to find a precise breakthrough [4] - The company identified a focus on coding as a direction, leading to the release of GLM-4.1 and GLM-4.5, with 150,000 developers using the GLM Coding Plan, resulting in an annual recurring revenue (ARR) exceeding 500 million, a 25-fold increase in just 10 months [4] Group 3: Future Focus and Innovations - The company plans to concentrate on the GLM-5 base model, a new model architecture design, and stronger generalization capabilities through reinforcement learning (RL) by 2026 [4] - The next generation of learning paradigms, including online learning and continual learning, will be strategically developed [4] - A new department, X-Lab, has been established to attract young talent for frontier exploration in model architecture and cognitive paradigms, while also expanding external investments to connect the industry [5]