据他透露，阶跃星辰积极探索新的技术范式，已经在万亿参数模型上实现了强化学习训练的方法论。同时，持续打磨更高性能的底层大模型，提升C端产品体验。近期阶跃星辰自研的Step-2万亿参数MoE语言大模型已接入智能助手“跃问”，展现出更强大的指令跟随、创作和推理能力。①OpenAI's large model o1 enables AI to have the ability of human slow thinking (actively reflecting, correcting complex thinking) through reinforcement learning, and the next key breakthrough will be to improve the generalization ability of the reinforcement learning model and accelerate the integration of multimodal understanding and generation, which is a further breakthrough. ②GPT-4o signifies the progress of multimodal integration, and it is the foundation for better modeling the physical world and thoroughly realizing the simulated world.
On September 19th, in the 2024 Yunqi Conference, Dr. Jiang Daxin, the founder and CEO of Star Leap, stated that the development of AI is undergoing a critical technological paradigm shift: OpenAI's large model o1 has explored a way to enable AI to have the ability of human slow thinking (actively reflecting, correcting complex thinking) through reinforcement learning. The next key steps for AI technology are to improve the generalization ability of the reinforcement learning model and to accelerate the integration of multimodal understanding and generation.
The person second from the left is Dr. Jiang Daxin, CEO of Star Leap.
He disclosed that Star Leap is actively exploring new technological paradigms and has already implemented the methodology of reinforcement learning training on trillion-parameter models. At the same time, they are continuously improving the performance of the underlying large models and enhancing the end-user product experience. Recently, Star Leap's self-developed Step-2 trillion-parameter MoE language model has been integrated into the intelligent assistant 'Leap Question,' demonstrating stronger command following, creation, and reasoning capabilities.
At the Yunqi Conference, the latest release of the o1 model by OpenAI has sparked discussions. The o1 model is considered to have explored a way to enable AI to have the ability of human slow thinking (actively reflecting, correcting, trying different strategies, and conducting complex reasoning) through reinforcement learning technology. This is the first time a large model has simultaneously possessed the ability of human brain System 1 (fast thinking, giving direct answers) and System 2 (slow thinking). Jiang Daxin believes that this is a critical step for large models to begin to possess the ability to infer the world.
"We divide the path to AGI into three parallel stages: simulating the world, exploring the world, and inferring the world. In the past year, there have been breakthrough technological advancements in all three areas, and the pace of development can be described as 'a day in AI, a year on Earth'". He stated that in addition to the progress made by o1 in inferring the world, GPT-4o signifies the progress of multimodal integration, which is the foundation for better modeling the physical world and completely realizing the simulated world. Regarding exploring the world, Tesla's release of the fully autonomous driving system FSDV12 indicates the technological direction for how embodied intelligent devices can be combined with large models, transitioning from the digital world to exploring the physical world.
In climbing the road to AGI, strong foundational models are needed for reinforcement learning, multimodal understanding, and industry models. This year, Star Leap has continued to iterate the trillion-parameter language model Step-2, and in March, it was the first in China to release a preview version of the trillion-parameter language model. In July, it was officially released. After testing, Step-2's comprehensive capabilities have increased by nearly 50% compared to the hundred-billion-parameter language model Step-1, with significant improvements in logical reasoning, mathematics, programming, and knowledge.
Meanwhile, YUE Xingchen announced the comprehensive upgrade of its smart assistant "YUEWEN", offering users the Step-2 trillion parameter MoE language model capability for free for a limited time, and launching the multi-modal search and question-answering feature "PAI ZHAO WEN". Users can achieve "capture and ask" through image interaction, solving the scenario demands that are difficult to efficiently describe in text and voice interaction modes.
