Open Source Securities: OpenAI o1 model is coming, with RL assisting models in improving reasoning abilities.

Open Source Securities released a research report stating that the early AI large language model performance improvement is mainly achieved through the Scaling law. With the continuous improvement in corpus dataset and model parameters, the performance of the model continues to break through. The emergence of the new generation o1 series models represents a way for the model to improve its performance through new technologies such as RL on the reasoning side, apart from the training side, introducing techniques like the Chain of Thought, providing more accurate answers for professional fields such as science, mathematics, and coding, and marking an important turning point in the development of generative AI. In addition, OpenAI stated that besides the new OpenAI o1 series, they are still continuing to develop the GPT series models. OpenAI releases o1-preview and o1-mini models, significantly improving inference capabilities On September 13, 2024, Beijing time, OpenAI released the new generation o series models o1 and o1-mini, introducing large-scale reinforcement learning (RL) during model training. With an increase in train-time compute time and extension of test-time compute time during training, the performance of the o1 model continues to improve. Through RL training, the o1 model uses the Chain of Thought method during inference to solve problems, breaking down complex issues into simple steps and promptly recognizing and correcting errors, significantly enhancing the model's inference capability. With maximum inference time, o1 outperforms GPT-4o in the vast majority of inference-intensive tasks. Model testing results show that in 57 MMLU subclasses, o1 outperforms GPT-4o in 54 subclasses and performs comparably to human experts. The o1 model has not been priced separately and is planned to provide access to o1-mini for free for ChatGPT users. In terms of inference cost, the o1 model is predominantly pre-trained on large text datasets, leading to high inference costs and slower speeds. The o1-mini model is optimized for STEM reasoning during pre-training, especially excelling in mathematics and coding fields, with smaller model parameters and lower latency inference, 80% lower inference cost than OpenAI o1-preview. Currently, ChatGPT Plus and Team users will have access to the o1 series models in ChatGPT, while Enterprise and Edu users will start access next week. OpenAI currently only provides preview versions of the models in ChatGPT and API, with plans to launch browsing, file and image uploads, and other features in the future. o1-mini access will be provided to ChatGPT Free users in the future. The o1 model is not priced separately, but there are usage restrictions, with a weekly limit of 30 messages for the o1-preview model and 50 messages for the o1-mini model. AI large models continue to iterate, with cloud giants at home and abroad increasing capital expenditures on AI infrastructure. The strength of the computing industry chain remains promising. Recommended: Shanghai Baosight Software (600845.SH), Zhongji Innolight (300308.SZ), Shenzhen Envicool Technology (002837.SZ), Eoptolink Technology Inc., (300502.SZ), Suzhou TFC Optical Communication (300394.SZ), ZTE Corporation (000063.SZ), Shengketong Communication (688702.SH). Beneficiaries: Range Intelligent Computing Technology Group (300442.SZ), Yuanjie Semiconductor Technology (688498.SH), Huagong Tech (000988.SZ), Unisplendour Corporation (000938.SZ), Accelink Technologies (002281.SZ), Sichuan Huafeng Technology (688629.SH), Wangsu Science & Technology (300017.SZ), Fiberhome Telecommunication Technologies (600498.SH), INESA Intelligent Tech Inc. (600602.SH), and others. Risk Warning: AI development falling short of expectations, insufficient construction of smart computing centers, intensifying industry competition.