北京时间2024年9月13日，OpenAI发布新一代o系列模型o1以及o1-mini，在模型训练时引入大规模reinforcement learning(RL)，随着训练时计算(train-timecompute)时间的增加以及推理(test-time compute)时间的延长，o1模型性能持续提升。Early AI large language model performance improvement mainly relies on the Scaling law, as the corpus dataset and model parameters continue to improve, the model's performance continues to break through. The emergence of the new generation o1 series models represents that the model can use RL on the inference side apart from the training side.
Smart Finance APP learned that Open Source Securities released research reports stating that the early AI large language model performance improvement mainly relies on the Scaling law, as the corpus dataset and model parameters continue to improve, the model's performance keeps breaking through. The introduction of new technologies like thinking chains on top of RL on the inference side, aside from the training side, enhances the model's performance, providing more accurate answers for professional fields such as science, mathematics, and coding, or being a crucial turning point in the development of generative AI. In addition, OpenAI mentioned that besides the new OpenAI o1 series, they are still continuing to develop the GPT series models.
OpenAI released the o1-preview and o1-mini models, significantly improving their inference capabilities.
On September 13, 2024, Peking Time, OpenAI unveiled the new generation o series models o1 and o1-mini, introducing large-scale reinforcement learning (RL) during model training. With the increase in training time compute and extension of test-time compute, the performance of the o1 model continues to improve.
Through RL training, the o1 model uses a Chain of Thought approach during inference to solve problems, breaking down complex issues into simple steps, promptly identifying and correcting errors, significantly enhancing the model's inference capabilities. In maximizing inference time, o1 outperforms GPT-4o in the vast majority of inference-intensive tasks. Model test results demonstrate that out of 57 MMLU subclasses, o1 performs better than GPT-4o in 54 subclasses, reaching human expert level performance.
The pricing for the o1 model has not been decided yet, with plans to provide o1-mini access to chatgpt free users.
In terms of inference cost, the o1 model is mainly pre-trained on large text datasets, resulting in higher and slower inference costs. The o1-mini model is optimized for STEM reasoning during the pre-training phase, especially excelling in the areas of mathematics and coding, with smaller model parameters and lower latency inference, offering an 80% lower inference cost compared to OpenAI o1-preview.
Currently, ChatGPT Plus and Team users will be able to access the o1 series models in ChatGPT. Enterprise and Edu users will have access starting next week. OpenAI currently only provides a preview version of the models in ChatGPT and API, and will launch browsing, file and image upload, and other features in the future. The o1-mini usage permission will be provided for ChatGPT Free users in the future. The o1 model is not priced separately, but there are usage limitations. The weekly usage limit for the o1-preview model is 30 messages, and the o1-mini has a weekly usage limit of 50 messages.
Large AI models are continuously iterated, and domestic and foreign cloud giants are increasing their capital expenditure on AI infrastructure. The industry chain of computing power is expected to continue to perform well.
Recommended symbols: 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), Shengke Communication (688702.SH).
Beneficiary symbols: Range Intelligent Computing Technology Group (300442.SZ), Yuanjie Technology (688498.SH), Huagong Tech (000988.SZ), Unisplendour Corporation (000938.SZ), Accelink Technologies (002281.SZ), Huafeng Technology (688629.SH), Wangsu Science & Technology (300017.SZ), Fiberhome Telecommunication Technologies (600498.SH), Inesa Intelligent Tech Inc. (600602.SH), and others.
Risk reminders: AI development falls short of expectations, the construction of smart computing centers falls short of expectations, and industry competition intensifies.
