OpenAI o1初探：或能成为引领AI Phenomenal Ride的LLM新范式

Investment Rating - The report maintains an "Outperform" rating for the information services industry [1]. Core Insights - OpenAI has launched the o1 series of large models, marking a new era in AI models with enhanced reasoning capabilities, particularly in complex tasks such as science, programming, and mathematics [6][14]. - The o1 model demonstrates significant improvements over the previous GPT-4o model, especially in reasoning accuracy and safety, indicating a potential for further advancements in AI technology [24]. Summary by Sections 1. OpenAI o1 Series Launch - On September 12, OpenAI announced the development of new AI models that excel in reasoning and problem-solving, achieving performance levels comparable to PhD students in challenging tasks [6]. - In the International Mathematical Olympiad qualification exam, OpenAI o1 achieved an 83% correct rate, significantly outperforming GPT-4o, which only managed 13% [6][9]. 2. o1-mini Model - OpenAI also introduced the o1-mini model, which is cost-effective and performs exceptionally well in STEM fields, particularly in mathematics and programming, achieving scores comparable to the o1 model while being 80% cheaper [26][30]. - The o1-mini model's reasoning speed is 3 to 5 times faster than that of o1-preview, making it suitable for applications requiring quick responses [27][30]. 3. New Reasoning Paradigm - The o1 model employs a new reasoning paradigm that utilizes chain-of-thought (CoT) prompting and self-taught reasoning techniques, enhancing its ability to tackle complex problems [40][41]. - This approach allows the model to optimize its reasoning process, identify errors, and break down complex tasks into simpler steps, significantly improving its reasoning capabilities [41][48].