OpenAI releases o1 model with reasoning capabilities

2024-09-13

OpenAI has released a new model called o1, which is one of the planned "inference" models by the company. This model aims to provide faster solutions to complex problems. o1 will be launched alongside a smaller and more cost-effective version called o1-mini. Although OpenAI labels this release as a "preview" to indicate its early stage, the model has garnered significant attention due to high expectations. For OpenAI, o1 represents an important step towards achieving human-level artificial intelligence. In terms of practicality, it outperforms previous models in coding and multi-step problem-solving. However, compared to GPT-4o, o1 is more expensive to use and slower in speed. OpenAI states that ChatGPT Plus and Team users can access o1-preview and o1-mini, while Enterprise and Edu users will gain access next week. As for free users, they will be granted access to o1-mini at some point in the future. The training method of o1 differs significantly from its predecessor. Jerry Tworek, the research lead at OpenAI, revealed that o1 utilizes a new optimization algorithm and a specially curated training dataset. Unlike the GPT models that learn by imitating patterns in training data, o1 employs reinforcement learning techniques for self-problem-solving and is trained through reward and punishment. Additionally, it utilizes a "thinking chain" to process queries, similar to the step-by-step problem-solving process of humans. It is claimed that this new training method makes the model more accurate and reduces instances of error generation, although completely eliminating errors remains a challenge. The new model demonstrates excellent performance in solving AP math tests and International Mathematical Olympiad problems, even reaching the 89th percentile of participants in Codeforces programming competitions. However, o1 falls short in terms of world knowledge compared to GPT-4o and lacks the ability to browse the internet or handle files and images. Nevertheless, OpenAI considers o1 as representing a new category of functionality. The model is named o1, which signifies "resetting the counter to 1". OpenAI hopes to convey a new naming logic through this name. Despite exhibiting stronger capabilities in handling complex problems, o1 is not equivalent to genuine human thought processes. OpenAI emphasizes that the model was not designed to be synonymous with human thinking but rather aims to showcase how the model can delve deeper into problem-solving through its interface. However, even in simulating human thought processes, o1 is not a true thinking entity.