Meet OpenAI o1: The Next Level AI Model with Human-Like Reasoning Skills

Previously, it was reported that OpenAI was preparing to launch a new artificial intelligence model that could perform some human-like reasoning tasks. This model, referred to as Strawberry both internally at the company and in media reports, finally revealed its official name with the official launch of the large model - OpenAI o1.

According to previous reports and the company's promotion, this new AI tool can solve multi-step problems. It will also think through the process of solving complex mathematics and programming problems like a human. For complex reasoning tasks, this is a significant advancement and represents a new level of AI capability. Given this, we are resetting the counter back to 1 and naming this series OpenAI o1."

Experimental data also shows that compared to its predecessors, o1 indeed demonstrates stronger reasoning capabilities: this new model performs on par with doctoral students (i.e., expert-level talents) in challenging benchmark tests in subjects like physics, chemistry, and biology. Moreover, in a qualifying exam for the International Mathematical Olympiad (IMO), its accuracy rate is as high as 83%, nearly six times higher than the 13% accuracy rate of GPT-4o. Additionally, it outperformed 89% of human competitors in the Codeforces online programming competition.

OpenAI stated that o1 is suitable for individuals solving complex problems in science, programming, mathematics, and similar fields. At the same time, because it is also adept at accurately generating and debugging complex code, it can also perform efficiently in programming. Of course, its powerful reasoning capabilities also enable it to support applications that require reasoning but do not need extensive common sense.

However, OpenAI has only recently launched the o1-preview preview version and the o1-mini mini version, and it is rolling out to paid users, free users, and developers in stages. The preview version of the model is already available to paid Plus and Team users through OpenAI's ChatGPT, while ChatGPT Enterprise and Edu users will be able to use both models next week, and at some unknown time in the future, all users will have free access to o1-mini.

At the same time, since the currently released o1 is only the initial version, it also has many issues such as the inability to browse web information, upload images, and usage limits. Some users have also found that in some cases, o1's reasoning capabilities are not so impressive because it seems to take a long time to answer questions.

In a technical paper, OpenAI also acknowledged that the current o1 has some flaws, and the company's research scientist Noam Brown also said that releasing the preview version of the model now is just to better understand how people will use it and areas it needs to be improved.

OpenAI has been working hard to get computers to perform multi-step operations. For example, in May 2023, the company released a blog post and an accompanying research paper introducing its efforts to improve the ability of AI systems to solve mathematical problems. According to the paper, the company trained the model by adopting a technique called chain of thought reasoning, breaking down complex problems into smaller logical steps, and rewarding the model for each correct step in the process of solving a problem, rather than just rewarding it for generating an accurate answer.

OpenAI also admits that there is a lot of pressure in training AI model reasoning capabilities. The San Francisco-based unicorn is reportedly seeking to raise billions of dollars in funding to develop more complex artificial intelligence systems to face increasingly fierce industry competition: in addition to OpenAI, Anthropic, and Google have also demonstrated the reasoning skills of their advanced AI models.