OpenAI Unveils Revolutionary AI Model “o1” with Enhanced Problem-Solving Capabilities

OpenAI has released its latest AI model, named “o1”, which is generally believed to be the secret project code-named “Strawberry” that was reported earlier. According to OpenAI, this model can spend more time thinking and deliberating on questions than past models; try different strategies to solve problems; and recognize errors and correct answers to improve the accuracy of answers, “just like real people.”

The new model name not only starts with “1” again, but also removes “GPT”. This means that the o1 and GPT series adopt different strategies. They no longer rely on statistical predictions to “guess” what the next word should be, but instead consider the logic of the questions and answers as a whole. OpenAI says it performs similarly to PhD students in the fields on solving difficult problems in physics, chemistry and biology, and even better in areas such as mathematics and programming. OpenAI’s current flagship model, GPT-4o, only correctly answered 13% of the questions in the International Mathematical Olympiad qualifying exam, while o1 solved a whopping 83%.

Another “initiative” of o1 is to open “thinking time” as another dimension to enhance AI results. In existing AI models, most of the time is spent on model training. Once the model is launched, the computing resources consumed for each use are relatively low. However, the emergence of o1 proves that investing more computing resources in the use stage can also strengthen the answers to individual questions. Unlike GPT, which generally emphasizes speed, o1’s current calculation time can take ten seconds or even tens of seconds, so it will be more suitable for applications that require response quality higher than speed.

The o1 model does not have the ability to browse the Internet, upload files or images, and there is no significant improvement in text generation compared to GPT-4o. In addition, the speed is slow, and it is not suitable for most people’s general applications and business applications. It is said that GPT-4o may still be a more suitable choice. At present, OpenAI is more aimed at academia, especially fields such as theoretical physics that require complex mathematical operations. However, despite having a longer time to consider its output, o1 has not yet completely solved the problem of “hallucinations”, so it will ultimately be up to humans to do the final check.

The development of o1 is still in its infancy, with two models, o1-preview and o1-mini initially launched. The former is a “preview version” of the complete o1, with a slight discount in capabilities; while the o1-mini is a lightweight model that is faster, cheaper, and is particularly good at writing programs. Due to the significant increase in the computing power requirements of o1, it is currently only open to paid users for trial, and there are very strict usage restrictions – o1-preview can only send a few 30 messages per week, and o1-mini can only send 50 messages.

Key Features of OpenAI’s o1 Model

• Enhanced problem-solving capabilities
• Ability to spend more time thinking and deliberating on questions
• Improved accuracy in answering questions
• Adoption of different strategies compared to GPT series
• More suitable for applications that require response quality higher than speed