OpenAI officially launches o1-preview: AI wisdom upgrade!
OpenAI, a leader in the field of artificial intelligence, released its latest AI model "o1-preview" on Thursday. This new model emphasizes "spending more time thinking" before answering user questions, allowing it to better solve complex multi-step problems, especially in programming and mathematics problems.
New training method, greatly improved accuracy
The o1 model adopts a brand-new training method, teaching the system to gradually solve problems through "reinforcement learning" technology. According to OpenAI research director Jerry Tworek, o1 uses a training data set specially tailored for it and handles problems through a "thinking chain" so that AI can perform logical reasoning like humans. This new method not only greatly improves the accuracy of the model, but also reduces the "AI hallucination" phenomenon that was common in the past.
Excellent math and programming skills
According to Bob McGrew, chief research officer of OpenAI, o1 has demonstrated amazing capabilities in the fields of mathematics and programming. In the International Mathematical Olympiad qualifying test, o1's correct answer rate was as high as 83%, far exceeding the 4% of the previous generation GPT-13o. In addition, o1's performance in the famous online programming competition Codeforces was also among the best, reaching the top 89% of the participants. This proves its powerful reasoning and calculation capabilities in solving complex problems.
Early access for ChatGPT Plus users
Currently, ChatGPT Plus and Team users have early access to the o1-preview and o1-mini versions, while enterprise and education users will gain access early next week. It is worth mentioning that OpenAI plans to make o1-mini free to all ChatGPT users in the future, so stay tuned!
It’s more expensive to use, but worth the wait
Although the o1 model performs well in some aspects, compared with GPT-4o, the cost of use is indeed higher. The API usage cost of o1-preview is US$15 per million input tokens and US$60 per million output tokens, which are twice and three times that of GPT-4o respectively. For developers, this means paying a higher price for more powerful inference capabilities.
The functionality is not as good as GPT-4o, but it is a milestone
It should be noted that the o1 model is not as good as GPT-4o in some aspects (such as factual knowledge, web browsing, file and image processing). However, OpenAI regards this development as an important step for artificial intelligence to move towards human-like reasoning capabilities, and resets the counter to "1", symbolizing a new starting point for AI capabilities.
OpenAI believes that the launch of the o1 model represents a major advancement in artificial intelligence in reasoning tasks, laying a solid foundation for more complex AI applications in the future.
future outlook
OpenAI plans to further improve the reasoning capabilities of the o1 model in subjects such as physics, chemistry, and biology in future updated versions, and expects these capabilities to reach a level comparable to that of doctoral students. This series of progress is bound to allow OpenAI to continue to lead the forefront of global artificial intelligence technology.
