这种能力使得o1能够在给定更多的“思考”时间时，回答更难的问题，类似于在国际象棋比赛中关键时刻考虑可能的走法。这种推理能力是AI发展的一大步，它使得AI能够在更广泛的领域中应用，包括但不限于编程、数学和科学问题。As the first big model with true general reasoning ability, the O1 model shows the ability to handle complex problems and reasoning tasks, or heralds a historic turning point in artificial intelligence technology.
September 12, 2024 will probably be remembered as a historic turning point in the history of artificial intelligence. On this day, OpenAI announced its latest series of AI models — O1. According to officials, O1 is the smartest model in history, and its way of thinking is closer to humans, marking a new path for humans to enter the age of reasoning.
The core of the o1 model lies in its ability to reason. Unlike models based on predictions, o1 thinks deeply before answering questions, exploring different paths until it finds the best answer.
This ability allows o1 to answer more difficult questions when given more time to “think”, similar to considering possible moves at critical moments in a chess game. This reasoning ability is a major step in the development of AI. It enables AI to be applied in a wider range of fields, including but not limited to programming, mathematics, and scientific problems.
The subtle differences between the O1 model and the GPT series may be difficult for ordinary users to notice in everyday use, but for researchers studying AI technology in depth, this change is revolutionary.
According to an article in “The Atlantic” last week, researchers have recognized that compared to older models such as GPT, the O1 model is more than a simple upgrade, but represents a “completely different game rule” and “real progress.” As the first big model with true general reasoning ability, the O1 model shows the ability to handle complex problems and reasoning tasks, or heralds a historic turning point in artificial intelligence technology.
Mark Chen, vice president of research at OpenAI, pointed out earlier this year that o1 is fundamentally different from standard ChatGPT because it can “reason”, which is a sign of human intelligence. Shortly thereafter, OpenAI CEO Altman announced “the age of intelligence is coming,” and artificial intelligence will help humans solve climate problems and colonize space. Last week, the startup released to the public a complete first-generation version of the o1 with fully mature reasoning capabilities.
The first big model with true general-purpose reasoning ability
According to OpenAI, the O1 model is the first large model with true general reasoning capabilities, and this is fully reflected in GPQA-Diamond, a benchmark for testing expertise in chemistry, physics, and biology.
According to OpenAI's assessment, o1 completely surpassed human doctoral experts in this test, with an accuracy rate of 78.3%, while human experts scored 69.7%. According to OpenAI, “there are no obstacles to the path of general artificial intelligence (AGI).”
Furthermore, o1 excelled in several difficult reasoning benchmarks. For example, in the American Math Tournament (AIME), o1's accuracy rate reached 83.3%, far surpassing the performance of its predecessor GPT-4o.
The launch of the O1 model has also enabled domestic and foreign AI companies to pay close attention to each other's dynamics with the intention of achieving breakthroughs in the field of AI big models. A number of emerging big models have emerged on the market, such as Kimi's k0math, DeepSeek's DeepSeek's DeepSeek R1-Lite, and Kunlun World Wide's “Tiangong University Model 4.0” o1 version. They are trying to compete with O1 in terms of logical thinking and deep reasoning abilities.
