OpenAI finally released the new model O1 on September 12. Users can directly access the preview version o1-preview, or the small version o1-mini. Such a highly anticipated product is bound to have a profound impact on the industry and even society.
OpenAI o1 technical innovation
Compared with GPT-4o, o1-preview has improved its ability to solve mathematical and programming problems by more than 5 times, while o1, which has not yet been released, has improved by more than 8 times! The success rate in solving doctoral-level scientific problems has exceeded the level of human experts. In the International Mathematical Olympiad (IMO) qualifying exam, GPT-4o only correctly solved 13% of the problems, while the reasoning model scored 83%; the programming ability exceeded 89% of human players in the Codeforces competition.
In actual operation, the reasoning process of the new model is very different from before. We can see that there is an additional Show chain of thought box that can be opened and closed, showing the entire thinking process. Just like a human thinking for a long time before answering a difficult question, o1 will decompose the problem by thinking when trying to solve it, and think about each small task repeatedly, recognize and correct errors. When one method does not work, it will try another method, which greatly improves the reasoning ability of the model.
o1 represents an important progress in the field of artificial intelligence, but if we look at it from a long-term perspective, we will find that OpenAI still has a long way to go. First of all, in terms of technical barriers, OpenAI's greatest contribution is to be firm and cost-effective in taking the lead in practice. Like Sora, once OAI has clarified the technical direction, it is only a matter of time before the engineering is reproduced, and it is simply an impossible task for OAI to win the world in all directions.
Judging from the results of the whole network evaluation, the model effect can only be said to be unsatisfactory. In many scenarios, it is not as good as the results of the thinking chain method under other engineering means (such as Claude3.5). It can even be said that it is often just a slightly higher probability of drawing cards, and its practical value is still difficult to determine.
Finally, the cost. The model that was basically completed last year was delayed for so long. In addition to the well-known security reasons, it may be because o1, like Sora, consumes too much computing power and is not feasible for large-scale commercial use. And it is well known that the OpenAI team has been trying to find a solution. They have been waiting for a long time, hoping that the computing power cost can decrease with the advancement of technology. They are also raising funds around the world to buy or rent more computing resources.
But the fact is that even after these efforts, the products launched still take several minutes or even dozens of minutes for a single reasoning, the unit price is several times higher than 40, and the token consumption often increases several times. In this context, whether OpenAI's industry status and valuation can be maintained has become quite uncertain. The high R&D and operating costs, coupled with the obstruction of commercialization, may affect investors' confidence and market expectations.
Despite this, it is undeniable that OpenAI o1 is an important breakthrough in the research of artificial intelligence technology. In the future, artificial intelligence will gradually evolve from a single large model to a flexible combination of different ability modules such as knowledge, reasoning, creativity, and emotion, and form a closer and more efficient collaboration with humans. o1 is just the beginning, and we look forward to a flourishing future.