OpenAI's New o1 Model is Transitional, Not a Breakthrough

Asianfin -- I saw the news that OpenAI has released the long-awaited new model Orion o1 today, but I was a little disappointed. I said in my last video that the iteration of the GPT large model has entered a bottleneck period. And o1 seems to be just a transitional product, not a breakthrough.

Essentially, o1 is trained through reinforcement learning in complex reasoning capabilities and security mechanisms. Indeed, significant progress has been made in these two aspects. However, the existing fundamental limitations of GPT and the underlying architecture based on Transformer have not changed.

The new model is not named Orion exactly, nor the project code Strawberry, but o1. Unlike GPT, the o1 model marks its own reasoning process and thinking time, but it seems that its knowledge base was only updated as of October 2023. The reason may be that it is still in the beta stage. From a functional point of view, it is similar to what I predicted in the last video.

First, it has more enhanced complex reasoning capabilities and is specially designed to deal with complex reasoning tasks, such as solving mathematical problems and generating complex code. In International Mathematical Olympiad, GPT 4o can only answer 13% of the problems correctly, while the correct rate of o1 model reaches 83%.

Second, its reinforcement learning is closer to the way human thinking and reasoning, and it introduces new security protocols in the security mechanism, becoming more compliant with the rules. For example, in one of the hardest jailbreaking test of OpenAI, the o1 Preview model scored 84, while the GPT 4o scored 22. These improvements all indicate that the o1 model will provide more possibilities for ToB scenario applications, rather than for ToC.

It is foreseeable that OpenAI will earn more from ToB business. The o1 model series focuses more on specialized complex tasks than general-purpose AI, making it an important tool for coding mathematics and scientific research experts, especially in scientific research, engineering calculation, medical treatment, data analysis and other scenarios. The advanced reasoning capabilities of the o1 model make it particularly suitable for enterprise users.

In addition, OPAI also released the o1 mini version, and the cost of o1 Mini is roughly 80% lower than that of o1 Preview. Designed to provide more cost-effective options for developers and professional teams, enterprises can also use these models to develop and deploy complex workflows and solutions, and these requirements are usually the mainstream requirements in the B2B market.

It should be noted that ToB has never been the main area for China’s application innovation. When we say that China leads in application innovation, we are talking about the ToC application market and individual user applications, not ToB. The U.S. ToB market is way more mature than the Chinese ToB market, so for a long time to come, we may no longer be able to claim that we lead in the AI applications, and it will become more difficult to catch up.