OpenAI's latest model, o1, has stirred significant waves in the AI community. This model, once jokingly referred to as "Strawberry" within the company, offers ChatGPT users a new opportunity to experience the AI's "thinking" process. However, is this highly anticipated model truly worth trying? The answer seems to be: it has its pros and cons.
The o1 model has indeed shown significant progress in certain aspects. Its performance in reasoning and solving complex problems is impressive, particularly its "multi-step reasoning" capability. This ability allows the model to "think" before answering, breaking down large problems into smaller steps and attempting to identify the correctness of each step. Although this approach is not a new concept, o1's implementation makes it a viable technology.
However, the o1 model also has some notable flaws. Compared to GPT-4o, o1 is inferior in tool usage, multimodal capabilities, and processing speed. More importantly, the cost of using o1 is about four times that of GPT-4o, which will undoubtedly affect its widespread adoption. OpenAI even admits that for most everyday tasks, GPT-4o remains a better choice.
The o1 model has demonstrated its unique value in practical applications. For example, in helping plan a Thanksgiving dinner, o1 can provide detailed and thoughtful suggestions, considering factors such as oven management and cost control. However, when dealing with simple problems, o1 seems to "overthink," providing lengthy and unnecessary answers.
Industry experts have mixed reviews on the o1 model. Professor Ravid Shwartz Ziv of NYU's AI Model Research believes that although o1 is impressive, the overall improvement is not significant. Research engineer Rohan Pandey from ReWorkd states that while o1's reasoning capabilities are strong, they do not represent a revolutionary advancement.
The pricing strategy of the o1 model has also sparked discussions. In addition to the regular input and output fees, users also need to pay extra for the model's "thinking process" in the form of "reasoning tokens." This pricing model emphasizes the necessity of using o1 cautiously to avoid unnecessary high costs on simple problems.
Despite these limitations, the o1 model still showcases the potential of AI in complex reasoning. It could become a powerful tool for users to examine major decisions, especially in scenarios that require in-depth analysis and consideration from multiple angles. However, whether o1 is worth its high price remains a debatable issue.
OpenAI's CEO, Sam Altman, has also tempered external expectations, emphasizing that o1 is not a general artificial intelligence and still has many flaws and limitations. This candid attitude also reflects the AI industry's rational perception of technological progress.