OpenAI has officially launched its highly anticipated AI model, previously referred to by the code name “Strawberry.” The newly introduced model, named “OpenAI o1-preview,” is engineered to “spend more time thinking” prior to delivering responses, which gives it the capability to tackle more intricate tasks and challenging problems.
The o1 model has been widely speculated to be a major advancement towards achieving artificial general intelligence, a theoretical stage where AI surpasses human capabilities. The primary objective is to equip the model with a form of “reasoning” that enables it to resolve more complicated mathematical challenges, among other tasks.
According to OpenAI, the model exhibits impressive academic performance, claiming that it “performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology.” However, it is important to note that the o1-preview is still in the early stages of development, and numerous updates are anticipated.
The company acknowledged that as an early model, it does not yet have many of the features that make ChatGPT useful, like browsing the web for information and uploading files and images. For many routine tasks, it suggests that GPT-4o will remain more capable in the short term.
Currently, the OpenAI o1 model is accessible to users of ChatGPT Plus and Team. Additionally, there are plans to introduce a lighter version named o1-mini for all free ChatGPT users, although a release date has yet to be disclosed.
The model is also subject to usage limits, with a cap of 30 messages per week for o1-preview and 50 for o1-mini.
Chain of Reasoning
The o1 model addresses several reasoning challenges that typically hinder generative AI models by effectively fact-checking itself through a more deliberate consideration of all aspects of a question. OpenAI emphasises that what distinguishes o1 from other generative AI models is its capacity to “think” before answering queries.
When allowed additional time to “think,” o1 can approach a task comprehensively, planning and executing a series of actions over an extended period to arrive at a solution. This capability makes o1 particularly adept at tasks requiring the synthesis of multiple subtasks, such as identifying privileged emails in a lawyer’s inbox or developing a product marketing strategy.
OpenAI o1’s Limitations
Despite its advancements, OpenAI o1 still has its shortcomings. The model may exhibit slower response times compared to others, as it can sometimes take over 10 seconds to respond to certain queries.
Furthermore, OpenAI’s technical documentation revealed anecdotal feedback from testers indicating that o1 tends to hallucinate—confidently generating incorrect information—more frequently than GPT-4o, and it is less likely to admit when it does not know the answer to a question.
While OpenAI is the first to launch the o1 model, the real challenge will be making it widely accessible and affordable, especially as competitors are likely to develop similar models. The company aims to experiment with o1 models capable of reasoning for extended periods to enhance their “thinking” abilities.