OpenAI Releases New AI Models That Think Before They Speak
OpenAI on Thursday released its new o1 series of artificial intelligence (AI) models. The AI company calls these reasoning models for their advanced capabilities in solving mathematical and complex reasoning problems. There are two models: o1, which is available in preview, and the o1-mini. The company said these AI models are trained to spend time thinking before responding, similar to humans. Interestingly, this is believed to be the same AI model that was previously reported to be Strawberry.
OpenAI o1 series AI models released
In a blog post, the AI company writes introduced new AI models with advanced reasoning capabilities. These models differ from standard generative AI because they do not process the entire prompt at once, but instead work through the problem systematically, similar to how humans would. This also allows the AI model to try different strategies and correct any errors. OpenAI points out that these models are slower than the GPT-4o model because they need an extra moment to think.
OpenAI o1 translates a corrupted sentence. photo.twitter.com/E37e4SOuq4
— OpenAI (@OpenAI) September 12, 2024
So, what does this mean for an average user? Users can ask complex questions to the AI that often require multi-level reasoning and critical evaluation. For example, a question like “Look at this series: 12, 11, 13, 12, 14, 13, … What number should come next?” that requires multi-step thinking can now be answered accurately by the AI.
A man walks into a library and asks the librarian for a book. The librarian points to a specific shelf. The man thanks her and leaves without taking a book. Why?
OpenAI claimed that the o1-preview model performs at a similar level to PhD students when answering questions in the subjects of physics, chemistry and biology. The model also shows a similar output when solving mathematical problems. “In a qualifying exam for the International Mathematical Olympiad (IMO), GPT-4o solved only 13 percent of the problems correctly, while the reasoning model scored 83 percent,” the post added.
Sam Altman, CEO of OpenAI, marked with an X (formerly known as Twitter) after that the o1 models were able to achieve a score of 78.3 on the PhD-level science benchmark GPQA Diamond. However, he added that the large language model (LLM) is still flawed, as it is the emerging version of the model. OpenAI plans to roll out updates to consistently improve it.
The o1 series AI models are currently available for ChatGPT Plus and Team users in preview. However, there is a weekly rate limit of 30 messages for o1 and 50 messages for o1 mini. The company stressed that these limits may be increased in the future. One of the reasons for imposing the rate limits is that the models are more expensive to operate compared to the standard transformer-based architecture.
Eligible developers can also use the new AI models with a rate limit of 20 requests per minute (RPM). However, developers will not be able to use this for function calls, streaming, system message support, and more. Additionally, ChatGPT Enterprise and Edu users will get access to the models next week.
Users on the free version of ChatGPT will soon have access to the o1-mini AI model, but this model is also expected to have a lower speed limit than GPT-4o.