OpenAI is introducing a new artificial intelligence model named “Strawberry” that is capable of performing human-like reasoning tasks in order to remain competitive in the crowded market of AI rivals. This new model, referred to as o1, is specifically designed to take more time to compute answers before responding to user queries. With this model, OpenAI’s tools are expected to tackle complex multi-step problems, such as intricate math and coding questions.
While this early version of the model may not have all the features of the ChatGPT, like web browsing and file uploading capabilities, it represents a significant advancement in AI capability for complex reasoning tasks. Therefore, OpenAI is restarting the series and naming it OpenAI o1. A preview of this model will be accessible to paid Plus and Team users through OpenAI’s ChatGPT chatbot starting this Thursday.
The release of this model comes at a time when OpenAI is seeking substantial funding and facing increased competition in the race to develop more advanced AI systems. Competitors like Anthropic and Google have also showcased similar “reasoning” capabilities in their AI models. OpenAI’s blog post provided examples of the AI model’s responses to questions on various subjects, like coding, English, and math, including solving a simple crossword puzzle. Noam Brown, a research scientist at OpenAI, mentioned releasing the model in preview to gather feedback on its usage and areas for improvement.
The updated AI system from OpenAI will introduce a slight change in user experience compared to ChatGPT. Before responding to prompts, the software will pause briefly to consider related prompts and generate the best response using a “chain of thought” prompting technique. OpenAI has been focusing on enhancing AI systems’ abilities to handle multi-step actions, as showcased in their previous efforts to improve math problem-solving capabilities. This entails training the model to reward each correct step towards solving a problem, rather than solely focusing on generating an accurate final answer.
