On Thursday, OpenAI announced the launch of its "Strawberry" series of AI models, aimed at advancing the ability of artificial intelligence to reason through complex problems and deliver more accurate solutions. The models, codenamed Strawberry during development, are designated as o1 and o1-mini. The o1 model will be integrated into ChatGPT and its API starting today.
The launch of these new models represents a significant leap in AI problem-solving capabilities. According to OpenAI's blog post, the o1 model achieved an impressive 83% score on the qualifying exam for the International Mathematics Olympiad, a notable improvement from the 13% score of its predecessor, GPT-4o. Additionally, the o1 model demonstrated enhanced performance in competitive programming and surpassed human PhD-level accuracy on a benchmark of science problems.
Noam Brown, an OpenAI researcher focused on reasoning improvements, confirmed on social media that the models are indeed the same as the internally codenamed Strawberry project. He expressed enthusiasm about the models' general reasoning capabilities, highlighting their potential to solve more challenging problems.
The advancement in the Strawberry models is attributed to their use of "chain-of-thought" reasoning. This technique involves breaking down complex problems into smaller, logical steps. OpenAI has automated this process in the new models, allowing them to independently decompose problems and refine their solutions without user prompting.
OpenAI’s blog post elaborated on how the models were trained to emulate human-like problem-solving approaches. The models are designed to spend more time processing and refining their answers, learning from mistakes, and exploring different strategies before responding. This enhanced reasoning process helps improve the models' accuracy and effectiveness in handling difficult tasks.
The Strawberry project, initially referred to as Q*, was first reported by Reuters in November 2023. The project's development and subsequent renaming to Strawberry reflect OpenAI’s commitment to pushing the boundaries of AI reasoning and problem-solving capabilities.
