OpenAI has introduced its latest innovations, the 'Strawberry' bots, named o1 and o1-mini. These models are designed to excel in reasoning and solving intricate problems in science, coding, and mathematics. Impressively, the o1 model achieved an 83% score on the International Mathematics Olympiad qualifying exam and surpassed human PhD-level accuracy on scientific tests. These advancements will be accessible through ChatGPT starting Thursday.
The 'Strawberry' bots employ 'chain-of-thought' reasoning, a method that allows them to think through problems more methodically before providing answers. This approach marks a significant improvement over previous models, such as GPT-4o, which only managed a 13% score on the same math exam. OpenAI's blog post elaborates on how these models have been trained to refine their abilities by learning from mistakes.
In addition to their impressive performance in STEM tasks, the o1-mini offers a cost-effective alternative while maintaining high proficiency. The development of these models involved extensive reinforcement learning, and they are expected to set a new benchmark in AI reasoning. Reuters first reported on this project, highlighting the potential of these models to close the gap between human and artificial intelligence.
