A New Line AI Models With “Reasoning” Capabilities Released By OpenAI

OpenAI announced on Thursday the launch of its highly-anticipated “Strawberry” series of AI models, which have been specifically crafted to allocate additional time for in-depth processing of queries, enabling them to tackle complex problems more effectively. 

Initially unveiled by Reuters, these advanced models boast the ability to navigate intricate challenges across science, coding, and mathematics, surpassing the problem-solving capabilities of their predecessors. 

Through a recent blog post, the AI company detailed the models’ exceptional aptitude for critical reasoning, indicating a significant leap forward in the realm of artificial intelligence innovation.

OpenAI used the code name Strawberry to refer to the project internally

Internally, OpenAI referred to the project as Strawberry, and called the models that were revealed on Thursday o1 and o1-mini. According to the corporation, the o1 will be accessible through ChatGPT and its API on Thursday.

The models were the same as the Strawberry project, according to a post made by OpenAI researcher Noam Brown on social media platform X. Brown’s work focuses on enhancing reasoning in the company’s models.

Brown wrote “I’m excited to share with you all the fruit of our effort at OpenAI to create AI models capable of truly general reasoning,”

OpenAI’s o1 model achieved an 83% score on the International Mathematics Olympiad qualifying exam, surpassing the 13% score of its previous model, GPT-4o, and outperforming human PhD-level accuracy on a benchmark of science problems.

“We trained these models to spend more time thinking through problems”

According to Brown, the models achieved the scores by using a method called “chain-of-thought” reasoning, which divides difficult issues into more manageable logical steps. Researchers have found that when the method is applied as a prompting strategy, AI model performance on challenging problems typically improves. 

Now that OpenAI has automated this feature, models are able to solve problems without assistance from the user.

OpenAI said “We trained these models to spend more time thinking through problems before they respond, much like a person would. Through training, they learn to refine their thinking process, try different strategies, and recognize their mistakes,”

In November 2023, OpenAI’s progress on the reasoning project—then known as Q*—was first covered by Reuters. In July, it was revealed that the project had adopted the name Strawberry.

Follow Wat-Not on Facebook, Twitter, and Instagram