OpenAI launches new sequence of AI fashions with 'reasoning' talents

OpenAI launches new sequence of AI fashions with 'reasoning' talents

(Reuters) -Microsoft-backed OpenAI stated on Thursday it was launching its "Strawberry" sequence of AI fashions designed to spend extra time processing solutions to queries so as to clear up arduous issues.

The fashions, first reported by Reuters, are able to reasoning via complicated duties and might clear up more difficult issues than earlier fashions in science, coding and math, the AI agency stated in a weblog submit.

OpenAI used the code identify Strawberry to discuss with the venture internally, whereas it dubbed the fashions introduced on Thursday o1 and o1-mini. The o1 shall be out there in ChatGPT and its API beginning Thursday, the corporate stated.

Noam Brown, a researcher at OpenAI centered on enhancing reasoning within the firm's fashions, confirmed in a post on social media platform X that the fashions have been identical because of the Strawberry venture.

"I am excited to share with you all of the fruit of our effort at OpenAI to create AI fashions able to really normal reasoning," Brown wrote.

In its weblog submit, OpenAI stated the o1 mannequin scored 83% on the qualifying examination for the Worldwide Arithmetic Olympiad, in contrast with 13% for its earlier mannequin, GPT-4o.

The mannequin additionally improved efficiency on aggressive programming questions and exceeded human PhD-level accuracy on a benchmark of science issues, the corporation stated.

Brown stated the fashions have been in a position to accomplish the scores by incorporating a way often known as "chain of thought" reasoning, which includes breaking down complicated issues into smaller logical steps.

Researchers have known that AI mannequin efficiency on complicated issues tends to enhance when the strategy has been used as a prompting approach. OpenAI has now automated this functionality so the fashions can break down issues on their very own, with out person prompting.

"We educated these fashions to spend extra time considering via issues earlier than they reply, very like an individual would. By coaching, they will be taught to refine their considering course of, strive completely different methods, and acknowledge their errors," OpenAI stated.

Reuters was the primary to report OpenAI's work on the reasoning venture, then referred to as Q*, in November 2023. It reported in July that the venture had come to be often known as Strawberry.

(Reporting by Akash Sriram in Bengaluru, Katie Paul in New York and Anna Tong in San Francisco; Modifying by Alan Barona and Leslie Adler)