OpenAI mentioned the o1 mannequin scored 83% on the qualifying examination for the Worldwide Arithmetic Olympiad, in contrast with 13% for its earlier mannequin, GPT-4o [File]
| Photograph Credit score: AP
Microsoft-backed OpenAI mentioned on Thursday it was launching its “Strawberry” series of AI models designed to spend more time processing answers to queries in an effort to remedy onerous issues.
The fashions, first reported by Reuters, are able to reasoning by means of complicated duties and may remedy tougher issues than earlier fashions in science, coding and math, the AI agency mentioned in a weblog publish.
OpenAI used the code title Strawberry to check with the challenge internally, whereas it dubbed the fashions introduced on Thursday o1 and o1-mini. The o1 will probably be obtainable in ChatGPT and its API beginning Thursday, the corporate mentioned.
Noam Brown, a researcher at OpenAI targeted on bettering reasoning within the firm’s fashions, confirmed in a publish on social media platform X that the fashions have been the identical because the Strawberry challenge.
“I am excited to share with you all of the fruit of our effort at OpenAI to create AI fashions able to really basic reasoning,” Brown wrote.
In its weblog publish, OpenAI mentioned 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 corporate mentioned.
Brown mentioned the fashions have been in a position to accomplish the scores by incorporating a method referred to as “chain-of-thought” reasoning, which includes breaking down complicated issues into smaller logical steps.
Researchers have famous that AI mannequin efficiency on complicated issues tends to enhance when the method has been used as a prompting method. OpenAI has now automated this functionality so the fashions can break down issues on their very own, with out person prompting.
“We skilled these fashions to spend extra time pondering by means of issues earlier than they reply, very similar to an individual would. By means of coaching, they be taught to refine their pondering course of, strive totally different methods, and acknowledge their errors,” OpenAI mentioned.
Reuters was the primary to report OpenAI’s work on the reasoning challenge, then known as Q*, in November 2023. It reported in July that the challenge had come to be referred to as Strawberry.
Printed – September 13, 2024 08:11 am IST
