OpenAI is actively developing a new approach to enhance its artificial intelligence models under the project codename “Strawberry,” according to sources familiar with the matter and internal documents reviewed by Reuters.
The project, previously unreported, underscores OpenAI’s urgency to demonstrate advanced reasoning capabilities in the models it offers. Teams within OpenAI are currently engaged in the Strawberry project, as outlined in recent internal documents seen by Reuters in May. The details of these plans indicate ongoing research efforts, although the exact timeline for public availability of Strawberry remains uncertain.
The specifics of how Strawberry operates are closely guarded within OpenAI, emphasizing its innovative nature even within the company. Described as enabling AI to not only answer queries but also to autonomously navigate the internet for in-depth research, Strawberry aims to tackle challenges that existing AI models have struggled with, such as planning and reliable multi-step problem-solving.
OpenAI’s ambitions with Strawberry include enhancing AI models’ reasoning capabilities significantly. This involves specialized methods applied after initial training on extensive datasets, a process known as “post-training.” The company is also reportedly exploring techniques similar to Stanford’s Self-Taught Reasoner (STaR), which allows models to iteratively enhance their intelligence levels.
As OpenAI prepares to potentially unveil technology with advanced reasoning capabilities, it is actively engaging with developers and stakeholders outside the company. Strawberry’s development includes plans to handle long-horizon tasks and potentially autonomously perform complex actions over extended periods, leveraging what OpenAI terms a “deep-research” dataset.
The ultimate goal for OpenAI is to equip its models with capabilities that extend beyond current AI limitations, potentially revolutionizing fields from scientific research to software engineering.
