Strawberry, a closely guarded project within OpenAI, aims to enable AI to autonomously navigate the internet and perform deep research, a capability that has so far eluded existing AI models.(Source: Image buy RR)

New Initiative to Enable Autonomous Deep Research and Planning

OpenAI is developing a new project, code-named “Strawberry,” aimed at enhancing the reasoning capabilities of its AI models, according to internal documents and a source familiar with the matter. This initiative seeks to improve the ability of AI to plan ahead and conduct deep research autonomously, a challenge that has previously eluded AI models. The Strawberry project, which remains a closely guarded secret within OpenAI, involves specialized post-training techniques to adapt pre-trained models for advanced reasoning tasks.

The project builds on previous advancements, such as the Q* initiative, which demonstrated the potential of AI to answer complex science and math questions. A recent internal meeting showcased a research project with new human-like reasoning skills, though it’s unclear if this was Strawberry. As noted in reuters.com, OpenAI’s approach includes creating and evaluating models on a “deep-research” dataset, with the goal of enabling the AI to browse the web autonomously and conduct long-horizon tasks, such as planning and performing a series of actions over time.

Strawberry aims to address the current limitations of AI in handling common-sense problems and logical reasoning, areas where models often produce inaccurate or nonsensical results. By leveraging a method similar to Stanford’s “Self-Taught Reasoner,” OpenAI hopes to bootstrap its AI models to higher levels of intelligence, potentially surpassing human reasoning capabilities. This technique involves iteratively creating training data to enhance the AI’s performance.

OpenAI’s efforts are part of a broader industry push to improve AI reasoning, with companies like Google, Meta, and Microsoft also exploring similar advancements. While there is skepticism about whether large language models can achieve human-like reasoning, OpenAI is optimistic that projects like Strawberry will pave the way for significant breakthroughs in AI capabilities, enabling applications from scientific discoveries to advanced software development.

read more at reuters.com