OpenAI, the company behind ChatGPT, is developing a new project called "Strawberry" to enhance the reasoning abilities of its AI models. This information comes from an insider and documents reviewed by Reuters. The project is still in progress and its launch date remains unclear.

Strawberry is designed to help OpenAI’s models not just answer questions, but also autonomously navigate the internet to conduct what OpenAI calls "deep research." This level of advanced reasoning is something current AI models struggle with.

According to an internal OpenAI document from May, the company aims for Strawberry to significantly improve its AI’s ability to understand and plan tasks, much like humans. However, the specifics of how Strawberry works are kept confidential, even within the company.

A spokesperson from OpenAI confirmed that continuous research is key to improving AI, but did not provide specific details about Strawberry.

Previously known as Q*, Strawberry has shown promise in internal demos by answering complex science and math questions that current models can't handle. During a recent company meeting, OpenAI demonstrated a project with new reasoning skills, though it's unclear if this was Strawberry.

Strawberry involves a special method of processing AI models after they've been initially trained on large datasets. This post-training, similar to a technique called "Self-Taught Reasoner" (STaR) developed at Stanford, allows models to improve their intelligence by generating their own training data.

One of Strawberry’s goals is to handle long-term tasks that require planning and multiple steps. OpenAI is testing these capabilities with a "deep-research" dataset and aims to have its models perform tasks autonomously on the internet, guided by a "computer-using agent."

This project is part of OpenAI's broader effort to advance AI reasoning, a critical step toward achieving human or super-human level intelligence. Other companies like Google, Meta, and Microsoft are also working on similar advancements.