OpenAI, the company behind the widely known ChatGPT, is developing a new approach to artificial intelligence (AI) through a project named “Strawberry.” This project, which has not been publicly detailed before, aims to significantly enhance the reasoning capabilities of AI models, a crucial area where current models often fall short.
The project “Strawberry” has been in development for some time, as revealed by internal OpenAI documents reviewed by Reuters in May. However, the exact date of these documents remains unclear, and details about how Strawberry works are closely guarded even within the company. The project’s goal is to enable AI not only to generate responses to queries but also to plan and navigate the internet autonomously for deep research. This capability has eluded AI models so far.
An OpenAI spokesperson commented on the project, emphasizing the company’s ongoing efforts to improve AI’s ability to understand and interact with the world in a human-like manner.
Previously known as “Q*,” Strawberry has already been viewed internally as a significant breakthrough. Earlier demos of Q* showcased AI models capable of answering complex science and math questions that current commercially available models struggle with. There are reports of AI models scoring over 90% on a challenging MATH dataset, although it’s not confirmed if this is directly related to Strawberry.
OpenAI demonstrated a research project with new human-like reasoning skills during an internal meeting. However, details of this demonstration were not disclosed, and it’s unclear if it was related to Strawberry.
Strawberry aims to dramatically improve AI models’ reasoning capabilities. This involves a specialized post-training process where the AI models are fine-tuned after initial training on large datasets. This process, similar to a method developed at Stanford called “Self-Taught Reasoner” or “STaR,” allows AI to iteratively create its training data, potentially achieving higher levels of intelligence.
Strawberry’s objectives include performing long-horizon tasks (LHT), which require planning and executing a series of actions over an extended period. To achieve this, OpenAI is developing, training, and evaluating models on a “deep-research” dataset. Although the contents of this dataset and the specifics of the tasks remain undisclosed, the company intends for its AI models to autonomously conduct research and perform functions like those of software and machine learning engineers.
As OpenAI continues to push the boundaries of AI capabilities, it faces competition from other tech giants like Google, Meta, and Microsoft, all of which are exploring ways to enhance AI reasoning. The debate within the AI research community continues about whether large language models can achieve human-like reasoning. While some researchers are optimistic, others, like Yann LeCun from Meta, remain skeptical about the current capabilities of these models.
OpenAI’s Strawberry project represents a significant step towards addressing the challenges of AI reasoning. It aims to enable models that can reliably perform complex, multi-step tasks. As the company potentially prepares to release this technology, it signals to developers and stakeholders the promise of AI models with advanced reasoning capabilities, marking a crucial development in artificial intelligence.
