OpenAI is working on a groundbreaking project code-named "Strawberry" aimed at dramatically enhancing the reasoning capabilities of its AI models. According to an internal document reviewed by Reuters and its sources, the project seeks to enable AI not only to generate answers but also to autonomously navigate the internet and conduct deep research.
Strawberry represents a strategic effort by OpenAI to address some of the most significant challenges in artificial intelligence. Current AI models, despite their ability to summarise texts and compose prose, often struggle with common-sense problems and logical fallacies. By incorporating advanced reasoning capabilities, OpenAI hopes to overcome these limitations, potentially paving the way for AI to make scientific discoveries and develop new software applications.
The project, which was formerly known as Q*, is described as utilising a specialised post-training process to refine AI models after their initial training on large datasets. This method, similar to Stanford's "Self-Taught Reasoner" technique, involves the models creating their own training data to iteratively boost their intelligence. While the specifics of how Strawberry operates remain closely guarded within OpenAI, the internal documentation highlights its focus on long-horizon tasks, requiring extensive planning and execution.
OpenAI's internal communications suggest that the company is nearing a public release of this technology, which it claims will bring significantly more advanced reasoning capabilities to its AI models. The company aims to use Strawberry to perform complex research autonomously, with the help of a computer-using agent capable of taking actions based on its findings. The document also indicates plans to test the models on tasks typically handled by software and machine learning engineers, as per Reuters.
OpenAI CEO Sam Altman has previously highlighted reasoning ability as a crucial area for AI progress. As the competition among tech giants like Google, Meta and Microsoft intensifies, the race to develop AI with human-like reasoning capabilities continues to accelerate. However, researchers remain divided on whether current large language models can incorporate the necessary elements for long-term planning and predictive reasoning.
(Inputs from Reuters)