OpenAI’s Q* Becomes Strawberry Project, Aiming to Advance AI Reasoning

OpenAI is embarking on a notable AI research venture known internally as “Strawberry.” According to insights from sources and documents examined by Reuters, this initiative targets major advancements in AI’s reasoning abilities, an area where current models often fall short. Limited Information on Strawberry The specifics surrounding Strawberry are tightly held within OpenAI. Most team members have limited access to detailed project data. Originally named Q*, this project has been in the pipeline for a while, with documents indicating plans for significant research applications. However, the timing for a public rollout remains uncertain. In November, we wrote about the company’s Q* model, a research project within the company. What seems to set Q* apart from GPT and potentially any other AI model is the level of self-awareness it is reportedly exhibiting. Last week it emerged that OpenAI classifies its AI General Intelligence into five tiers.  Advancing AI Reasoning Capabilities The main focus of Strawberry is to significantly upgrade AI models’ ability to handle complex reasoning tasks. In contrast to existing AI, which can produce responses, this project aspires to allow AI to independently explore the web and conduct comprehensive research. Success in this area would represent a considerable leap forward, enabling AI to plan and solve intricate problems. Earlier this year, OpenAI showcased Q*’s prowess in tackling tough science and math queries. These demonstrations hinted that the new technology could address challenges beyond the capabilities of today’s commercial models. Furthermore, internal tests revealed that an AI system, potentially linked to Strawberry, achieved over 90% on a challenging MATH dataset. Industry Context and Challenges Enhancing AI’s reasoning ability remains a top priority for many, including researchers at companies like Google, Meta, and Microsoft. Although large language models are proficient in summarizing text and writing coherent sentences, they often falter in logical and common sense reasoning. It is widely believed that improving these skills is essential for AI to reach human or even super-human intelligence. Strawberry employs a particular post-training method designed to adapt pre-trained AI models for better performance. This methodology resembles Stanford’s “Self-Taught Reasoner” (STaR) approach, which enables AI to generate its own training data iteratively. According to OpenAI’s documentation, Strawberry will be evaluated on tasks requiring extended planning and execution. Autonomous Research Capabilities A prominent objective of Strawberry is to enable AI models to perform autonomous research by navigating the web using a computer-using agent (CUA). This functionality could extend to tasks currently managed by software engineers. Internal communications suggest OpenAI is close to launching such technology with advanced reasoning features. If successful, this project could lead to AI systems that make significant scientific discoveries and create new software solutions.