OpenAI CEO Sam Altman has hinted at a new initiative, nicknamed "Project Strawberry," that could improve the reasoning powers of the company's artificial intelligence models. Altman posted a mysterious photo of strawberries on X with the statement, "I love summer in the garden," raising curiosity. According to a Reuters story, Project Strawberry, also known as Q*, could represent a significant step towards artificial general intelligence (AGI).
Project Strawberry is described as an innovative strategy aimed at improving AI models' ability to do long-horizon tasks (LHT) and conduct autonomous research. The research includes a new "post-training" step that improves base AI models' performance in reasoning tasks, comparable to Stanford's "Self-Taught Reasoner" (STaR) approach. This upgrade may solve present AI constraints, such as common-sense reasoning and logical consistency, which are frequently stumbling blocks for existing models.
According to Reuters, OpenAI is developing Q*, which has the capacity to surf the internet autonomously, allowing the AI to undertake in-depth study and planning. This approach could help overcome obstacles in AI's grasp of complex, real-world events. Jimmy Apples, an AI insider, said that Q* has not yet been introduced because to continuous optimisations for latency and other technical concerns.
The initiative is also expected to include a specialised "deep-research" dataset for training and testing these advanced models. While the specifics of this dataset are unknown, it is predicted to dramatically improve the AI's ability to collect and interpret data autonomously. OpenAI is also putting Strawberry to the test in jobs that are traditionally handled by software and machine learning specialists, emphasising the project's breadth.
OpenAI recently introduced a five-level classification system to track progress towards AGI, with the business now ranked as Level 1. The invention of Q*, which combines features of Q-learning and A* search algorithms, could be critical in moving to the next level. Q-learning, a key component of reinforcement learning, seeks to optimise decision-making by identifying the appropriate actions to maximise cumulative rewards. The addition of A* search to this framework improves the AI's goal-directed reasoning and problem-solving abilities.
As conjecture about Project Strawberry spreads, OpenAI continues to push the boundaries of AI research.
