July 12, OpenAI
OpenAI is working on a new project called “Strawberry.” This project aims to improve the reasoning abilities of their AI models. Details about Strawberry are kept secret. The goal is for AI to not just answer questions but to plan and navigate the internet on its own. This is a big challenge for AI, which hasn’t been done before. A spokesperson said they want AI to see and understand the world more like humans do.
Strawberry was previously called Q*. It showed promising results in early demos, answering tough science and math questions. At a recent meeting, OpenAI demonstrated new AI reasoning skills. They hope Strawberry will greatly improve their AI’s reasoning. The project uses a special post-training process after the AI is trained on large datasets. This process helps the AI plan ahead and solve complex problems. Improving AI reasoning is crucial for advanced tasks like scientific discoveries and software development. Other companies like Google, Meta, and Microsoft are also working on improving AI reasoning.
Strawberry is similar to a method from Stanford called “Self-Taught Reasoner” (STaR). STaR helps AI improve itself by creating its own training data. OpenAI wants Strawberry to perform long-horizon tasks, which require planning and multiple steps. They are testing the AI on a special “deep-research” dataset. The AI will also be able to browse the web and perform tasks like a software engineer.
OpenAI hopes Strawberry will unlock new capabilities for AI models.