Photo: Camilo Concha/Shutterstock.com
OpenAI, the company behind the popular ChatGPT, is reportedly working on a new AI initiative called Strawberry, which aims to address the issue of reasoning in AI models.
First reported by Reuters, the current generation of AI models often struggle with tasks that require common-sense reasoning or long-term planning. However, they offer quick solutions to tasks such as generating text based on a prompt, summarising the information provided by the user, or composing coherent responses.
OpenAI’s Strawberry project is designed to overcome these limitations, enabling its AI models to navigate complex, multi-step problems autonomously and conduct what the company terms “deep research.”
Internal documents indicate that Strawberry’s main objective is to enable AI to go beyond merely answering questions. Instead, it will be capable of planning and executing tasks that require continuous engagement, such as browsing the Internet to gather information and solve long-horizon tasks (LHT) — activities that demand foresight and extended action.
Despite the intrigue surrounding Strawberry, many details remain closely guarded. According to sources cited by Reuters, even within OpenAI, much of the project’s inner workings are a mystery. The document outlines OpenAI’s ambitions but is vague on specific methodologies and timelines.
Notably, Strawberry builds on a previous initiative, codenamed Q*, already viewed as a significant breakthrough within OpenAI.
Two anonymous sources revealed that early demonstrations of Q* showed it could tackle difficult math and science problems that other commercially available models could not handle.
Additionally, a separate source indicated that OpenAI’s internal tests of the AI scored over 90% on the MATH dataset, a benchmark for advanced mathematical problem-solving. It remains unclear whether this success is directly tied to Strawberry.
One of Strawberry’s unique aspects is its post-training phase. Sources familiar with the project suggest that OpenAI uses a specialised ” post-training ” approach to enhance the model’s performance. This process involves fine-tuning the AI after its initial training on massive datasets.
While many researchers affirm that upgrading the AI’s reasoning capabilities is the first step in unlocking human-like intelligence in the AIs, some have expressed scepticism.
While companies like Meta are already working on developing artificial general intelligence (AGI), a hypothetical human-like AI model, OpenAI also seems to take a step in that direction.
In the News: Meta is scraping Australia’s data including public photos of children
