OpenAI, the creator of ChatGPT, is working on a new reasoning technology under the code name ‘Strawberry,’ according to a report by Reuters. This initiative aims to address the limitations of current AI models, which often struggle with tasks requiring advanced reasoning and common-sense problem-solving.
Project Goals and Capabilities
The internal document outlining this project, which has not been publicly disclosed, describes Strawberry as a tool designed to enable OpenAI’s AI to perform deep research by autonomously navigating the internet. This capability is seen as a significant step in improving the advanced reasoning functions of AI models.
Secretive Development and Demonstrations
Details about Strawberry’s specific workings remain closely guarded within OpenAI. Internally, they have praised the project, previously known as Q*, as a breakthrough. Demonstrations earlier this year showed the model answering complex science and math questions that are currently beyond the reach of commercially available AI.
Human-Like Reasoning Skills
During an internal meeting last week, OpenAI showcased a research project featuring new human-like reasoning skills, as reported by Bloomberg. Strawberry’s technology involves a specialised processing method applied to an AI model after pre-training on vast datasets.
Addressing Common AI Challenges
While existing large language models can summarize texts and generate prose, they often fail at intuitively easy tasks for humans, such as identifying logical fallacies or playing games like tic-tac-toe. Strawberry aims to improve these capabilities, which are crucial for applications ranging from scientific breakthroughs to software development.
Focus on Long-Horizon Tasks (LHT)
The project emphasizes enabling long-horizon tasks (LHT), requiring planning and executing a series of actions over time. OpenAI is developing and testing these models on a “deep-research” dataset, the specifics of which remain undisclosed.
In addition to conducting web-based research, OpenAI aims to leverage Strawberry’s capabilities to perform tasks typically done by software and machine learning engineers. The company is exploring the use of a computer-using agent to act on the AI’s findings, further pushing the boundaries of autonomous AI technology.
See also: NATO Releases Revised AI Strategy To Combat Emerging Threats
