OpenAI, the company that created ChatGPT, is working on a new project called "Strawberry" intended to significantly improve the reasoning capabilities of artificial intelligence. This project is supported by Microsoft and aims to make AI much more capable in planning and solving complex problems. "Strawberry" isn't just limited to generating answers, but aims to allow AI to autonomously navigate the Internet to perform "deep research." This would allow the models to deal with complicated problems that require long-term planning, an ability that current models lack. The project is a continuation of an earlier project called "Q*", which has already demonstrated advanced capabilities in answering scientific and mathematical questions.
The "post-training" method used in "Strawberry" is similar to the "Self-Taught Reasoner" (STaR) technique developed at Stanford University, one of the most prestigious universities in the world. This allows the AI to continuously improve its performance by creating its own training data. This process could help models achieve higher-than-human intelligence. However, creating an AI with advanced reasoning capabilities and autonomy also raises serious questions about safety and control. If these models get out of control, they can pose a major risk to humanity, with the potential to make decisions without human oversight and cause unpredictable consequences. It's an impressive achievement, given that current AI models have failed to reach this level of reasoning
What is Project Strawberry?
"Strawberry" is an extension of an earlier OpenAI project known as Q*, which has demonstrated impressive capabilities in solving math and science problems. This suggests that the models developed in this project can outperform current commercial AI, paving the way to artificial general intelligence (AGI).
The project involves refining existing language models through advanced machine learning techniques. A similar method was described in a 2022 Stanford study that uses "chains of thought" to improve reasoning. This method allows the AI model to generate and train on its own reasoning data, leading to significantly improved performance.
Capabilities and Objectives
Planning and Complex Reasoning: Models will be able to develop detailed plans and solve problems that require multiple steps to reach a solution.
Autonomous Internet Browsing: Models will be able to search and analyze information on the Internet autonomously, enhancing their research capabilities.
Deep research: Models will be able to perform detailed and comprehensive research, similar to how a human researcher would do.
Progress and Challenges
Currently, the specific details of how OpenAI achieved these improvements are being kept under wraps. "Strawberry" is known to involve refining existing models through continuous learning techniques, possibly using self-generated data to allow the model to self-improve. This concept, known as "recursive self-improvement", is considered a major goal in AI research.
However, it is important to note that information about "Strawberry" is limited and many of the details of the project remain unconfirmed. The public and the scientific community are eagerly waiting to see if these advances will lead to significant achievements in the field of artificial intelligence.
If OpenAI succeeds in achieving the goals of the "Strawberry" project, it could represent an important step toward creating AI models capable of complex reasoning and deep research. This could transform the way we use artificial intelligence, opening up new opportunities in various fields, from the exact sciences to academic research and industry. In conclusion, "Strawberry" could be the next big step in the evolution of AI, and the global community is eagerly waiting to see the results of this ambitious project.