OpenAI developing new reasoning tech, codenamed ‘Strawberry’

OpenAI's Project "Strawberry" Aims to Revolutionize AI Reasoning

OpenAI, the company behind ChatGPT, is developing an innovative approach to artificial intelligence called "Strawberry." This project, not previously disclosed, signifies OpenAI's effort to significantly enhance the reasoning capabilities of its AI models. Internal documents reveal that Strawberry aims to enable AI to autonomously navigate the internet and perform what the company calls “deep research.” This level of functionality has been a significant challenge for existing AI models.

According to a recent internal document, Strawberry involves a specialized post-training process designed to refine AI models after their initial training on large datasets. Importantly, this method is geared towards improving the AI's ability to plan and reason more effectively, which is essential for tasks that require a series of well-thought-out steps.

AI researchers believe that improved reasoning could lead to monumental advances, such as deep scientific discoveries and the development of new software applications. Despite the impressive capabilities of current AI, such as summarizing complex texts and generating creative content, they often fail at tasks requiring common sense or logical consistency. This is where Strawberry aims to make a breakthrough.

Strawberry: Improving AI's Deep Research Abilities

The primary goal of Strawberry is to achieve advanced reasoning skills in AI, enabling it to handle tasks that require long-term planning and sophisticated decision-making. The project intends to push the boundaries of what AI can accomplish, going beyond simple query responses to conducting autonomous research online. This includes a computer-using agent (CUA) capable of taking action based on its findings, thereby introducing a layer of autonomous decision-making.

Although the specifics of how Strawberry functions remain a closely guarded secret within OpenAI, the project supposedly employs a method influenced by the "Self-Taught Reasoner" (STaR) developed at Stanford. This technique involves AI iteratively creating its training data to achieve higher levels of intelligence.

Breaking New Ground in AI Reasoning

Strawberry's objective is long-horizon tasks (LHT), which require extensive planning and execution over a prolonged period. OpenAI is training its models using a "deep-research" dataset, although the exact contents and the duration defined as "long-term" remain undisclosed. The project also aims to equip AI with the ability to perform tasks typically handled by software and machine learning engineers, further showcasing its potential versatility.

OpenAI's spokesperson emphasized the industry-wide consensus on the importance of continuous research to enhance AI reasoning capabilities. The company's ambition aligns with efforts from other major tech entities like Google, Meta, and Microsoft, all exploring ways to improve AI's cognitive functions.

Challenges and Future Implications

While Strawberry represents a significant stride towards achieving human-like reasoning in AI, experts remain divided on whether large language models can fully integrate such sophisticated reasoning and long-term planning. Yann LeCun of Meta, for instance, has expressed skepticism about LLMs achieving human-equivalent reasoning.

Nevertheless, the development of Strawberry is seen as a critical component of OpenAI's strategy to overcome these hurdles. As the company signals impending advancements, the AI community and industry at large brace for potential breakthroughs that could redefine the scope of artificial intelligence. Despite the excitement, the prospect also raises profound ethical and practical considerations for the future, as highlighted by experts like Stanford's Noah Goodman.

OpenAI's commitment to refining AI reasoning through Strawberry could usher in a new era of autonomous and highly intelligent AI, poised to tackle complex, multi-step challenges with unprecedented efficiency.