OpenAI developing new reasoning AI called ‘Strawberry,’ aiming for breakthrough in human-level problem solving

OpenAI, the renowned AI company backed by Microsoft, is covertly developing a cutting-edge AI project named “Strawberry,” which aims to revolutionize the field with advanced reasoning capabilities and human-level problem solving.

Short Summary:

• OpenAI’s new project “Strawberry” focuses on enhancing AI’s reasoning skills.
• Strawberry builds on the previous project named Q* and aims for autonomous, deep research capabilities.
• OpenAI seeks to tackle long-horizon tasks and potentially achieve super-human intelligence through advanced AI training techniques.

Introduction

In an ambitious effort to push the boundaries of artificial intelligence, OpenAI is currently working on a new AI model under the code name “Strawberry.” Supported by Microsoft, this project aims to improve the reasoning abilities of AI, enabling it to autonomously perform complex tasks requiring long-term planning and “deep research.” The information, revealed by Reuters through internal documents, paints a picture of a highly guarded and groundbreaking initiative within OpenAI.

The Genesis of Strawberry

According to sources, the Strawberry project evolved from a preceding initiative known as Q*, which was already perceived as a significant breakthrough within OpenAI. Q* demonstrated impressive capabilities in tackling complex science and math problems, surpassing the abilities of existing commercially available models. This progression signifies OpenAI’s continuous efforts to refine and enhance its artificial intelligence systems.

“We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time,” an OpenAI spokesperson said in a statement to Reuters.

Strawberry’s Aims and Methodology

The core goal of Strawberry is to enable AI models to perform “deep research” autonomously, navigating the internet and planning ahead effectively. Unlike current AI models that generate responses based on existing data, Strawberry aspires to develop reasoning skills that allow it to tackle multi-step problems and long-term tasks without human intervention.

This endeavor bears similarity to the “Self-Taught Reasoner” (STaR) technique developed at Stanford University, which allows AI models to iteratively generate their own training data, thereby bootstrapping their intelligence levels. Noah Goodman, a Stanford professor and one of the creators of STaR, expressed both excitement and caution about the direction AI is heading:

“I think that is both exciting and terrifying…if things keep going in that direction, we have some serious things to think about as humans,” Goodman told Reuters.

Enhancing AI Reasoning

Central to Strawberry’s strategy is the deployment of specialized post-training processes. These methods involve refining the AI models after their initial training on vast datasets. The intent is to hone AI’s performance in specific areas, improving its ability to reason and solve complex, multi-step problems effectively. This is crucial for achieving human or even super-human intelligence, an aspiration that has significant implications for the future of AI.

“Reasoning is key to AI achieving human or super-human-level intelligence,” a researcher interviewed by Reuters mentioned.

Sam Altman, the CEO of OpenAI, has underscored the importance of advancing AI reasoning capabilities, stating:

“The most important areas of progress will be around reasoning ability,” Altman said earlier this year.

Long-Horizon Tasks and Autonomous Web Research

One of the ambitious goals of Strawberry is to enable the AI to perform long-horizon tasks (LHT), which require planning and executing actions over extended periods. This involves training and evaluating AI models on a “deep-research” dataset designed to mimic complex real-world tasks. The specifics of the dataset and the duration of these tasks remain confidential, but the overarching goal is to develop AI that can autonomously conduct web research and execute tasks based on its findings.

To facilitate this, OpenAI is considering the deployment of a “computer-using agent” (CUA) that can make decisions and act independently based on the gathered information. This marks a significant step towards creating an AI capable of long-term autonomous functionality.

“How Strawberry works is a tightly kept secret even within OpenAI,” a source familiar with the matter revealed to Reuters.

Comparisons with Other AI Efforts

OpenAI is not alone in its quest to enhance AI reasoning. Major tech companies such as Google, Meta, and Microsoft are also exploring various techniques to improve the reasoning capabilities of AI models. However, there is an ongoing debate within the AI community about the feasibility of incorporating long-term planning and reasoning into large language models (LLMs).

Yann LeCun, a prominent AI researcher at Meta, has often expressed skepticism regarding the ability of LLMs to achieve human-like reasoning. His perspective highlights the varied opinions and challenges that exist within the field.

Despite these differing viewpoints, the development of Strawberry represents a critical part of OpenAI’s strategy to address the limitations of current AI models and potentially redefine what AI can accomplish.

Post-Training and Human Feedback

Strawberry’s development involves a series of post-training techniques, including fine-tuning through human feedback and iterative learning processes. These methodologies are designed to refine AI models and significantly enhance their performance on specific tasks, pushing the boundaries of what AI can achieve. This approach aligns with OpenAI’s broader goal of creating systems that can think and reason more like humans.

Future Implications and the Road Ahead

The advancements brought by Strawberry could herald a new era of intelligent, autonomous AI systems capable of performing tasks that were once thought to be exclusively within the realm of human intelligence. The potential applications are vast, ranging from scientific discoveries to the development of sophisticated software tools. While the path forward is filled with challenges, the potential rewards are immense.

“We want our AI models to see and understand the world more like we do. If Strawberry succeeds, it could bring us one step closer to realizing that vision,” an OpenAI spokesperson remarked optimistically.

OpenAI’s Five-Tier AI System

To track its progress towards achieving Artificial General Intelligence (AGI), OpenAI has introduced a five-tier system. This framework ranges from Level 1, representing current conversational AI, to Level 5, which envisions AI capable of managing and performing the work of an entire organization. As per the latest updates, OpenAI believes it is nearing Level 2, which involves problem-solving capabilities akin to a PhD holder without the use of external tools. This systematic approach helps in understanding and developing AI systems that could eventually surpass human intelligence.

The Industry’s Competitive Landscape

In addition to internal endeavors, OpenAI is navigating a competitive landscape where multiple tech giants are vying to enhance AI reasoning. At the same time, OpenAI faces regulatory scrutiny, prompting strategic decisions such as Microsoft and Apple withdrawing their board seats at OpenAI. Despite these challenges, OpenAI remains steadfast in its mission to push the frontiers of AI technology.

The collaboration with Los Alamos National Laboratory adds another intriguing layer to OpenAI’s endeavors. Together, they are exploring the potential applications of AI in bioscience research, reflecting the broadening scope of AI’s impact across various fields.

Conclusion

OpenAI’s Strawberry project represents a significant leap forward in the quest for advanced AI reasoning and autonomous problem-solving capabilities. By building on the foundation laid by the Q* project, Strawberry aspires to enable AI systems to navigate the complexities of the real world with human-like reasoning and planning skills. As the project progresses, the potential implications for science, technology, and society are profound, marking a pivotal moment in the evolution of artificial intelligence.