To maintain its leading position in the rapidly evolving field of artificial intelligence, OpenAI is secretly developing a new AI model codenamed "Strawberry."

This news comes from internal documents revealed by Reuters and an informed source. The Microsoft-backed startup, known for its ChatGPT product, is currently showcasing its model's advanced reasoning abilities, which could represent a significant leap forward in AI technology.

01. Inside Look at the "Strawberry" Project

According to the latest internal document seen by Reuters in May, the OpenAI team is deeply engaged in research on the "Strawberry" project. While the exact timeline of this document is unclear, it outlines OpenAI's plans to utilize "Strawberry" for advanced AI research.

The project is described as "ongoing" and has been kept under wraps even within the company. The goal of "Strawberry" is to enable AI not only to generate answers but also to autonomously and reliably navigate the internet, engaging in what OpenAI calls "deep research."

"This is something that AI models have not been able to achieve so far," the source noted, highlighting the project's ambition.

When asked about "Strawberry" and the details in this report, a spokesperson for OpenAI stated in a declaration, "We aspire for our AI models to perceive and understand the world as we do. Continuous research into new AI capabilities is a common practice in the industry, and we share the belief that over time, the reasoning abilities of these systems will continue to improve."

However, the spokesperson did not directly address questions about "Strawberry."

02. From Q to "Strawberry": A New Era of Reasoning

Sources indicate that "Strawberry" is the successor to a previous project named Q.

According to two informed sources, OpenAI internally viewed Q as a breakthrough due to its ability to answer complex scientific and mathematical questions, surpassing the capabilities of most currently commercialized models.

Bloomberg reported that during an internal all-hands meeting this year, OpenAI showcased a research project demonstrating new human-like reasoning abilities.

While Reuters could not confirm whether the showcased project was "Strawberry," it aligns with the company's ongoing efforts to enhance AI reasoning capabilities.

OpenAI CEO Sam Altman has emphasized the importance of reasoning in AI, stating earlier this year that "the most important areas of progress will be around reasoning."

03. Challenges in AI Reasoning

Researchers believe that enhancing AI models' reasoning abilities is crucial to achieving human- or superhuman-level intelligence. While large language models can efficiently summarize text and write articles, they often make mistakes on commonsense questions and logical tasks, leading to so-called "hallucinations" or generating incorrect information.

According to AI researchers, reasoning involves AI planning, understanding the physical world, and solving multi-step problems.

OpenAI's "Strawberry" project aims to overcome these challenges by adopting specialized post-training processes. This includes fine-tuning AI models after pre-training them on vast datasets.

According to an informed source, "Strawberry's" approach bears similarities to Stanford University's "Self-Taught Reasoning" (STaR), which allows AI models to iteratively create their own training data, potentially enabling them to reach higher levels of intelligence.

Noah Goodman, one of the creators of STaR and a professor at Stanford University, commented, "I think it's both exciting and scary... If things continue to move in this direction, we humans have some serious thinking to do."

04. Long-Term Task Planning and Autonomous Research

One of the ambitious goals of the "Strawberry" project is the ability to execute long-term tasks (LHTs), requiring AI to plan and execute a series of actions over an extended period.

Internal documents reveal that OpenAI is training and evaluating models on a "deep research" dataset to achieve these capabilities.

While the specific content and duration of this dataset remain undisclosed, the objective is clear: to enable AI to autonomously conduct research with the help of computer usage agents (CUAs) and take actions based on research findings.

05. A Competitive AI Landscape

OpenAI is not alone in its efforts to enhance AI reasoning capabilities. Major tech companies like Google, Meta, and Microsoft, along with numerous academic labs, are also exploring various technologies to improve AI reasoning.

However, opinions differ regarding whether large language models can incorporate long-term planning and advanced reasoning into their predictions. Yann LeCun, a pioneer in modern AI at Meta, has often expressed skepticism about the ability of large language models (LLMs) to achieve human-like reasoning.

"Strawberry" represents a crucial component of OpenAI's strategy aimed at addressing the limitations of current AI models. By developing more advanced reasoning abilities, OpenAI aims to unlock new possibilities for AI, ranging from scientific discoveries to creating new software applications.

Simultaneously, the company has been signaling to developers and partners that it will soon release technologies with significantly enhanced reasoning capabilities.

The development of "Strawberry" involves post-training methods such as fine-tuning, which incorporate human feedback and iterative learning processes. These techniques aim to refine AI models and improve their performance in specific tasks.

Advancements through "Strawberry" technology could redefine AI capabilities and set new standards for what these models can achieve.

While the path forward is fraught with challenges, the potential rewards are immense, heralding a new era of intelligent, autonomous AI systems.

In the words of the OpenAI spokesperson, "We aspire for our AI models to perceive and understand the world as we do. If the 'Strawberry' project succeeds, we will be one step closer to realizing this vision."

OpenAI has introduced a five-level system to track its progress towards achieving Artificial General Intelligence (AGI). These levels range from Level 1, representing current conversational AI, to Level 5, envisioning systems capable of managing and executing the work of entire organizations, encompassing different levels of AI capabilities.

Below are the five AI levels defined by OpenAI:

1. Chatbot: AI with conversational language capabilities

2. Reasoner: AI with human-level problem-solving abilities

3. Agent: A system capable of taking actions

4. Innovator: AI that can assist in invention and creation

5. Organizer: AI capable of completing organizational work

OpenAI believes that "Strawberry" is approaching Level 2, which involves problem-solving akin to a Ph.D. without tools. This framework aims to provide a structured approach to understanding and developing AI systems, ultimately surpassing human intelligence.

The Chinese content is compiled by the MetaverseHub team. Please contact us for reprint permission.