A bomb under the table.

The outside world has long been intrigued by OpenAI's "Strawberry" project, but for a long time, it remained elusive. This may soon change.

The Information recently reported that two individuals who have tested the "Strawberry" model claim that OpenAI may release it within the next two weeks, along with the launch of ChatGPT. However, unlike GPT-4o and GPT-4o mini, which focused on optimizing user experience and cost, "Strawberry" aims to propel large models closer to AGI once again.

Image/ The Information

Users who have canceled ChatGPT Plus subscriptions may need to resubscribe, but reportedly, the new model will have an "amazing" pricing strategy.

As "Strawberry" Ripens, Large Models Must Mature Too

The news that OpenAI will launch "Strawberry" in the next two weeks has undoubtedly garnered widespread attention in the industry. This not only represents OpenAI's latest endeavor in the field of large language models but also signals a potential revolution in AI reasoning capabilities.

Based on currently available information, "Strawberry" demonstrates unprecedented potential in tackling complex problems and executing multi-step tasks, marking a significant step towards AGI.

Firstly, the core highlight of the "Strawberry" project is its significantly enhanced reasoning abilities. Reports indicate that "Strawberry" can autonomously complete in-depth research tasks, suggesting that it goes beyond merely generating answers; it can plan ahead, navigate the internet automatically, and even solve highly complex scientific problems independently.

Compared to existing GPT-4 series models, "Strawberry" holds great promise in helping AI better mimic human thought processes, particularly in mathematics and science, successfully tackling the multi-step reasoning challenges that have plagued large models in the past.

Beyond its reasoning breakthroughs, "Strawberry" also introduces a new technique called "Post-training," which fine-tunes the model after pre-training on large datasets to excel at specific tasks.

Paper on STaR, Image/ arXiv

This approach resembles Stanford University's "Self-Taught Reasoner (STaR)" model, which iteratively improves its intelligence by self-generating training data. This enables "Strawberry" to perform more flexibly and efficiently in complex tasks, particularly long-horizon tasks (LHTs), significantly enhancing its planning and execution capabilities.

However, despite "Strawberry"'s potential, early tests have revealed some issues. For instance, response times can occasionally be slow for simple tasks, and the model's conversation memory function remains unstable. These issues undoubtedly impact user experience but are inherent in the exploration of new technologies.

The crux lies in whether OpenAI can address or improve these issues upon launch.

Regardless, "Strawberry" is arguably the most promising large model capable of qualitatively transforming AI reasoning capabilities, fueling heightened expectations both within and outside the industry. It signifies not only a new phase in the development of large language models but also a glimmer of hope for achieving AGI.

The Catalyst for OpenAI's "Palace Coup": From Q* to Strawberry

The birth of the "Strawberry" project was not overnight; its roots trace back to OpenAI's early secret project codenamed "Q*." While "Q*" may be unfamiliar to many, the OpenAI palace intrigue at the end of last year is well-known.

At the end of 2023, OpenAI experienced a management shakeup akin to a palace coup, with CEO Sam Altman briefly ousted by the board of directors. The immediate trigger was the security risks and controversies surrounding the "Q*" project.

"The Soul of OpenAI": Sam Altman (second from left) and Ilya Sutskever (fourth from left), Image/ OpenAI

The story began during the initial testing phase of "Q*," when the project demonstrated astonishing results in solving mathematical and scientific reasoning problems, particularly complex multi-step issues, surpassing existing large models. This sparked excitement among researchers but also deep concerns.

These breakthroughs hinted that OpenAI might be rapidly approaching AGI, a development beyond some board members' expectations. They feared the rapid evolution of AI posed uncontrollable risks and potentially profound societal impacts. Internal debates within OpenAI about whether to proceed with "Q*" intensified.

Some board members argued that the project's advancement was too aggressive, especially concerning AGI's potential dangers, and urged caution. Ultimately, disagreements reached a climax when Altman accelerated "Q*" without sufficient board communication, leading to his brief resignation.

However, the conflict effectively ended with the departure of the opposing side. After Altman's return, Ilya Sutskever, OpenAI's chief scientist and spiritual leader, never appeared at official launches again and officially resigned after GPT-4o's release.

Image/ X

Three days later, Jan Leike, the head of OpenAI's "Superalignment" team, also announced his resignation. Meanwhile, the "Q*" project accelerated, evolving into the widely known "Strawberry" project.

This summer, Altman shared a real photo of a strawberry on X, sparking speculation about "Strawberry"'s imminent release.

Image/ X

Marching Towards AGI Again: Will "Strawberry" Redefine AI?

Over the past two years, the direction of large model vendors has shifted. Initially, the industry revered OpenAI's Scaling Law as gospel, with each generation of generative AI models evolving alongside massive parameter increases, based on the belief that larger parameters foster greater intelligence emergence potential.

However, by 2024, large models began slowing their pursuit of "parameter scale," turning to medium to small parameter sizes for more cost-effective models. In July, even OpenAI introduced GPT-4o mini, following similar moves by renowned domestic and international large models that had already unveiled medium or smaller-scale versions.

Against this backdrop, OpenAI's "Strawberry" project stands out. Rather than stacking parameters for intelligence emergence, "Strawberry" aims to break through traditional AI performance bottlenecks by optimizing reasoning logic, enhancing reasoning capabilities, and leveraging self-training algorithms.

Based on current information, "Strawberry" appears to possess human-like reasoning abilities, offering new possibilities for solving complex multi-step problems. In fields like scientific research, complex decision-making, and data analysis, traditional language models often rely solely on simple text generation, whereas "Strawberry" can autonomously generate solutions tailored to problem complexity and validate them through reasoning.

Image/ DALL·E Generation

This transformation is crucial in advancing AI's role from a "tool" to an "intelligent assistant," benefitting numerous enterprises through enhanced AI reasoning capabilities.

Furthermore, "Strawberry" exhibits a degree of self-iteration and self-improvement, approaching the elusive "Recursive Self-Improvement" in AI. By self-generating training data and iteratively optimizing, "Strawberry" "learns" how to solve complex problems more effectively and leverages prior experience to quickly adapt to new tasks.

However, the "Strawberry" project faces notable challenges. For instance, stability in large-scale usage and the "security" issues that triggered OpenAI's palace coup. A more practical concern is that compared to existing multi-modal models like GPT-4, "Strawberry" reportedly lacks multi-modal processing capabilities, limiting its handling of complex data like images and videos.

Then there's the price.

It's crucial to note that the actual price of the "Strawberry" model remains unknown until its official announcement. However, it's almost certain that its training and inference costs will be higher. According to The Information's sources, OpenAI executives once discussed setting subscription prices for "Strawberry" and the "Orion" model (potentially GPT-5) at:

$2,000 per month.

Yes, you read that right—not $20 (ChatGPT Plus's current pricing) or even $200, but $2,000. While this was merely a discussion, possibly aimed at enterprise users, it underscores "Strawberry"'s high costs and OpenAI's confidence.

Current Plus Pricing, Image/ ChatGPT

Balancing high reasoning capabilities, costs, and user experience will be a challenge for the "Strawberry" project. Thus, OpenAI may consider offering "Strawberry" at different tiers and prices, similar to how it did with ChatGPT (GPT-3.5) at the end of 2022.

Regardless, the prerequisite remains "Strawberry"'s exceptional and impressive performance.

In the coming two weeks, all eyes will be on OpenAI to see if "Strawberry" can once again redefine the boundaries of AI.

Source: LeTech