OpenAI's "strawberry" model focused on inference to be released in two weeks? It's the next big breakthrough in artificial intelligence

The author of this article: Du Yu

Source: Hard AI

On Tuesday, September 10, OpenAI, the "AI darling" on the cusp of the times, plans to release the latest artificial intelligence product called "Strawberry" in the next two weeks as part of the services provided by ChatGPT, according to The Information, a well-known technology and business media that has tested the model. This is earlier than the rumored "autumn release" (September-November) two weeks ago.

According to the above report, there are still some urgent problems to be solved before and after the release of the artificial intelligence "strawberry" focusing on reasoning ability, but the seemingly "hasty launch" seems to indicate that OpenAI feels the fierce competitive pressure in the field of large language model-driven products, and hopes to restore the momentum of ChatGPT for most of the two years since its debut through the long-awaited strawberry project in recent months.

What is the "strawberry" model? Focusing on reasoning is the next big breakthrough in artificial intelligence

Two weeks ago, it was reported that OpenAI was accelerating the launch of its "strawberry" inference AI model to drive the chatbot business. Four weeks ago, it was reported that the CEO of perplexity, a well-known AI search company, hinted that OpenAI's latest model "Strawberry" had been used in advance.

Wall Street has mentioned that as early as May, there were rumors that OpenAI was mysteriously developing a project codenamed "Strawberry". Two months ago, on July 13, it was confirmed that "Strawberry" is a new artificial intelligence model that OpenAI is developing, hoping to provide more advanced inference capabilities to form a major breakthrough towards artificial general intelligence (AGI).

Specifically, the "strawberry" model is better able to handle complex scientific and mathematical problems that have caused previous AI models to "scratch their heads" with their enhanced reasoning capabilities, not only allowing large language models (LLMs) to generate answers based on complex queries of users, but also planning ahead to autonomously and reliably navigate the Internet for "deep research" as defined by OpenAI.

Commentators generally agree that OpenAI's "Strawberry Project" is expected to be the next major breakthrough in artificial intelligence. More advanced reasoning capabilities mean that the technology can perform more complex tasks "autonomously" rather than just relying on the user's step-by-step prompts. Reasoning has always been one of the most intractable problems for AI in the pursuit of human-like intelligence, and AI models with inference capabilities can not only improve math and programming skills, but also revolutionize scientific research, enhance decision-making in complex fields such as healthcare and finance, and accelerate technological innovation.

Will it be significantly more expensive? The exact form and pricing model of the "strawberry" model is uncertain

It is reported that although "Strawberry" is a part of ChatGPT as a whole after its release, it is also an independent product, and may be included in the drop-down menu of AI models selected by customers to support ChatGPT operation, but the specific form of provision and pricing model are uncertain.

Unlike regular conversational AI, Strawberry "thinks" for 10 to 20 seconds before responding to the user's prompt, rather than giving an answer right away. For queries that are complex or contain multiple execution step requirements, "Strawberry" will be easier to use than the existing version of GPT-4o, and will not require the user to explicitly tell ChatGPT to "complete an intermediate inference step to arrive at the final answer" and other additional actions.

According to some analysts, this means that "Strawberry" will become a more intelligent assistant, not only better at solving math problems and coding, but also better at more "subjective" business tasks, such as brainstorming product marketing strategies. And for such creative tasks, the model will provide more detailed recommendations that are more specific to the needs of the user's company, such as generating a weekly execution plan.

However, assuming that the release goes well, in the initial version, "Strawberry" can only receive and generate text information, but not images, which means that it does not have "multimodal" functions like other OpenAI models, which is one of the "shortcomings".

At the same time, while the "thinking" phase helps prevent AI from making mistakes and predicts how to answer user questions in a comprehensive manner, some people who have tested the model say that some simple questions can also make strawberries "think" and not be able to give quick answers. The quality of the answers is only marginally better and not worth the extra wait; And sometimes it's not possible to remember and integrate previous chats with users before answering new questions.

In terms of pricing, it was reported last week that ChatGPT Plus's current paid price is $20 per month, but OpenAI is considering paying $2,000 per month for new large models such as "Strawberry" and "Orion", which is known as "ChatGPT 5", which is equivalent to a 100-fold increase in pricing. According to the analysis, OpenAI can consider such a high price span, mainly because of the lack of money, the new product is strong enough, and it has hundreds of millions of users.

Why is the "strawberry" model important? Or redefine the boundaries of AI capabilities to provide GPT 5 with high-quality synthetic training data

OpenAI CEO Sam Altman said earlier this year that "the most important direction of progress in the field of artificial intelligence will be reasoning ability."

The "strawberry" project is believed to be OpenAI's "Q*" plan, which was revealed last year, and the project has been seen as a breakthrough within the company. Those who have seen the demo say it can answer tough scientific and mathematical questions that current commercial AI models can't solve.

Ultraman's post in early August was considered by the outside world to be a warm-up for the "Strawberry" project

OpenAI gave a five-stage roadmap for future AI development on August 11, the first stage is the AI chatbot with conversational language, the second stage is the "reasoner" with human-level problem-solving ability, the third stage is the "AI agent" who can take autonomous action, followed by the "innovator" who can assist in invention and the "organizer" who can complete the organization's work.

Once successful, it will redefine the boundaries of AI's capabilities, autonomously perform complex tasks, solve complex problems independently, and verify the output content, representing a step closer to AGI for human beings. Strawberry, which focuses on reasoning, may also be the way to lay the foundation for the next generation of AI model Orion, or ChatGPT 5.

Ghazenfer Mansoor, founder and CEO of Technology Rivers, a software development organization, pointed out that with the "Strawberry" project, OpenAI's pursuit of human reasoning is not only a technological miracle, but also a breakthrough that can push boundaries and stimulate new possibilities in every field. In business, for example, AI will dramatically change the way supply chains are managed, predict market trends with great accuracy, and largely personalize the customer experience.

For example, AI models with inference capabilities can not only provide better travel plans and health recommendations, but can even analyze users' spending habits and create fully customized budget plans to become smarter assistants.

All in all, in an ideal scenario, "Strawberry" would be able to solve problems and tasks beyond the capabilities of current AI models and conduct "deep dives" that "outperform any existing chatbot".

There are also analysts who point out that the "Strawberry" program aims to raise more money, and OpenAI needs this money to support the development of the next generation of cutting-edge models codenamed "Orion". Moreover, generating high-quality synthetic training data for the Orion model is considered one of the key applications of Strawberry, "which is significant because most of the training data on the Internet has been exhausted":

The large datasets on which AI models are based can be prone to bias and errors, or contain incomplete or inaccurate information. The high-quality synthetic data generated by Project Strawberry can fill gaps in real-world datasets and provide a more comprehensive, inclusive, and balanced training set.

Many believe that the use of synthetic data can help future AI models be more neutral and fair, and reduce noise, hallucinations, and irrelevant information, thereby improving the efficiency of training and the accuracy of models.

Based on what is known, Project Strawberry improves reasoning, logic, and the ability to plan and conduct research, allowing models to autonomously conduct experiments, analyze data, and formulate new hypotheses. This could lead to scientific breakthroughs, including the discovery of new drugs. These models can also provide personalized education, create educational content, and interactive lessons.

According to people familiar with the matter, the "Strawberry" project includes a specialized "post-training" approach, in which generative AI models have been pre-trained on large datasets and then further tuned to improve their performance on specific tasks. This is similar to the "Self-Taught Reasoner" (STaR) method developed by Stanford University in 2022.

Noah Goodman, a professor at Stanford University and one of the creators of STaR, has said that STaR allows AI models to "guide" themselves to a higher level of intelligence by iteratively creating their own training data, which can theoretically be used to enable language models to achieve intelligence beyond humans. But he also admits that "it's both exciting and scary, and there are some questions that humanity needs to think about [in terms of its own situation]." ”

This article is from Wall Street News, welcome to download the APP to see more