Project “Strawberry”: OpenAI develops new inference AI technology

As part of an ambitious effort to maintain a leading position in the rapidly evolving field of artificial intelligence, OpenAI has been secretly working on a new approach to AI models codenamed “Strawberry.”

The revelation, revealed in internal documents seen by Reuters and people familiar with the matter, comes as the Microsoft-backed startup best known for its ChatGPT product makes strides to demonstrate that its models have advanced inference capabilities, which could mark a major leap for AI technology.

A look inside the Strawberry project

A recent internal document seen by Reuters in May said the OpenAI team was hard at work on the Strawberry project. While the exact timeline in the document is unclear, it outlines how OpenAI plans to use Strawberry for advanced AI research.

The project is said to be ongoing and shrouded in secrecy even within the company. Strawberry’s goal is to enable an AI not just to generate answers, but to navigate the internet autonomously and reliably, conducting what OpenAI calls “deep research.”

“This is something that has never been possible with an AI model before,” the source said, noting the ambitious nature of the project.

Asked about Strawberry and the details reported in this article, an OpenAI spokesperson responded in a statement: “We want our AI models to see and understand the world in the same way humans do. Continuous research into new AI capabilities is common practice in the industry, and there is a shared belief that these systems will improve in their reasoning capabilities over time.”

The spokesman did not directly respond to questions about Strawberry.

From Q to Strawberries: A New Era of Inference

Strawberry is the successor to an earlier project known as Q, which was already considered groundbreaking within OpenAI for its ability to answer complex science and math questions that are beyond the scope of current commercially available models, according to two sources.

According to Bloomberg, OpenAI demonstrated a research project at an all-hands meeting this year showcasing new human-like reasoning skills. Reuters could not confirm whether the project demonstrated was Strawberry, but this is in line with the company’s ongoing efforts to bolster its AI reasoning.

OpenAI CEO Sam Altman emphasized the importance of inference in AI, saying earlier this year that “the most important area of ​​advancement will be in inference capabilities.”

Challenges of AI inference

Researchers believe that improving the reasoning capabilities of AI models is key to achieving human or even superhuman levels of intelligence. Large language models can summarize text and produce prose efficiently, but they often perform poorly on common sense problems and logical tasks, leading to “hallucinations” and the generation of misinformation.

Reasoning, as AI researchers describe it, includes an AI’s ability to make plans, understand the physical world, and solve multi-step problems.

“Inference is key to enabling AI to achieve human or even superhuman levels of intelligence,” said an AI researcher interviewed by Reuters.

OpenAI’s Strawberry project aims to overcome these challenges by employing a specialized post-training process, which involves fine-tuning a pre-trained AI model on an extensive dataset.

One source said Strawberry’s approach is similar to Stanford University’s Self-Taught Reasoner (STaR), which has the potential to achieve higher levels of intelligence by allowing AI models to iteratively create their own training data.

“I think that’s exciting and frightening at the same time… if things continue to move in that direction, we as humanity have some serious reasons to think,” said Stanford University professor Noah Goodman, one of STaR’s developers.

Long-term projects and autonomous research

One of Strawberry’s ambitious goals is the ability to perform Long-Term Tasks (LHTs), which require the AI ​​to plan and execute a series of actions over an extended period of time.

To enable these capabilities, OpenAI is training and evaluating models on a “deep research” dataset, according to internal documents. Details of the dataset and the length of the extension were not disclosed, but the goal is clear: to enable the AI ​​to conduct research autonomously with the assistance of a computer-used agent (CUA) that can act on the findings.

The competitive AI industry

OpenAI is not alone in its efforts to improve AI inference: major technology companies such as Google, Meta, and Microsoft, as well as numerous academic labs, are also exploring different techniques to improve AI inference capabilities, although there is disagreement over whether large-scale language models can incorporate long-term planning and advanced reasoning into predictions.

Yann LeCun, a pioneer of modern AI who works at Meta, has frequently expressed skepticism about the ability of large language models (LLMs) to achieve human-like reasoning.

Strawberry is a key element of OpenAI’s strategy to address the limitations of current AI models. By developing more advanced inference capabilities, OpenAI aims to unlock new possibilities for AI, from scientific discovery to creating new software applications. The company has been telling developers and partners that it is on the verge of releasing technology with significantly enhanced inference capabilities.

Strawberry’s development includes post-training methods such as fine-tuning, which involves human feedback and an iterative learning process. These techniques are designed to refine AI models and improve their performance on specific tasks.

The advances made through Strawberry have the potential to redefine AI capabilities and establish a new standard for what these models can achieve. While the road ahead is fraught with challenges, the potential benefits are enormous, ushering in a new era of intelligent, autonomous AI systems.

“We want AI models to see and understand the world the way humans do,” an OpenAI spokesperson said. If Strawberry is successful, it could be one step closer to realizing that vision.