OpenAI, the company behind ChatGPT, is reportedly working on an innovative AI project called “Strawberry.”
The project, formerly known as Q, would be a major step forward for the company, according to sources familiar with the matter and internal documents seen by Reuters.
OpenAI hasn’t publicly revealed any details about Strawberry, but the startup is working to demonstrate that its models can provide advanced inference capabilities. The OpenAI team is actively working on Strawberry, but the exact release date for the technology is unclear.
The details of Strawberry’s capabilities are a closely guarded secret even within OpenAI. The project aims to enhance AI’s ability not just to generate answers but to autonomously navigate the internet and conduct in-depth research — a capability that, as several AI researchers have noted, has been a challenging goal for AI models thus far.
When asked about Strawberry, an OpenAI spokesperson said, “We want our AI models to see and understand the world the same way we 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.”
A demonstration of Q earlier this year reportedly showed the device’s ability to answer complex scientific and mathematical questions that current commercially available models cannot handle.
During an internal meeting on Tuesday, OpenAI reportedly presented a demo of a research project that claims new human-like reasoning capabilities.
OpenAI hopes that Strawberry will significantly improve the inference capabilities of AI models, which involves a specialized post-training process that takes place after an AI model is pre-trained on an extensive dataset. Researchers agree that improved inference capabilities are essential for AI to attain human or even superhuman intelligence.
Large language models can efficiently summarize dense text and produce elegant prose, but they often struggle with common-sense problems that humans can intuitively solve. Addressing these problems is essential for AI models to achieve advanced reasoning capabilities.
Strawberry is part of OpenAI’s broader strategy to overcome these challenges. The company is signaling to developers and other stakeholders that it is close to releasing technology with significantly improved inference capabilities. Strawberry includes specialized post-training techniques to improve the performance of generative AI models after initial training on generalized data.
The project shares similarities with Stanford University’s 2022 Self-Taught Reasoner (STaR) method, which allows AI models to iteratively create their own training data to reach higher levels of intelligence. One of its creators, Stanford professor Noah Goodman, pointed out the potential and implications of such advances:
OpenAI aims to use Strawberry to perform long-term tasks (LHTs), where the model must plan and execute a series of actions over an extended period of time. The company is developing and testing the model on its “Deep Research” dataset, but the details of the dataset and the duration of these tasks are still unknown.
Additionally, OpenAI plans for its models to conduct research autonomously online with the help of “CUAs” (Computer-Use Agents) that can act on the findings. The company also plans to test these capabilities on software and machine learning engineering tasks.
