Company OpenAIKnown for creating the popular chatbot ChatGPT, he is working on a new project under the code name strawberrydesigned to significantly improve the “reasoning” capabilities of its AI-based models. The Strawberry project is a closely guarded secret even within the company, according to a source familiar with the situation and internal documents reviewed by Reuters.

Project strawberry It aims to enable the model to not only produce answers to queries, but also have offline access to the internet and perform “deep research.” This is achieved thanks to a special way of processing the AI ​​model after pre-training it on large datasets.

Reasoning is key to AI reaching human-level intelligence. Although large language models can generalize and generate text much faster than humans, they often fail to tackle “common sense problems” whose solutions seem intuitive to humans.

Improving the logic of AI models is seen as the key to unlocking their power to do everything from making scientific discoveries to planning and building new software programs.

Project Strawberry was previously known as Q*, and as Reuters reported last year, the company already saw it as a breakthrough. Two sources said earlier this year that demo versions of Q* showed it could answer complex scientific and mathematical questions beyond today’s commercially available models.

OpenAI CEO Sam Altman said earlier this year that “the most important areas of progress” in AI will be around the ability to think.

Other companies, such as Google, Meta, and Microsoft, are also experimenting with different techniques to improve reasoning in AI-based models, but researchers disagree on whether large language models have the capacity to combine research and long-term planning.

In recent months, the company has hinted to developers and other external parties that it is on the verge of releasing technology with much more advanced “reasoning” capabilities. The Strawberry project involves a special method of OpenAI’s generative AI models called “post-training,” or adapting the underlying models to improve their performance in a specific way after they have already been trained on generalized datasets.

An OpenAI representative confirmed this fact at an internal general meeting where a demo version of the research project “with human-like reasoning skills” was shown. However, Reuters was unable to determine whether the Strawberry project had been introduced. A representative from the company did not directly answer questions about the new project, but said that “continuous research into new AI capabilities is common practice in the industry, and we all believe that these systems will eventually improve their reasoning.”

Reuters could not confirm the exact date of the document, which details a plan for how OpenAI plans to use Strawberry to conduct research. The source said the plan was in the works. The news agency could not determine how close Strawberry was to going public. Teams within OpenAI are working on Strawberry, according to a copy of an internal document seen by Reuters in May. The source said work on the project was a closely guarded secret, even within OpenAI.