OpenAI is gearing up to develop a new reasoning tool for its large language models (LLMs) called 'Strawberry', according to Reuters, citing internal business documents and sources familiar with the topic. According to reports, OpenAI is gearing up for Strawberry to significantly increase the reasoning powers of its AI models.

Strawberry, according to the report, is a well-guarded secret within the company. It was previously known as Q* and was regarded as a significant breakthrough within the company. However, OpenAI has provided Q* demos to some of its employees, demonstrating that the LLMs are capable of addressing difficult science and math issues that are currently above the scope of commercially available models.

According to the source, the document discusses a project employing Strawberry models to enable the company's AI to not only generate answers to enquiries, but also to plan ahead enough to explore the internet autonomously and dependably to undertake what OpenAI refers to as "deep research".

OpenAI is gearing up
Image Source: Youtube

Capabilities of ‘Strawberry’

According to reports, Strawberry will identify a certain method of processing an AI model after it has been thoroughly pre-trained on enormous datasets. It contains a specific method for "post-training," or modifying, OpenAI's generative AI models to enhance their performance in particular ways even after they have been "trained" on a broad set of data.

According to reports, OpenAI intends to employ Strawberry for long-horizon tasks (LHT), which call for an AI model to plan ahead and carry out a sequence of operations over a considerable amount of time.

More specifically, OpenAI hopes that its models will employ these capabilities to do research through autonomous web browsing, supporting "computer using agent," or CUA, which will be able to act on its discoveries.

You might also be interested in - OpenAI co-founder Schulman quits, Joins rival Anthropic