OpenAI working on a new breakthrough reasoning technology code-named ‘Strawberry’

Codenamed strawberry, Sam Altern’s OpenAI is working on a new reasoning technology for its large language models (LLMs). It is reported that the company is hoping that Strawberry will dramatically improve the reasoning capabilities of its AI models. Strawberry is a closely guarded secret even within the organization. Previously known as Q*, it was seen as a breakthrough within the company.

However, OpenAI has shown Q* demos to some of its staff which show the LLMs capable of answering tricky science and maths questions that are currently beyond the reach of commercially available models.

The document describes a project using Strawberry models to enable the company's AI not only to generate answers to queries, but also to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI calls "deep research".

Strawberry will reportedly mark a specialized way of processing AI models after it has been pre-trained on very large datasets. It includes a specialized way of ‘post-training’ OpenAI's generative AI models or adapting them in order to improve their performance in specific ways even after they have been ‘trained’ on generalized data.

OpenAI reportedly wants to use Strawberry for performing long-horizon tasks (LHT), which require an AI model to plan ahead and perform a series of actions over an extended period of time.

Specifically, OpenAI wants its models to use these capabilities for conducting research by browsing the web autonomously with the support for ‘computer using agent’ or CUA which will be able to take action based on its findings.