OpenAI’s New ‘Strawberry’ Project
According to an unnamed source and internal documentation reviewed by Reuters, OpenAI has a project designed to deliver advanced reasoning capabilities. The source described it to Reuters as a work in progress with no indication of availability.
The gist is that Strawberry models will enable the company’s AI to generate answers to queries and plan enough to navigate the internet autonomously and reliably. This is what OpenAI calls “deep research.” OpenAI said, “We want our AI models to see and understand the world more like we do. Continuous research into new AI capabilities is a common practice in the industry, with a shared belief that these systems will improve in reasoning over time.”
The project was formerly known as Q* and, according to Reuters, OpenAI staffers told them there are Q* demos capable of answering tricky science and math questions beyond today’s models.
An OpenAI spokesperson confirmed a recent internal hands-on meeting that showed a research project demo, which may or may not be Strawberry. The indication is that Strawberry involves a specialized way of processing an AI model after pre-training on very large datasets.
Improving reasoning is key to unlocking the ability of AI to perform major scientific discoveries, as well as planning and building new software applications. OpenAI is not alone in going down the “reasoning” path. Google, Meta, and Microsoft are also experimenting to improve reasoning in AI models. Approaches differ on whether LLMs can incorporate ideas and long-term planning into prediction.
In recent months, the company is signaling that it is close to releasing technology with significantly more advanced reasoning capabilities. Strawberry’s post-training phase is akin to “fine-tuning.”
OpenAI wants Strawberry to perform long-horizon tasks (LHT) that require a model to plan and perform a series of actions over an extended period. Thus, OpenAI is creating, training, and evaluating the models on what the company calls a “deep-research” dataset. OpenAI specifically wants its models to be able to conduct research by browsing the web autonomously with the assistance of a “CUA,” or a computer-using agent, that takes actions based on the findings. The company also plans to test its capabilities on performing the work of software and machine learning engineers.
