OpenAI is almost ready with the revolutionary AI Strawberry – it can plan and reason

OpenAI is developing an artificial intelligence system that uses a new approach – the project is codenamed Strawberry. What makes the new model different is its ability to reason. This was reported by Reuters with reference to an internal OpenAI document, which the agency’s journalists reviewed back in May.

Image source: Mariia Shalabaieva / unsplash.com

The exact date of the document could not be determined, but it details how the company intends to use Strawberry for research – the model is currently under development, a source told the publication. It was also not possible to establish how close the Strawberry model is to being released to the general public. It is classified and access to it is strictly guarded even within OpenAI. The document describes a project in which Strawberry not only provides answers to questions, but creates a plan for autonomous AI navigation on the Internet to perform some kind of “deep research”.

OpenAI did not remain silent or deny the existence of the project. “We want our AI models to see and understand the world the same way we do. Continuous exploration of new AI capabilities is common practice in the industry, and we share confidence that these systems will improve their reasoning abilities in the future,” a company spokesperson said. Work on the project was carried out last year, but then it was called Q* (“Q with an asterisk”), and the incident with the dismissal of Sam Altman occurred shortly after its launch and the receipt of the first results. Two OpenAI employees reported witnessing demonstrations of Q*’s capabilities this year, where the model successfully answered complex scientific questions and solved mathematical problems.

Image source: Lukas / pixabay.com

On Tuesday, the company held an internal general meeting, at which a certain research project was shown – AI with new reasoning skills similar to humans. An OpenAI representative confirmed the meeting took place, but declined to say what happened at the meeting; Reuters was unable to determine whether this was about the Strawberry project. The next-generation system is expected to set a new benchmark in AI’s ability to reason, thanks to a new way of processing the model, which has been pre-trained on very large data sets.

In recent months, OpenAI has privately signaled to developers and other third parties that it is on the verge of releasing technology related to significantly more advanced AI reasoning abilities, anonymous sources said. A special feature of Strawberry is a unique method for processing the AI ​​system after the training procedure – most often this means “fine-tuning” the model. In the case of Strawberry, we are talking about similarities with the StaR (Self-Taught Reasoner) method, which was developed in 2022 at Stanford University (USA): it describes the self-learning of AI and the model’s iterative preparation of its own data sets for subsequent additional training – this scheme is in theory can be used to create an AI model that surpasses human-level intelligence.

Strawberry’s most important ability is to perform tasks that require planning ahead and performing a series of actions over a long period of time. To do this, OpenAI creates, trains and evaluates models using “deep research” data – journalists were unable to determine the composition of this data set and the duration of the period for which the AI ​​makes a plan. This model implements its own research projects, autonomously surfing the web with the help of a special agent – a computer user (Computer-Using Agent, CUA). As part of the verification process, such a model will perform tasks assigned to software and machine learning engineers.