OpenAI’s Strawberry Aims for Advanced Reasoning Capabilities
OpenAI is developing a new approach to AI with a project code-named “Strawberry,” according to a person familiar with the matter and internal documentation reviewed by Reuters. This novel project aims to enhance the reasoning capabilities of AI models, marking a significant step forward in AI research.
What is OpenAI Strawberry?
Internal documents reveal that the OpenAI Strawberry project enables AI to plan and navigate the internet autonomously. It is conducting what OpenAI terms “deep research.” This level of functionality has eluded AI models to date, despite advancements in the field.
An OpenAI spokesperson commented on the company’s broader research goals, stating, “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.“.
However, the spokesperson did not directly address specific questions about Strawberry. Previously known as Q*, early demos showcased its ability to solve complex science and math problems. This exceeded the capabilities of current commercially available models.
One source revealed that an internal AI test scored over 90% on a MATH dataset, a benchmark for championship math problems, though it remains unclear if this was related to Strawberry.
At a recent internal meeting, OpenAI demonstrated a project with new human-like reasoning skills, according to Bloomberg. While the specifics of the demonstration remain undisclosed, it indicates significant progress in advanced reasoning research.
How are they doing it?
This post-training, or “fine-tuning,” involves adapting base models with targeted feedback and examples of good and bad answers. OpenAI Strawberry’s methodology bears similarities to Stanford’s 2022 “Self-Taught Reasoner” (STaR).
The way it works is that it iteratively creates its own training data to boost intelligence levels. Stanford professor Noah Goodman, a creator of STaR, remarked, “I think that is both exciting and terrifying…if things keep going in that direction, we have some serious things to think about as humans.”
One of Strawberry’s key goals is to perform long-horizon tasks (LHT), which require planning and executing a series of actions over extended periods. OpenAI is training and evaluating Strawberry models on a “deep-research” dataset to achieve this, although specific details about the dataset remain undisclosed.
OpenAI also aims to use Strawberry to conduct research autonomously on the web with the help of a “computer-using agent” (CUA) that can act based on its findings. This capability could revolutionize the way AI models perform complex research tasks, potentially handling the work of software and machine learning engineers.
With that said. It’s good to note something that OpenAI CEO Sam Altman emphasized. Earlier this year, he stated the importance of reasoning, “the most important areas of progress will be around reasoning ability.“
Originally posted on OpenDataScience.com
Read more data science articles on OpenDataScience.com, including tutorials and guides from beginner to advanced levels! Subscribe to our weekly newsletter here and receive the latest news every Thursday. You can also get data science training on-demand wherever you are with our Ai+ Training platform. Interested in attending an ODSC event? Learn more about our upcoming events here.