OpenAI is working on an advanced AI technology called Strawberry
The project involves Microsoft-backed startups trying to show how different types of models can provide advanced reasoning capabilities. Strawberry is a separate team within OpenAI, and the details of its development are unknown even to other employees.
Strawberry's models will allow AI to not just generate answers to queries, but also plan "deep explorations" in advance.
An OpenAI spokesperson told reporters: “We want AI models to see and understand the world better than we do. Continuous exploration of new AI capabilities is common practice in the industry, and there is a general belief that these systems will get better at reasoning over time.”
The Strawberry project was previously called Q*. Sources who saw its presentation said the AI could answer complex scientific and mathematical queries that commercially available models could not handle.
This week, OpenAI showed off a demo of a research project that it claims has new human-like thinking skills. The company hopes the innovation will significantly improve the reasoning abilities of its AI models. The source notes that Strawberry includes a special way to process an AI model after it has been pre-trained on very large data sets.
AI researchers interviewed by reporters generally agree that reasoning in the context of AI involves forming a model that allows AI to plan ahead and reliably solve complex multi-step problems.
Strawberry involves a special way to do what’s called “post-training” for OpenAI’s generative AI models, or adapting baselines, which would allow AI models to hone their performance in specific ways after they’ve been “trained” on masses of generalized data, one of the people said. Strawberry is similar to a method developed at Stanford in 2022 called “Sentient Self-Taught,” or “STaR,” one of the people familiar with the matter said. STaR allows AI models to “bootstrap” to higher levels of intelligence by iteratively creating their own training data. Stanford professor Noah Goodman said, “I think it’s both exciting and scary … if things go in this direction, we as humans are going to have to think about some serious things.”
The company's paper says Strawberry will enable long-horizon tasks (LHT), which require the model to plan ahead and perform a series of actions over an extended period of time.
OpenAI plans to have models use these capabilities to conduct research by browsing the web autonomously using a “CUA,” or computer-based agent. The company also intends to test the AI’s capabilities in performing the functions of software and machine learning engineers.
OpenAI recently discussed its new roadmap for developing AGI (Artificial General Intelligence) during an internal meeting . The company admitted that it has only achieved the first stage of five possible stages.