ChatGPT developer OpenAI is working on a new approach to its artificial intelligence (AI) models in a project codenamed Strawberry, Reuters reported, citing its sources and OpenAI documentation.
► Read Telegram channel «Ministry of Finance»: main financial news
The new project, the details of which have not previously been reported, aims to push the limits of advanced AI models, in particular to improve their logical thinking abilities.
Details of the work on Strawberry are kept strictly confidential, even by OpenAI.
A Reuters document in possession describes a project using Strawberry models to allow the company's artificial intelligence to not just generate responses to queries, but also plan ahead to autonomously and reliably navigate the Internet to perform what OpenAI calls “deep exploration” “.
The Strawberry project was previously known as Q, which was already seen internally as a breakthrough. However, Reuters was unable to find out whether Strawberry would become publicly available.
Two Reuters sources said that earlier this year they saw what OpenAI employees called a demo of Q, capable of answering complex scientific and mathematical questions that could not be achieved for today's commercial models.
Artificial intelligence researchers say that thinking is key to achieving human- or superhuman-level AI.
While large language models can already summarize large texts and compose elegant prose far faster than any human, the technology often struggles with common-sense problems that seem intuitive to humans, like recognizing logical fallacies and playing tic-tac-toe.
When faced with such problems, the model often hallucinates false information.
In general, experts agree that thinking in the context of AI means forming a model that allows the AI to plan ahead, reflect how the physical world works, and reliably solve complex, multi-step problems.
Improving the thinking in AI models is seen as key to unlocking the models’ ability to do everything from make big scientific discoveries to planning and creating new software applications.