[“OpenAI Nearing Human-Level Reasoning with 'Strawberry' Project”]

 OpenAI is developing a new approach, code-named 'Strawberry,' to improve the AI model's reasoning abilities to a human level. This is related to OpenAI's recent announcement that they are close to reaching Stage 2 of AGI (Artificial General Intelligence) out of 5 stages. Stage 2 is beyond simple chatbots (Stage 1) and involves 'Reasoners' who can think like highly educated people, similar to those with a Ph.D. 'Strawberry' aims to enhance AI's reasoning skills, especially in solving math problems.

 Strawberry uses an approach similar to Stanford University's 'STaR' method to improve performance through model retraining. 'STaR,' which stands for 'Self-Taught Reasoner,' is a method where AI enhances its reasoning abilities through repeated learning. This project requires massive computing resources, which is why CEO Sam Altman is building a global network to secure AI chips. Altman has emphasized that reasoning ability will be the most crucial area of advancement in AI.

 OpenAI plans to build 'Computer-Using Agents (CUA)' and develop models capable of performing Long-Horizon Tasks (LHT) through the Strawberry project. Long-Horizon Tasks involve complex activities where the model plans ahead and executes a series of tasks over an extended period, which OpenAI refers to as 'deep research.' These efforts are expected to continually improve AI's reasoning abilities over time, playing a crucial role in enabling AI to perform high-level reasoning similar to humans.

Source: “OpenAI Nearing Human-Level Reasoning with 'Strawberry' Project”, AI times, July 13, 2024.