The way OpenAI will determine the strength of the AI system: Rating scale 1 to 5, self-assessment surprise score

OpenAI has just announced a new scoring system, aiming to assess the progress of large language models (LLMs) on the path to artificial general intelligence (AGI). This rating scale, from level 1 to level 5, promises to provide a more objective and transparent view of AI's future.

According to Bloomberg, current chatbots like are only at level 1. OpenAI is confident that they are nearing level 2, with the ability to solve basic problems equivalent to a PhD holder. Level 3 describes AI's ability to perform tasks autonomously. Level 4 is when AI can generate new innovations. Finally, level 5, also the ultimate goal of AGI, is AI's ability to replace an entire human organization. Previously, OpenAI defined AGI as 'highly automated systems that outperform humans in most economically valuable tasks.'

However, AGI remains a distant goal, requiring massive computational resources and long-term research. Experts, including those at OpenAI, still cannot accurately predict when AGI will be achieved. In October 2023, OpenAI CEO Sam Altman said it would take 'about 5 years' to reach this milestone.

Although still in development, this new rating scale was introduced just one day after OpenAI announced a partnership with the Los Alamos National Laboratory. The goal of this collaboration is to explore the potential of advanced AI models like GPT- in safely supporting biological research. According to a program manager at Los Alamos interviewed by The Verge, the project aims to test the capabilities of GPT- and establish safety standards and other factors for the US government.

OpenAI's lack of detailed information on how models are classified using this rating scale (and their refusal to comment to The Verge) has raised many concerns. Particularly in the context of OpenAI's safety research group being disbanded last May after the head of the group, OpenAI co-founder Ilya Sutskever, left the company. Jan Leike, a key researcher at OpenAI, also resigned shortly after, citing that 'safety culture and processes have been placed behind flashy products' at the company.

Despite OpenAI's denial, many still express concerns about the potential implications if the company truly achieves AGI. The discrepancy in opinions between CTO Mira Murati (who believes models in OpenAI's lab are not much better than what the public knows) and CEO Sam Altman (who believes recent AI models have made significant advancements) further unsettles public opinion.

Thang điểm mới của OpenAI, mặc dù còn gây tranh cãi, được kỳ vọng sẽ mang lại một phương pháp đánh giá rõ ràng và minh bạch hơn cho sự phát triển của trí tuệ nhân tạo. Thay vì chỉ dựa vào những tuyên bố mơ hồ, bây giờ chúng ta có thể đánh giá tiến bộ của AI dựa trên những tiêu chí cụ thể.

Trong cuộc họp nội bộ diễn ra vào thứ Năm vừa qua, ban lãnh đạo OpenAI đã trình bày một dự án nghiên cứu sử dụng mô hình AI GPT-4. Theo thông tin từ Bloomberg, dự án này đã cho thấy một số kỹ năng mới của AI, thể hiện khả năng lập luận gần giống với con người. Điều này cho thấy tiềm năng to lớn của AI trong tương lai, đồng thời đặt ra những thách thức mới về đạo đức và an toàn.