OpenAI has introduced a new leveling system to measure progress in creating artificial intelligence capable of surpassing humans. The move, the company says, is necessary to help people better understand security issues and the future of AI.
Image source: Andrew Neel/Unsplash
According to Bloomberg, at a recent general meeting at OpenAI, management presented a new classification that will be used for internal assessments and communication with investors. According to a company spokesperson, the levels range from “today” AI capabilities such as interaction in spoken language (Level 1) to AI capable of doing the work of an entire organization (Level 5).
OpenAI executives told employees that the company was currently at level one, but was close to achieving level two, which they called “Reasoners.” This level denotes systems that are capable of solving basic problems at the level of a person with a doctorate degree. A research project based on the GPT-4 model was also shown, demonstrating new AI skills.
Image source: David Paul Morris/Bloomberg
OpenAI is also working to create what it calls artificial strong intelligence (AGI), which can perform most tasks better than a human. Although such a system does not currently exist, OpenAI CEO Sam Altman says AGI could be created within this decade.
Interestingly, AI developers have long debated the criteria for achieving AGI. So, in November 2023, employees from Google DeepMind proposed their own system of five levels of AI, including criteria such as “expert” and “superhuman”. These levels are reminiscent of the system used in the auto industry to rate the degree of automation in autonomous vehicles.
However, OpenAI plans to use its level development to measure progress in creating safe and efficient AI systems that can solve complex problems while remaining under human control. The company has come up with a third layer, which will be called “Agents,” referring to AI systems that can perform actions under the guidance of the user. Level 4 AI will be able to generate innovative ideas. And the most advanced fifth level will be called “Organizations”.
All levels listed were compiled by OpenAI executives and senior staff. The company will collect feedback from developers, employees, investors and may adjust the levels over time.