OpenAI has created an internal scale to track how its large-scale language models are progressing towards artificial general intelligence, or AI with human-like intelligence, a spokesperson said. Bloomberg.
Today’s chatbots like ChatGPT are Level 1. OpenAI claims we’re approaching Level 2, which it defines as a system that can solve basic problems at the level of a person with a PhD. Level 3 refers to AI agents that can take actions on behalf of users. Level 4 is AI that can create new innovations. Level 5 is the final stage to achieving AGI, an AI that can perform the work of an entire organization. OpenAI previously defined AGI as “highly autonomous systems that outperform humans at most economically valuable tasks.”
OpenAI’s unique structure is centered around its mission to achieve AGI, and how OpenAI defines AGI is key. The company pledges to drop everything and support any value-aligned, safety-focused project that gets closer to building AGI than OpenAI does, rather than competing with it. While this language in OpenAI’s charter is vague and leaves room for judgment for a for-profit organization (governed by a non-profit), the scale at which OpenAI can test itself and its competitors could help it more clearly determine when AGI is achieved.
Still, AGI is still a long way off. Reaching AGI, if it’s even possible, will require billions of dollars worth of computing power. Experts’ timelines vary widely, even OpenAI’s. In October 2023, OpenAI CEO Sam Altman said it’s “about five years away” from reaching AGI.
The new criteria, which are still under development, were introduced a day after OpenAI announced a collaboration with Los Alamos National Laboratory to explore ways in which advanced AI models like GPT-4o can safely aid biological science research. The Los Alamos program manager in charge of the national security biology portfolio, who was instrumental in securing the partnership with OpenAI, said: The Verge The goal is to test GPT-4o’s capabilities and establish a set of safety and other parameters for the U.S. government, so that you can eventually test public or private models against these parameters and evaluate your own models.
In May, OpenAI disbanded its safety team after its leader, OpenAI co-founder Ilya Sutskever, left the company. Jan Reicke, a principal researcher at OpenAI, resigned shortly after posting that “safety culture and safety processes have taken a back seat to flashy products” at the company. OpenAI has denied that this is the case, but some worry about what this means if the company does indeed reach AGI.
OpenAI has not provided any details about how it assigns models to these internal levels (and The VergeIn response to a request for comment, the company said, “The GPT-4 AI model showcases new skills that allow humans to perform human-like reasoning.” However, the company said that company executives demonstrated a research project using the GPT-4 AI model during an all-hands meeting on Thursday, and that the company believes the project showcases several new skills that demonstrate human-like reasoning. Bloomberg.
This metric can help provide a strict definition of progress, rather than leaving it open to interpretation. For example, OpenAI CTO Mira Murati said in a June interview that the company’s lab models aren’t all that different from the models the public already has, while CEO Sam Altman said late last year that the company had recently “lifted the veil of ignorance,” meaning its models had gotten significantly smarter.
