On Monday, OpenAI announced the creation of a “Safety and Security Committee” to manage risks associated with its projects and operations. This announcement coincides with the company beginning the training of a new “frontier model,” which aims to advance towards artificial general intelligence (AGI). While OpenAI aspires to reach AGI—AI with human-like capabilities for various tasks—critics argue that this goal is still far off.
The term “frontier model” refers to cutting-edge AI systems that push current technological boundaries. It is unclear if OpenAI’s new model is GPT-5 or a different advancement.
The Safety and Security Committee, chaired by OpenAI directors Bret Taylor, Adam D’Angelo, Nicole Seligman, and CEO Sam Altman, will advise the board on AI safety. This includes preventing AI from becoming uncontrollable and implementing safeguards related to alignment research, child protection, election integrity, societal impact, and security.
The committee will spend the next 90 days evaluating and improving these safeguards, after which they will present recommendations to the board and provide a public update.
Technical and policy experts, including Aleksander Madry, Lilian Weng, John Schulman, Matt Knight, and Jakub Pachocki, will support the committee’s work.
This move follows negative publicity from the resignation of Ilya Sutskever and Jan Leike from OpenAI’s Superalignment team, which aimed to control highly advanced AI. Their departure has led to criticism about OpenAI’s commitment to safe AI development. Some, like Meta’s Yann LeCun, argue that concerns about AGI safety are premature.
There are also rumors that advancements in large language models (LLMs) have stalled at levels similar to GPT-4. Competing models from Anthropic and Google are on par with GPT-4. Recently, instead of a superior model, OpenAI released GPT-4o, which matches GPT-4’s abilities but is faster. This suggests that significant improvements beyond GPT-4 may not be imminent.
