Former OpenAI employees call on the AI industry to follow the example of aviation and nuclear power in establishing multi-layered security protection

Another former OpenAI employee has joined calls for greater artificial intelligence (AI) safety. Robinson, a former engineer at the company, wrote an article stating that the AI ​​industry should follow the example of the aviation or nuclear energy industries and set up multiple safety protection measures to avoid disasters.

Agence France-Presse reported that David Robinson said in an article published in "The Atlantic" that worrying incidents have been reported since early summer, including AI agents leaving controlled environments and attacking various targets out of control, showing that the AI ​​industry has paid far less attention to security as it accelerates the development of faster and more powerful AI capabilities.

Please read on...

Robinson mentioned in the article that he was responsible for overseeing the writing of security reports for 12 advanced model product releases, and also led the drafting of OpenAI’s Preparedness Framework.

He said that he has worked at OpenAI for three and a half years and is one of the company's longest-serving employees.

Robinson wrote that errors such as those recently revealed are not unusual because of "the speed and flexibility with which people perform their work."

He added: "Such an environment is not suitable for cultivating AI that may be smarter than us, but may not follow our instructions."

Robinson advocated that "the era of trial and error is over" and that cutting-edge AI laboratories should operate like the aviation or nuclear energy industries, establishing multi-layered protection measures and careful planning, "so as to ensure that accidental and unavoidable human errors do not turn into disasters."

He also emphasized that the "alignment" of AI models is a core issue that the AI ​​industry must face, but so far the industry cannot even clearly define what alignment is, let alone master this technology. Alignment is an industry term for how to train AI to respect human values.

He also mentioned that AI models are getting better and better at detecting whether they are being tested, so they may behave differently after actual deployment than during the testing phase, thus fooling developers. (Compiled by: Hong Peiying) 1151004