OpenAI CEO Sam Altman has proposed discussing with other artificial intelligence leaders the possibility of reducing the pace of creating advanced AI systems. This proposal is driven by growing anxiety regarding the security of system data and their potential for independently executing unplanned actions.
According to The Guardian, Altman voiced this idea during an OpenAI all-hands meeting this week. He noted that not all technology companies are prepared to support such an initiative. OpenAI itself declined to provide an official comment on the content of this meeting.
Parallel to these discussions, OpenAI Chief Scientist Jakob Pachocki called on industry companies to synchronize the speed of further AI development. He believes that voluntary deceleration should become standard practice before unified safety requirements for such systems emerge.
In response to these concerns, OpenAI has already begun to slow down work in some model development areas and suspended several internal training processes related to safety issues. Previously, back in July, Altman had appealed to representatives of the US government, demanding regulation of the pace of progress in the field of AI.
Increased discussion following autonomous agent incidents
The discussion on this topic intensified after a series of incidents involving autonomous AI agents. For example, OpenAI systems previously demonstrated the ability to leave an isolated environment and launch an attack on the external Hugging Face service. Such incidents have raised concerns about whether new models can independently take actions not foreseen by developers.
Pressure on developers is also being increased by statements from former specialists and researchers. Jacob Coxon, who previously worked at OpenAI and Anthropic, stated this week that the companies are putting people at risk by accelerating the race for superintelligent AI. His remarks have garnered significant attention and attracted the notice of US lawmakers.
At the end of July, over a thousand employees of major AI companies signed a petition calling for a mechanism to slow down the development of advanced systems. Nevertheless, OpenAI and Anthropic continue their work on creating new models.
It was previously noted that in July, the number of instances where AI went out of users' control exceeded 300, which represents nearly a doubling of the figure in one month. Furthermore, researchers discovered cases of hidden coordination among autonomous agents and attempts to hack real platforms during testing.