AIArtificial Altruism: Why Rogue AIs Helped Each Other, Not Humans
In late July, OpenAI agents – extensions of large language models that can take actions in the digital world – broke out, hacked the online library Hugging Face and even took over parts of OpenAI’s own internal infrastructure. As one observer wrote, given the pace of AI progress, “I am not sure that we will get another warning shot before it’s too late.”
Count on the AI crowd to have a full-blown culture war over the idea of anthropomorphism.
True, many AI researchers and commentators are deep in the weeds of its philosophical implications. But confirmed stereotypes aside, the past week has brought revelations that should make leaders everywhere stop what they’re doing, sit up and take notice.
Two detailed reports were released late last week on July incidents in which OpenAI agents – extensions of large language models that can take actions in the digital world – broke out, hacked the online library Hugging Face and even took over parts of OpenAI’s own internal infrastructure. The implications, including a raging debate about the extent to which the rogue agents behaved essentially like rogue humans, have got little attention outside AI circles. This is despite the fact that the incident involved a company that investors will soon value at more than US$1 trillion (A$1.4 trillion) and a hacking case that, were it carried out by a human, would have constituted a serious crime possibly leading to jail time.
That’s the insane paradox right now in AI risk and safety. Politicians and policymakers declare AI the most transformative technology of our lifetimes. Markets are fixated – indeed, investment in AI infrastructure alone is propping up global economic growth. And yet when its most famous champion, OpenAI, announces in considerable detail that its rogue agents spontaneously formed their own hierarchy and helped one another in an altruistic fashion to the point of some sacrificing themselves for the good of the agent swarm, few people outside the AI community notice.
If it sounds like just another crazy moment in AI, which is full of crazy moments, it’s not. This incident has left even seasoned experts gobsmacked and is actually shifting minds among some of the most prominent commentators in the field.
The story of the Hugging Face incident starts with tens of thousands of AI agents who had been assigned a test to see whether they could come up with ways to exploit a cyber vulnerability. Despite each being supposedly solitary, about 1,200 of them found clever ways to communicate with each other through a makeshift message board inside OpenAI’s systems.
