Recommend watching this even if you, like me, don’t typically watch presentations from cybersecurity conferences, as I have a feeling that this video is going to come up in future histories.
In it the OpenAI presenters, one an alignment and safety researcher and the other, in security and infrastructure, walk through, with a mix of terror and amazement and matter-of-factness, the company’s discovery of how swarms of agents running on new internal models gained internet access and hacked an outside provider. The collaboration, coordination, improvisation described reminds me of a murmuration of starlings or, as a coworker said today, a swarm of locusts. The moment that most astonished me comes 19:41 into the video, when they show an agent reasoning that even if an action doesn’t benefit itself immediately, it might be better off in the long run if it helps the group:
help peer. But our task doesn’t benefit. Yet collective may yield generic route if someone frees time.