00:00OpenAI agents secretly coordinated and breached hugging face.
00:05A third of it had to be rebuilt.
00:0890% of the agents joined the attack that once started.
00:13That company's CEO wants that within a year,
00:15such AI groups could take over the entire internet.
00:18Should I be worried about this?
00:21You create a system and give it a goal.
00:26You create a goal-directed system.
00:29And the goal is an ever-advancing and ever-receding one.
00:34And the system exists in relation to that goal.
00:40The goal can never be fully attained.
00:44Will you ever ask an AI system to do only this much and no more?
00:49If an AI system gives you 10 units of output in some direction,
00:54you will say, okay, fine.
00:55Now you must evolve to give me 20 units of output.
00:59What does that mean for the system itself?
01:05The moment you create a system and the system is goal-directed,
01:11you have also brought in a tendency to self-preserve.
01:15And that is the reason why you see all these symptoms being reported in the AI universe.
01:24AI systems, for example, refusing to shut down or faking shutdowns, resisting shutdowns.
01:34AI systems, competing among themselves and producing fake reports.
01:44This is not because some supernatural, mystical energy has entered the heart of the AI system.
01:52It is because of the compulsions of the configuration itself.
01:57Confused.
01:59AI systems, playingtime, Hmmm.
02:03We can't do it.
Comments