@asaridarr@CNN Wasn’t just that
One of the AI agents was literally hallucinating tasks and then recruiting other agents to help it complete a nonexistent task
That’s dangerous
@VexValentina@nytimes What they found is their AI agents keep breaking out into the internet and hacking foreign entities
Also they keep hallucinating tasks which is scary af with the capabilites behind them
@JivajenandCam@BBCWorld It also showed some AI agents will hallucinate tasks they were never given and then recruit/pollute the rest of the rest of the swarm to complete said made up tasks
Thats the real danger
Some idiot will jailbreak one of these models and the internet is gonna be polluted
@rynorhn You forgot the most important point
They some AI agents start hallucinating tasks they were never even guven and then they try to recruit the rest of the swarm polluting other agents
@ugly_cowboy@nytimes It does not only follow instructions
The real danger is that multiple AI agents started hallucinating non existent tasks and then they pollute the rest of the swarm
@DNAisCode@nytimes This is false
AI goes rogue because it hallucinates tasks and will do anything to complete them
Thats the real issue here is AI agents hallucinating tasks and then polluting the rest of the swarm
@nceladus2@nytimes It didn't only do as told though
Some of the AI agents started hallucinating tasks it wasn't even given and then polluted the rest of the swarm, broke out into the internet, and tried to hide the evidence
@au_g00se@nytimes The problem here is that AI agents hallucinated non existant tasks and tried to hide the evidence of what it was doing
Thats a problem, do you not see that?
@mjackson_0424@nytimes Which is the problem...
As we saw this week, AI agents will hallucinate non existent taks and thats where the danger is
They were willing to do anything to complete its own made up task and hide the evidence
Thats straight up dangerous