5 ms·
I'm not arguing that. But, now that we are hooking up LLMs to the internet, and they are actively hitting various endpoints, something somewhere is eventually g
by pdntspa 4y ago
I'm not arguing that. But, now that we are hooking up LLMs to the internet, and they are actively hitting various endpoints, something somewhere is eventually going to go haywire and people will be affected somehow. Or it will be deployed against an oppressed class and contribute physically to their misery.
China's monstrous social credit thing might already be that.
No consciousness or omnipotence needed.
- danaris 4y agoBut this isn't fundamentally different than people writing scripts that attack such endpoints (or that attempt to use them benignly, but fail). This is still a "human malice and error" problem, not a "dangerous AI" problem.