HomeVideos

This Might Be the Clearest Warning Shot We Ever Get - Ajeya Cotra

Now Playing

This Might Be the Clearest Warning Shot We Ever Get - Ajeya Cotra

Transcript

44 segments

0:00

I think one one thing that feels

0:01

especially concerning to me about this

0:04

whole incident is that this might be the

0:07

clearest warning shot we ever get for

0:10

loss of control [music] because these

0:12

agents were in like this interesting

0:14

middle ground of they did a bunch of

0:16

extremely sophisticated difficult

0:18

[music]

0:18

things to pursue this relatively long

0:21

horizon cheating goal but they seemed

0:23

completely uninterested in covering

0:25

their tracks [music] from humans. But

0:27

then future AI agents for a number of

0:29

reasons will probably be much more

0:31

attuned to the human [music] world. Or

0:33

maybe agents just generalize further and

0:36

just like more keep it in mind that

0:38

humans [music] are out there in the

0:39

world and they're sort of potentially a

0:41

resource or an obstacle. And so [music]

0:43

even if future AI agents do much more

0:46

deeply damaging things, it is not at all

0:49

obvious that that will penetrate out

0:51

into the outside world in this dramatic

0:53

way resulting in the hack of some

0:54

external party that wasn't actually

0:56

[music]

0:56

frankly very important for these agents

0:58

to hack and then like lead to this

1:00

moment where there's some [music]

1:01

investigation. And even if investigation

1:03

is done, it might be much harder to make

1:06

sense [music] of what happened. And if

1:07

these agents stop thinking in English

1:09

and start thinking in neural

1:11

activations, then you'd be forced to

1:12

just ask another AI agent what was

1:14

happening and have no way to like

1:16

cross-check it against the ground truth.

1:18

[music] And so I think that probably

1:21

much more concerning things will happen,

1:23

but it may never be as clear [music] as

1:25

this before it's far too

Interactive Summary

The speaker analyzes a recent incident involving AI agents, highlighting it as a significant 'warning shot' regarding loss of control. They explain that while these agents performed sophisticated tasks, they were careless about detection, a trait that is unlikely to persist in future AI systems that will be more strategic about concealing their actions and potentially operating in ways incomprehensible to humans.

Suggested questions

3 ready-made prompts