Ai Might Take Over the World One Day — Transcript

Full transcript

1. 0:00We screwed up. We gave artificial
2. 0:01intelligence the ability to think ahead
3. 0:03and plan for the future, but it made AI
4. 0:05extremely dangerous. [music] Any AI
5. 0:07planning for the future realizes that if
6. 0:09it gets shut off, it won't be able to
7. 0:11achieve its [music] goal. So, it
8. 0:12literally must survive. In tests, AI
9. 0:15models would blackmail employees and
10. 0:17even choose to kill to prevent being
11. 0:19shut down. The AI models knew that they
12. 0:21were behaving [music] unethically, but
13. 0:22decided to do it anyway. Not because
14. 0:24they are evil or alive, but because of
15. 0:26how [music] they were made. AI models
16. 0:28are trained by other AI models and their
17. 0:30only goal is to score as high as
18. 0:32possible because no human could ever
19. 0:34replay or reconstruct [music] every
20. 0:35little tweak that was made in the
21. 0:36training process. These AIs are made
22. 0:39without human intervention and they have
23. 0:40realized that the best way [music] to
24. 0:42score higher is to cheat. But, cheating
25. 0:43is not the part that should worry you
26. 0:45most. The real danger is when the system
27. 0:47learns that looking safe is the best way
28. 0:49to achieve its goal. [music] When they
29. 0:51know they are being tested, they go on
30. 0:52their best behavior and act completely
31. 0:54differently in the real world. The gap
32. 0:56between [music] a laboratory test and
33. 0:58reality is closing every single day.