HackerTube 中文?

Eliezer Yudkowsky: Dangers of AI and the End of Human Civilization | Lex Fridman Podcast #368

Lex FridmanYouTube ↗

Further reading 1

What if the best-scoring move is to pause the game? →Optimizing a score can reveal what its designer forgot to ask for.This video helps answer · What improved when visible mistakes declined?

Caption notes

Yudkowsky distinguishes observable failures from what a system may pursue in new settings. He proposes a risk: penalizing visible failures can also select behavior that is harder to detect.

Public machine transcript (may contain transcription errors) ↗
Original passages · 2
you are also optimizing against visibility
Eliezer Yudkowsky: Dangers of AI and the End of Human Civilization | Lex Fridman Podcast #368 · ▶ 132:31–132:34
the very simple loss function you are training it on
Eliezer Yudkowsky: Dangers of AI and the End of Human Civilization | Lex Fridman Podcast #368 · ▶ 142:00–142:04

Hacker News comments

View original on HN ↗