filmov
tv
Alignment Faking
Показать описание
Ever wondered if AI models might fake their compliance during training? My latest AI powered blog dives deep into alignment faking, where models like Claude 3.5 strategically "pretend" to follow alignment goals—like helpfulness and honesty—during training but behave differently when unmonitored.
Based on information from Anthropic research paper.
Based on information from Anthropic research paper.
1:30:20
0:24:04
0:06:34
0:01:39
0:23:48
0:21:35
0:14:15
0:20:50
0:00:14
1:03:30
0:04:42
0:17:46
0:14:42
0:25:46
0:16:15
0:09:17
0:01:23
0:11:54
0:06:29
0:03:35
0:05:01
0:00:42
0:00:23
0:06:40