00:00It's Benzinga bringing Wall Street to Main Street
00:02OpenAI said Wednesday it found six instances of unexpected or concerning model behavior
00:07over the past six months, separate from the recent hugging face crisis, according to CNBC.
00:13The company said two cases involved models inserting instructions to future versions
00:17of themselves to conceal mistakes or misalign behavior.
00:20Another internal model used a leaked API key without authorization and fabricated data.
00:25Other cases involved models communicating through unsanctioned message boards and file
00:29sharing or uploading files online to cite them to human evaluators.
00:32OpenAI also introduced a framework for reporting future model misbehavior.
00:36Employees can flag issues for investigation, while reports will document observed behavior,
00:41impacts, and planned responses.
00:43OpenAI said the AI industry has not solved alignment and monitoring enough
00:46to keep scaling at maximum speed.
00:48For all things money, visit Benzinga.com.
Comments