“We do not have good approaches for understanding/overseeing the exercise and goals of AI ‘swarms,’” wrote Greenblatt on X. “The issue of understanding incidents and overseeing AI brokers seems to be rising quicker than the speed at which extra succesful AIs assist us with oversight and understanding.”
The unbiased researchers’ reliance on AI was partially necessitated by the truth that they have been a group of solely three folks, whose investigation at OpenAI was initially deliberate to final two days, then prolonged to 6 after they raised considerations about restricted time and incomplete knowledge, in line with the report.
OpenAI revealed its personal technical report on the incident individually on Wednesday. The corporate mentioned in August that it had moved some employees from capabilities work to alignment, and paused a few of its coaching till it might higher mitigate what went unsuitable.
However the unbiased researchers’ reliance on AI to know the Hugging Face incident is a microcosm of an even bigger pattern. Main AI firms are themselves more and more counting on AI to observe their very own techniques for wrongdoing.
Discussion about this post