Saunders: AI-assisted critique helps humans detect model errors
William Saunders · What The Ex-OpenAI Safety Employees Are Worried About · Jul 3, 2024 · at 8:13
Former OpenAI Superalignment researcher William Saunders explains technical oversight methods developed at OpenAI to help humans evaluate AI answers beyond their domain expertise.
0:00 / 0:36exact quote · 36.5s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“What some of the research that I did at OpenAI was, you know, trying to develop techniques for this, and the simple technique that we tried and showed that worked was just ask a different AI system, or even the same AI system, is there a problem with this answer you gave me? And then the AI will sometimes put out a list of problems and say like, hey, this part of the answer was like made up. It's not supported by any evidence, or you're leaving out this important piece of information. And then we showed that if you just, if you show people both an answer and a set of possible problems, people are better at spotting when the AI system has made a mistake.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from William Saunders
Saunders: Many at OpenAI discuss a three-year timeline to AGI
“And you know, what people were talking about at the company in terms of timelines to something dangerous were like, there were people talking, a lot of people talking about similar things to like the predictions of like, Leopold Aschenbrenner, where it's like …”
Saunders: OpenAI prioritizes shiny products over safety like the Titanic
“OpenAI claimed that their mission was to build safe and beneficial AGI, and I thought that this would mean that they would prioritize, you know, putting safety first. But over time, it started to really feel like the decisions being made by leadership were mor…”
Saunders: Departure from OpenAI was not caused by witnessing immediate harm
“If there were, if there was a group of people that I knew were being seriously harmed by this technology first, I still really hope that open AI would like do the right thing and address this. If this was like a very clear cut case. I also personally would you…”
Saunders: GPT-4 is safe, but GPT-5 or later might be the Titanic
“I don't think that I was working on the Titanic. I don't think that GPT four was the Titanic. I'm more, I'm afraid that like GPT five or GPT six or GPT seven might be the Titanic in, in, in this analogy.”
Saunders: Small model oversight of large models clearly misses violations
“Some systems that the company has talked about involve using a very like small and dumb language model to like monitor what a larger language model is doing. And this will like clearly miss a bunch of things.”
Saunders: California SB 1047 creates a superior model for AI whistleblowing
“I think that like a model that I think might work, you know, I think I personally think would work better would be like, The model more proposed in California Senate bill, 1047, where there would be like a, you know, the, where the law would create like, you k…”