I noticed that the comments regarding AI safety (e.g. on Twitter, YouTube, HackerNews) often contain sentiments like:
"there's no evidence AGI will be dangerous" or
"AGI will do what we intend by default" or
"making AGI safe is not a real or hard problem" / "it's obvious, just do X"
I had a thought that it might make a compelling case for 'AI safety is real and difficult' if there were something akin to the Millennium Prizes for it. Something impressive and newsworthy. Not just to motivate research, but also to signal credibility ("if it's so easy, why don't you solve the X problem and claim the prize?").
I was wondering if there is/are already such prizes, and if not, what problem(s) might be good candidates for such prize(s)? I think ideally such problems would be:
possible for most people to understand at a high level (otherwise it wouldn't make a compelling case for AI safety being 'real')
solution can be verified unambiguously (to avoid claims of moving goal posts or not being a real problem)
easy to motivate its usefulness / necessity towards AGI; if we can't solve this, then it's obvious we can't make safe AGI
a difficult problem, like one where progress has been slow / non-existent, or we have no idea where to start (a problem that is solved after a few hours doesn't make AI safety much more credible)
I noticed that the comments regarding AI safety (e.g. on Twitter, YouTube, HackerNews) often contain sentiments like:
I had a thought that it might make a compelling case for 'AI safety is real and difficult' if there were something akin to the Millennium Prizes for it. Something impressive and newsworthy. Not just to motivate research, but also to signal credibility ("if it's so easy, why don't you solve the X problem and claim the prize?").
I was wondering if there is/are already such prizes, and if not, what problem(s) might be good candidates for such prize(s)? I think ideally such problems would be: