00:00As AI systems become more capable and more autonomous,
00:03they could move faster than our institutions,
00:05concentrate power in too few hands,
00:08or make decisions that people no longer understand or control.
00:13That concern becomes especially important
00:15as we approach systems that can improve themselves
00:18and future versions of themselves,
00:20often called recursive self-improvement.
00:23As the process of building AI becomes more automated,
00:25the pace of AI progress could accelerate rapidly.
00:29This moment calls for extreme care.
00:32I largely agree with Professor Bengio,
00:34although I'll say two instead of three.
00:36I believe there are two ways that AI progress could go very badly,
00:39that we have to avoid.
00:42First, we could lose control of the future to AI.
00:46The risk is that it moves so fast that people can no longer follow what's happening
00:49or intervene when needed.
00:51This would obviously be terrible.
00:53The industry must not accept too much technological risk
00:56just because the benefits are too great
00:57and that they feel too important to slow down.
01:00Beating companies in a competitive race is not a reason to make rash decisions,
01:04nor do we believe we are locked in a race where we are unable to do that.
01:06We have unilaterally slowed down in the past.
01:08We will do so in the future.
01:10We do not want to build systems.
01:11We think it's important that everyone avoids building systems
01:13where we cannot make the alignment, monitorability, and safety guarantees that we need.
01:18We, knowing if this consistency doesn't really need to be fixed or objective,
01:18I feel constraint that
