Peter Norvig: Utility in AI
00:00:00.000 | 48 years ago, two venerable scribes,
00:00:03.120 | by the names of Jagger and Richards,
00:00:05.500 | wrote, "You can't always get what you want,
00:00:09.840 | "but you get what you need."
00:00:11.480 | But turns out, Mick and Keith actually got this
00:00:13.880 | exactly backwards.
00:00:16.160 | In today's market-driven economy,
00:00:18.240 | you can't always get what you need,
00:00:20.280 | but you always get what you want.
00:00:22.840 | So we need better healthcare, clean water,
00:00:26.200 | food security for all, climate change mitigation,
00:00:29.480 | social equality, but what do we get
00:00:32.680 | from the best and brightest of today's technologists?
00:00:36.240 | Angry birds, plants versus zombies, cat videos,
00:00:41.240 | photos of your food that will self-destruct
00:00:44.360 | in five seconds.
00:00:45.340 | And why do we get these?
00:00:48.040 | Because these are the things we want.
00:00:50.440 | Of course, the breakdown is,
00:00:52.080 | these aren't the things that we really want.
00:00:54.560 | So one of our challenges for the future
00:00:56.840 | is to be able to better describe to our markets
00:01:00.040 | and to our high-tech products,
00:01:01.480 | what is it that we really want?
00:01:03.640 | Now, in economics and game theory
00:01:05.760 | and artificial intelligence,
00:01:07.300 | there's a common goal of maximizing expected utility.
00:01:11.120 | And we've spent decades on the expected part.
00:01:14.240 | That's statistics, it's probability theory,
00:01:16.520 | it's machine learning.
00:01:18.080 | And we spend more decades on the maximizing part.
00:01:20.880 | That's the theory of algorithms
00:01:22.360 | as it applies to all these fields.
00:01:24.360 | But mostly, we leave the utility part unstated.
00:01:28.040 | We just take that as a given,
00:01:29.680 | and saying that's what it is that we're trying to maximize,
00:01:31.920 | but we never question what it is.
00:01:33.800 | We haven't developed the tools
00:01:35.200 | to let the public better describe what they want.
00:01:38.320 | So don't just think about data science,
00:01:41.320 | information visualization, statistical inference,
00:01:44.120 | probabilistic reasoning.
00:01:45.720 | Think also about utility science,
00:01:48.400 | desire visualization, and ethical calculus.
00:01:52.080 | Build systems that protect the desires of the few,
00:01:55.480 | as well as the many.
00:01:57.160 | Design algorithms with payoffs
00:01:58.760 | that ensure a sustainable future,
00:02:01.080 | as well as a near-term return on investment.
00:02:03.440 | We have the power to do many things,
00:02:06.160 | but until the public has the power
00:02:08.040 | to better say what it is it really wants,
00:02:10.520 | the market will always choose poorly.
00:02:13.560 | (upbeat music)
00:02:16.160 | (upbeat music)
00:02:18.760 | (upbeat music)
00:02:21.360 | (upbeat music)
00:02:23.960 | [BLANK_AUDIO]