Yann LeCun: Human-Level Artificial Intelligence | AI Podcast Clips
00:00:00.000 | - What do you think it takes to build a system
00:00:04.040 | with human level intelligence?
00:00:05.380 | You talked about the AI system in the movie "Her"
00:00:08.900 | being way out of reach, our current reach.
00:00:11.360 | This might be outdated as well, but--
00:00:13.680 | - It's still way out of reach.
00:00:14.520 | - It's still way out of reach.
00:00:16.000 | What would it take to build "Her"?
00:00:19.660 | Do you think?
00:00:21.040 | - So I can tell you the first two obstacles
00:00:23.040 | that we have to clear,
00:00:24.160 | but I don't know how many obstacles there are after this.
00:00:26.120 | So the image I usually use is that
00:00:27.920 | there is a bunch of mountains that we have to climb
00:00:29.920 | and we can see the first one,
00:00:31.000 | but we don't know if there are 50 mountains
00:00:32.600 | behind it or not.
00:00:33.440 | And this might be a good sort of metaphor
00:00:36.200 | for why AI researchers in the past
00:00:39.640 | have been overly optimistic about the result of AI.
00:00:43.280 | You know, for example, Newell and Simon, right,
00:00:48.160 | wrote the general problem solver
00:00:50.680 | and they called it a general problem solver.
00:00:52.720 | - General problem solver.
00:00:53.560 | - Okay, and of course, the first thing you realize
00:00:55.840 | is that all the problems you want to solve are exponential
00:00:57.640 | and so you can't actually use it for anything useful.
00:01:00.440 | But, you know.
00:01:01.360 | - Yeah, so yeah, all you see is the first peak.
00:01:03.560 | So what are the first couple of peaks for "Her"?
00:01:06.560 | - So the first peak,
00:01:07.700 | which is precisely what I'm working on,
00:01:09.280 | is self-supervised learning.
00:01:11.080 | How do we get machines to learn models of the world
00:01:13.560 | by observation, kind of like babies and like young animals?
00:01:17.120 | So we've been working with cognitive scientists.
00:01:25.080 | So this Emmanuelle Dupou, who is at FAIR in Paris,
00:01:28.080 | half-time, is also a researcher in French University.
00:01:33.920 | And he has this chart that shows
00:01:38.200 | which, how many months of life baby humans
00:01:41.880 | can learn different concepts.
00:01:44.040 | And you can measure this in various ways.
00:01:46.960 | So things like distinguishing animate objects
00:01:52.680 | from inanimate objects.
00:01:54.240 | You can tell the difference at age two, three months.
00:01:57.040 | Whether an object is going to stay stable,
00:02:00.320 | is going to fall, you know,
00:02:01.640 | about four months, you can tell.
00:02:04.760 | You know, there are various things like this.
00:02:06.460 | And then things like gravity,
00:02:08.280 | the fact that objects are not supposed to float in the air,
00:02:10.440 | but are supposed to fall,
00:02:11.880 | you learn this around the age of eight or nine months.
00:02:14.420 | If you look at a lot of, you know, eight-month-old babies,
00:02:17.160 | you give them a bunch of toys on their high chair.
00:02:20.360 | First thing they do is they throw them on the ground
00:02:21.840 | and they look at them.
00:02:23.040 | It's because, you know, they're learning about,
00:02:25.240 | actively learning about gravity.
00:02:27.400 | - Gravity, yeah.
00:02:28.240 | - Okay, so they're not trying to annoy you,
00:02:31.000 | but they, you know, they need to do the experiment, right?
00:02:33.960 | So, you know, how do we get machines to learn like babies?
00:02:37.880 | Mostly by observation with a little bit of interaction
00:02:40.520 | and learning those models of the world,
00:02:42.520 | because I think that's really a crucial piece
00:02:45.040 | of an intelligent autonomous system.
00:02:47.640 | So if you think about the architecture
00:02:48.840 | of an intelligent autonomous system,
00:02:50.800 | it needs to have a predictive model of the world.
00:02:52.640 | So something that says, here is a world at time T,
00:02:55.360 | here is a state of the world at time T plus one
00:02:56.800 | if I take this action.
00:02:57.940 | And it's not a single answer, it can be a--
00:03:01.000 | - Yeah, it can be a distribution, yeah.
00:03:02.560 | - Yeah, well, but we don't know how to represent
00:03:04.520 | distributions in high-dimensional continuous spaces,
00:03:06.160 | so it's gotta be something weaker than that, okay?
00:03:08.520 | But with some representation of uncertainty.
00:03:11.040 | If you have that, then you can do
00:03:13.920 | what optimal control theorists call
00:03:15.760 | model predictive control, which means that
00:03:17.400 | you can run your model with a hypothesis
00:03:20.200 | for a sequence of action and then see the result.
00:03:23.160 | Now, what you need, the other thing you need
00:03:24.560 | is some sort of objective that you want to optimize.
00:03:27.340 | Am I reaching the goal of grabbing this object?
00:03:30.080 | Am I minimizing energy?
00:03:31.360 | Am I whatever, right?
00:03:32.520 | So there is some sort of objectives
00:03:34.800 | that you have to minimize.
00:03:36.240 | And so in your head, if you have this model,
00:03:38.040 | you can figure out the sequence of action
00:03:39.560 | that will optimize your objective.
00:03:41.260 | That objective is something that ultimately
00:03:44.640 | is rooted in your basal ganglia,
00:03:46.920 | at least in the human brain, that's what it is.
00:03:48.520 | Basal ganglia computes your level of contentment
00:03:52.160 | or miscontentment, I don't know if that's a word.
00:03:55.320 | Unhappiness, okay?
00:03:56.680 | - Yeah, yeah.
00:03:57.920 | - Discontentment.
00:03:58.760 | - Discontentment, maybe.
00:03:59.720 | - And so your entire behavior is driven towards
00:04:03.640 | kind of minimizing that objective,
00:04:06.240 | which is maximizing your contentment,
00:04:08.720 | computed by your basal ganglia.
00:04:10.600 | And what you have is an objective function,
00:04:14.560 | which is basically a predictor of what
00:04:16.360 | your basal ganglia is gonna tell you.
00:04:18.440 | So you're not gonna put your hand on fire
00:04:20.400 | because you know it's gonna burn
00:04:23.560 | and you're gonna get hurt.
00:04:25.040 | And you're predicting this because of your model of the world
00:04:27.360 | and your sort of predictor of this objective, right?
00:04:31.400 | So if you have those three components,
00:04:34.800 | you have four components,
00:04:36.440 | you have the hardwired contentment objective
00:04:40.600 | computer, if you want, calculator.
00:04:45.200 | And then you have those three components.
00:04:46.400 | One is the objective predictor,
00:04:48.000 | which basically predicts your level of contentment.
00:04:50.200 | One is the model of the world.
00:04:53.800 | And there's a third module I didn't mention,
00:04:55.360 | which is the module that will figure out
00:04:58.520 | the best course of action to optimize an objective
00:05:01.800 | given your model.
00:05:02.640 | Okay?
00:05:04.840 | - Yeah.
00:05:05.720 | - Glissa policy, policy network,
00:05:08.560 | or something like that, right?
00:05:10.720 | Now, you need those three components
00:05:13.000 | to act autonomously, intelligently.
00:05:15.240 | And you can be stupid in three different ways.
00:05:17.400 | You can be stupid because your model of the world is wrong.
00:05:20.680 | You can be stupid because your objective is not aligned
00:05:23.800 | with what you actually want to achieve.
00:05:26.360 | Okay?
00:05:27.200 | In humans, that would be a psychopath.
00:05:30.440 | - Right.
00:05:31.320 | - And then the third thing,
00:05:33.800 | the third way you can be stupid
00:05:34.920 | is that you have the right model,
00:05:36.240 | you have the right objective,
00:05:37.640 | but you're unable to figure out a course of action
00:05:40.120 | to optimize your objective given your model.
00:05:42.360 | (silence)
00:05:44.520 | (silence)
00:05:46.680 | (silence)
00:05:48.840 | (silence)
00:05:51.000 | (silence)
00:05:53.160 | (silence)
00:05:55.320 | (silence)
00:05:57.480 | [BLANK_AUDIO]