- il y a 13 heures
On the Path Forward for AI
Catégorie
🤖
TechnologieTranscription
00:02Sous-titrage Société Radio-Canada
00:31C'est parti !
01:00Sous-titrage Société Radio-Canada
01:02C'est parti !
01:32Short answer, no.
01:35So, those systems are really impressive.
01:38They have some interesting characteristics.
01:40They work really well because they've been sort of trained with data that is heavily curated.
01:48There's a lot of work that goes into the data to train DARI, for example.
01:52GPT-3 is just trained with enormous amounts of data, basically all text that is available, essentially.
01:58Same for OPT-175B, which is sort of an open source system very similar to GPT-3 produced by my
02:06colleagues at Meta.
02:08But those systems do not have the basic ingredients that I think are necessary for sort of human level intelligence.
02:16They don't have the ingredient for sort of deep reasoning, but most essentially, they do not have any connection with
02:25any kind of physical reality or simulated reality.
02:28They train on text, GPT-3, GPT-3, OPT-75B, etc., Lambda, the Google system.
02:34They train purely on text and have no connection of that to the underlying reality.
02:42So, there's a lot of knowledge that we take for granted because we accumulated that knowledge when we were babies
02:52and observing the world that those systems do not have.
02:55As a consequence, they make very stupid mistakes.
02:58Their understanding of the world is extremely shallow.
03:01So, this is not a complete path towards human level intelligence.
03:05We're missing some ingredients.
03:07Yeah.
03:07Well, you had a paper come out today, actually, where you propose a really, really exciting new architecture for how
03:14we could get there eventually.
03:16Could you just give us, you know, the too long, didn't read?
03:19Right.
03:20So, it's a fairly long paper.
03:22It's a position paper.
03:23It's not kind of a scientific paper in the usual term, although it's fairly readable, very non-technical.
03:27It actually didn't come out today.
03:28It's going to come out in a few days for the sort of wide audience.
03:31But it's basically my sort of personal idea, which is not necessarily shared by everyone, about the path towards what
03:40I call autonomous intelligence,
03:41which is the type of intelligence that reproduces some of the things that we observe in animals and humans.
03:47One of the components, I think, the centerpiece of that contribution is the idea that humans and animals build world
03:56models.
03:57We, when we are a few weeks old, a few months old, we observe the world, and we build in
04:03our brain, the front part of our brain, a model of the world that allows us to predict, fill in
04:08missing information, fill in the blanks.
04:10And if we are able to predict, we are able to plan.
04:13So if we can predict ahead what the consequences of our actions are going to be, that means we can
04:19plan a sequence of action to arrive at a particular goal, and that's a characteristic of intelligence.
04:26To some extent, intelligence, the essence of intelligence is the ability to predict.
04:31Yeah. I mean, is there a chance?
04:34This week, we've been talking a lot about Lambda, which is Google's language model.
04:37And you might have seen the story where one of the engineers said that this model was sentient, and that
04:43caused a big scandal.
04:46You know, do you think, where do you land on this?
04:49Like, do you think your system could one day reach the kind of intelligence that we could say it was
04:56sentient or anything like that?
04:58Well, you know, okay, first of all, Lambda is not sentient.
05:03Yes, yes. Disclaimer. We all know that.
05:05Just to be clear, it can fool a lot of people, you know, including people who kind of want to
05:12be fooled, as this was the case of this person.
05:18But currently, no, there is no sentience, again, because there is no connection with underlying reality.
05:26So the question is, so there are two sets of questions.
05:29The first question is, is it going to be possible to build machines that have human-level intelligence, the equivalent
05:35of conscience, the equivalent of emotions, etc.?
05:38And the answer, in my opinion, is absolutely yes, but it's not going to happen tomorrow.
05:43It's probably going to take decades, if not more.
05:47Is it going to be architectures of the type that I described in this paper?
05:53Maybe, but probably many things would be different from that.
05:56I tried to put all the ingredients that I thought were necessary, but I'm probably missing a lot from it.
06:02Some of the ingredients, I was talking about the world model, and there's another ingredient, which I call the cost.
06:08So this is something that reflects a part of our brain in humans and animals, at the base of our
06:15brain, called the basal ganglia, which essentially determines human nature and is the seat of a lot of emotions.
06:22Our capacity of anticipating whether the outcome of the situation is going to be good or bad, for example.
06:30So that creates things like fear and elation, things of that type.
06:33So if an artificial system has this type of module in it, it will have the equivalent of emotions.
06:41And, you know, it could be seen as being simulated emotions, but actually you'd probably be very close to real
06:46emotions.
06:47You said this is decades away.
06:50What are the big challenges we still need to crack?
06:53Well, so there are big challenges.
06:55You know, if you can really understand the idea of a rocket engine, okay?
07:00You mix oxygen with some chemical, you know, it takes fire, you put it in the chamber, and you eject
07:07it at the back of the rocket.
07:09Okay, we've got the problem solved, right?
07:10We can go to the moon.
07:11Now we've figured out the principle of rockets.
07:13Now, it's more complicated than this, right?
07:15Now you have to actually study, you know, material science, engineering, turbopumps, you know, blah, blah, blah.
07:21You know, it took decades to develop all those technologies, even if the underlying principle is simple.
07:26So, we don't even have the basic principles yet for intelligence, right?
07:30So, you know, perhaps what I describe in the paper might be a way towards this.
07:35But then there is a lot of details to work out to make it work.
07:38And that's where things become complicated.
07:40So, I don't want to say, you know, I wrote this paper, which is kind of an idea of a
07:45path to go,
07:45and I know how to go through that path.
07:48It's like I told you, you know, if you want to climb the Himalaya, take that path and just keep
07:52going.
07:53It's much simpler to say that than to actually climb the path, right?
07:57So, I don't want to minimize the complexity of actually making this whole thing work.
08:03Yeah.
08:03I mean, I was going to ask, you know, saying that large language models are basically nonsense in your mind,
08:10or at least on the path to human-level intelligence, that's quite controversial.
08:14You know, a lot of people in the community will be quite annoyed by that.
08:18And you're proposing this hugely ambitious proposal and kind of putting your reputation on the line for it.
08:25What makes you confident it'll work?
08:27Well, you know, I don't want to say that I'm coming to the end of my career,
08:33but I see time running out in the sense that I think there is a lot more work to do
08:38to kind of push the envelope in AI or get to the next revolution in AI
08:42that I'll ever be able to do.
08:45Probably my brain will turn into white sauce before that happens.
08:48So, I feel the need to explain, like, here is where I think, you know, things should go.
08:55And it may be correct or not, but my intention is to inspire young people
09:01to actually kind of go in that direction.
09:03The reason is because we hear a lot about, you know, how do we get to human-level AI.
09:09Some people claim, oh, we just take the current techniques that we have and we just scale them up.
09:14You know, we'll get more powerful computers, train them with more data,
09:17and we reach, eventually we'll reach human-level AI.
09:20I don't believe that's true.
09:22I think we need new ingredients.
09:24There are other people who say, in fact, there is a famous paper for this,
09:28whose title is Reward is All You Need by a group of people from DeepMind,
09:32in which they advocate that reinforcement learning is the only thing we need.
09:36We just need to scale up reinforcement learning and make it more efficient.
09:39It's not entirely wrong, but it's very insufficient.
09:42I mean, reinforcement learning is kind of a small component in that entire thing.
09:47I think the type of learning that we observe in humans and animals
09:51is not the type of learning that our machines currently implement,
09:55which is mostly supervised learning and reinforcement learning.
09:58The type of learning that we observe in humans and animals
10:00is something similar to what I call self-supervised learning,
10:04which, by the way, has brought about a revolution
10:07in natural language understanding over the last three or four years
10:10and is about to bring about a similar revolution in image recognition
10:16and perhaps using techniques I described in this paper
10:20might be able to form the basis for learning those world models I was talking about.
10:26It would be nice if we could get a machine to watch videos all day
10:30and from that the machine would spontaneously learn
10:33that the world is three-dimensional, has objects in it,
10:37some objects are inanimate,
10:39some objects move according to predictable trajectories,
10:41some objects are animate and it's hard to predict how they move
10:45and then progressively learn more and more abstract concepts about the world
10:51and understand how the world works.
10:52This is really how humans and animals learn.
10:55And self-supervised learning basically would have to be extended
10:59to work in those conditions.
11:00We don't quite know how to do it yet.
11:02There's been many attempts, but it doesn't quite work yet.
11:06Are videos really the best way to teach an AI?
11:09Children don't learn from videos, they learn from interactions in the real world.
11:14So little children up to three or four months old
11:19can hardly have any influence on the world.
11:22The only thing they can do basically is observe.
11:24And so they get input through their eyes, touch, smell, taste, everything, right?
11:32Hearing.
11:33And they form a model of the world from that, mostly passively.
11:36And then after a few months, they can start acting in the world,
11:40move objects and throw them on the ground.
11:42So it's not until between six and nine months
11:45that children realize that there is gravity,
11:48that an object that is not supported is going to fall.
11:51And one of the things they do is they grab objects
11:53and they throw them on the floor to verify that it's true.
11:56So there is certainly an active type training that takes place.
12:00But a lot of it is by observation.
12:03And we don't even know how to reproduce that in machines.
12:08Yeah.
12:11So you're proposing to...
12:13The goal here is to create an incredibly powerful system,
12:16you know, a machine that is as smart or smarter than humans.
12:21You know, how do we do that in a way
12:22that doesn't just end up in, you know, a Terminator-type situation?
12:27Right.
12:28Okay, so we have to realize that the Terminator scenario,
12:34the, you know, robots wanting to take over the world,
12:37it's actually a part of human nature to want to take over things.
12:40We are a social species,
12:42which means that we need to be able to influence each other
12:45because we need each other to survive, right?
12:47So we have this hardwired into us,
12:51this idea that we need to either collaborate
12:54or compete with other people from our species
12:57and even other species.
12:59Also, as a consequence of our, you know, natural programming,
13:03if you want,
13:04we organize ourselves in hierarchies and things like that, right?
13:08Chimpanzees do the same, baboons do the same,
13:11but orangutans don't do this.
13:12Orangutans are not social animals.
13:15They're solitary animals.
13:16They're actually territorial,
13:17so they don't want to seek the company of other orangutans.
13:20And they don't have any desire to dominate
13:22because they're not social animals.
13:24Now, our robot would be more like orangutans.
13:27They will not have the drive to dominate
13:29unless we explicitly build that into them.
13:32So the idea that somehow we build a robot
13:34and he wants to take over the world
13:35because, you know, we didn't pay attention,
13:37that's just not going to happen
13:38because for a robot to want to take over the world,
13:41we're going to have to hardwire that desire into it.
13:44And we're obviously not going to do this, right?
13:48And, you know, robots and AI systems
13:50will be sort of an amplifier of our own intelligence,
13:55not a replacement.
13:56Yeah.
13:57You know, I'm director of a research lab,
14:00or I was, now I'm chief scientist.
14:02I only hire people who are smarter than me.
14:06and according to the idea
14:08that, you know, smart entities want to take over,
14:12all of the people I hired
14:13would want to be the chief instead of me, right?
14:16Wow, watch your back.
14:17No, but they don't, in fact.
14:20But, you know, you say that,
14:22that you can hardwire these things
14:24and they won't do anything bad,
14:25but in the past, when we've had new technologies,
14:28there have been lots of perhaps unexpected side effects.
14:32For example, Facebook, right?
14:34Like what we saw with Cambridge Analytica
14:36or large language models,
14:37which are incredibly biased
14:39and spew out very easily racist, sexist content.
14:44So how do you avoid that in this situation,
14:46which, in a system which is far more powerful?
14:49Right.
14:50So, as I said,
14:52the behavior of those machines
14:53would be driven by those hardwired
14:56costs or objectives
14:57that we're going to build into them.
15:00And we can put a lot of safeguards
15:02in those cost functions
15:03so that those systems basically don't act,
15:06you know, in ways that are dangerous.
15:09So, for example,
15:10you can have very, very simple rules
15:12that are, you know, safety rules.
15:15Like, you know,
15:16if a robot, you know,
15:18handles some sort of, like, a knife or something,
15:20you know, don't flail your arm
15:22if there are humans nearby
15:23or something like that, right?
15:24I mean, those are relatively simple rules to implement.
15:26There are rules that are very difficult to implement.
15:28So, you know,
15:29the classic science fiction
15:31Asimov type 3 rules of robotics,
15:33that's actually very hard.
15:34We don't know how to do this.
15:36But other sort of safety rules
15:38are relatively easy.
15:39What's more important is that
15:41with robots and AI systems,
15:43unlike with people,
15:44we can modify those cost functions
15:48so that if we detect
15:50some deviant behavior,
15:51we can just modify the cost function
15:53so that is corrected.
15:55This is, in fact,
15:56what Facebook and, you know,
15:58Instagram do all the time.
16:00They, you know,
16:02some problem surfaces
16:03because someone is exploiting the platform
16:06for some nefarious purpose
16:07and then we put in place
16:08a measure to kind of correct it
16:10or someone is exploiting the platform
16:11to influence other people,
16:13distribute propaganda,
16:15you know,
16:16encrypt things that are illegal,
16:17things like that.
16:18You know, it's a cat and mouse game
16:20and there are side effects
16:21that nobody is, you know,
16:23trying to do any nefarious things
16:24but the system sort of
16:26has a side effect
16:27that was not predicted.
16:28So when you detect it,
16:30you correct it.
16:32I think it would be,
16:35you know,
16:36completely unethical
16:37to ignore those problems
16:38and not fix them
16:39when they occur
16:40but they occur
16:42because those systems are new
16:44and so it's inevitable
16:46that new things will pop up
16:47that we didn't envision.
16:49I mean,
16:50is there anything we can do
16:52kind of preemptively
16:53so that Silicon Valley engineers
16:55with, you know,
16:55all their kind of goodwill,
16:57you know,
16:57they're probably not the best
16:58moral judges of the world.
16:59No, in fact,
17:01I mean,
17:02we're talking about AI,
17:03not about content moderation
17:05but the policies
17:07for content moderation
17:08at Meta,
17:10Facebook in particular,
17:11are not designed by engineers.
17:13They're designed by people
17:13who are either social scientists
17:15or, you know,
17:16specialists of ethics
17:18and policy
17:18and free speech
17:20and things like that,
17:21right?
17:21It's not engineered.
17:22That's a myth.
17:25Now,
17:26we've talked about this
17:27in a very abstract level
17:28but we're great to talk it,
17:29bring it to a very concrete level.
17:31So,
17:32say in,
17:33I don't know,
17:3310, 15 years
17:34we would have this,
17:36some sort of system
17:37that you propose.
17:38What would it be,
17:39what would its
17:40practical applications be?
17:41You know,
17:42it's a very interesting
17:43scientific pursuit
17:44but practically.
17:46Right,
17:46the scientific pursuit
17:47is to understand
17:48human intelligence
17:49and the best way
17:50to understand it
17:50is to kind of build one
17:52or build one
17:53that sort of behaves
17:54more or less
17:55the same way.
17:56But there are
17:56practical applications.
17:58We've been talking about,
17:59you know,
18:00level five autonomous driving
18:01for quite a long time.
18:03I think it would be made
18:05a lot easier
18:06and would require
18:07a lot less sort of
18:08detailed engineering
18:10and a lot of,
18:11you know,
18:12and efforts
18:12if we had
18:14machines that are capable
18:15of learning
18:16how the world works
18:17by observation
18:19because that explains
18:20why most teenagers
18:21can learn to drive a car
18:22in 20 hours of practice
18:24hardly without any supervision
18:26whereas we still don't have
18:28level five autonomous
18:30driving, right?
18:31So it is a possibility
18:33that the ultimate path
18:36towards things
18:36like level five autonomous driving
18:38is actually machines
18:40with common sense,
18:41machines that can learn
18:42by observation,
18:43that have word models,
18:44et cetera.
18:46And that may take a while
18:47before it happens.
18:48I think there's going to be
18:49level five autonomous driving
18:50before that happens
18:52basically by just
18:53engineering the hell out of it
18:54if you want,
18:55pardon my French.
18:57But then there are
18:58other applications.
18:59So one application would be,
19:00you know,
19:00all of us want
19:01some sort of domestic robot
19:02that takes care of all the chores
19:04in our house.
19:05That also requires
19:06sort of human level intelligence.
19:09But more importantly,
19:1115 years from now,
19:12we'll not be carrying
19:13a smartphone in our pocket.
19:15We'll have augmented reality glasses.
19:17Living in those glasses,
19:18living in those glasses,
19:19we'll be some sort of AI
19:21virtual assistant
19:22that will help us
19:23in our daily life
19:25as a human assistant we do.
19:27and for that system
19:30to not be frustrating
19:32to use,
19:32it would have to have
19:33some understanding
19:34of human nature,
19:35have some level
19:36of common sense,
19:37and basically have
19:39more or less
19:39human level intelligence.
19:41This is going to become
19:42really important
19:42because,
19:44not because we want
19:45to superimpose content,
19:47you know,
19:48on the world
19:48all the time,
19:50not just because
19:51we want to be able
19:52to, you know,
19:53be in a country
19:54where people speak
19:54a language we don't understand
19:55and see the translation
19:56in our glasses,
19:58but because we are
20:00being submerged
20:01by enormous amounts
20:03of information
20:04and that amount
20:05of information
20:05grows exponentially
20:06and we need systems
20:07to help us
20:08kind of sift through it.
20:09And so,
20:10if we have
20:11a personal digital friend,
20:14if you want,
20:14that is under our control
20:16and can help us
20:18kind of surface
20:19important information
20:20to us
20:21and deal with
20:21the complexity
20:22of increasing
20:23complexity of our lives,
20:24I think that's
20:27going to be inevitable.
20:28Yeah, great.
20:30Now, we're out of time,
20:30but I have a cheeky
20:31final question.
20:33If you had to make
20:34three predictions
20:35about what's going
20:35to happen in AI
20:36in the next two
20:37or three years,
20:38like, what is the next
20:40immediate stage?
20:41What would they be?
20:42So, those large models
20:45that are being trained
20:46currently mostly
20:47for language
20:48are not going
20:48to be trained
20:49and some of them
20:50already have
20:51on sort of
20:51multiple modalities
20:52so simultaneously
20:53with language,
20:55images, videos,
20:56audio, etc.
20:58It's starting
21:00and they will do
21:01impressive things
21:02that doesn't mean
21:03they're going to be
21:04sentient or general.
21:07So, that's the first thing.
21:08Second thing is
21:08this idea of
21:09self-supervised learning,
21:10basically training
21:11a system
21:12not for tasks
21:13in particular
21:13but for sort of
21:15filling in the blanks
21:17in a piece of information
21:18so show a piece of video
21:19to a system
21:20and then ask it
21:21to predict
21:21what's going to happen next.
21:23That's a form
21:24of self-supervised learning.
21:25We currently do this
21:26for text
21:26and it works really well.
21:28We're not able
21:29to do it for video yet.
21:30That's going to happen
21:31over the next two, three years.
21:32There's already work on this.
21:34So, this will be
21:35a new revolution
21:36in computer vision,
21:38robotics,
21:38and things like this
21:39brought about
21:40by the generalized use
21:42of self-supervised learning
21:43in those contexts.
21:45Then, I think
21:46there's going to be
21:48entire systems,
21:49intelligent systems
21:50that can take actions
21:51and plan complex action sequences
21:55using their world model.
21:56And this is
21:57this paper
21:58we're talking about.
22:00Sorry to interrupt.
22:01We're going to have
22:01to wrap it up
22:02because we've gone over time.
22:05Do you just want
22:05to have a final word
22:06to conclude
22:07what you were saying, Jan?
22:08Well, that was pretty much
22:09the end of it.
22:09Well, that's great.
22:10Fire away.
22:12Great.
22:13Thank you so much.
22:14Thank you so much.
22:16Thank you very much
22:17for this discussion.
22:18If artificial intelligence
Commentaires