English translation of the French transcript, made with AI and not yet proofread — click a paragraph to jump to it. Show the French original
Diderot: "speak, and I baptize you" 3:55 I've had reactions like that, informal, one-on-one, either by phone. Some are really enthusiastic to see that we're finally going to talk about artificial intelligence with someone who doesn't talk about it from the outside, and who is finally going to tell us what it is, rather than chattering about everything it isn't.
4:25 Which is the most common case media-wise, let's say, or at the corner café, which amounts to the same thing. I remind you, as always, that I have no social contempt for people at the corner café. It's not the place where truth blossoms, but that's not its function either. But you can also do very nice things there. And then there are, on the contrary, those who say, I don't understand, Pagani is going off in all directions a bit.
4:55 That was quite something, those analytic reveries. He gave us a whole number on Mignon, Goethe's Mignon, that kind of… Now we're going to be facing a machine. I'm caricaturing a bit on purpose, I trust Christophe to set things straight. But I'm speaking here on behalf of a certain unprepared audience. And whom I'd like to reassure right away, I see no discontinuity between, and yet, between our series of analytic reveries and our regular weekly Wednesday workshop, for which we're starting a new session tonight.
5:44 The proof, listen, I'm going to give you the proof by way of an anecdote I'm borrowing from Diderot. Right after Rameau's Nephew, sorry, D'Alembert's Dream, a bit as an epilogue to D'Alembert's Dream, Diderot offers us a conversation in the form of a dialogue. He's the one who somewhat launched the fashion at the time, and it would later inspire Rousseau, Judge of Jean-Jacques a little later. A dialogue essentially between, let's say, a doctor to put it quickly, and a young aristocrat, Mademoiselle de Lespinasse.
6:28 Her salon was famous, well, that's not our subject. And it deals quite a bit with the relationship between man and animal. You see, we're starting to get there, to artificial intelligence, not yet to the proof that I'm in continuity with my remarks on Mignon's song. But let's come back to it. And to illustrate his point, the doctor says to Mademoiselle de Lespinasse, listen, have you seen the orangutan at the Jardin des Plantes?
7:06 It was the very first time people had seen this kind of animal, in a cage of course, shown to the public of the northern hemisphere, in this case a Parisian public. And of course Mademoiselle de Lespinasse says, yes, yes, I've seen it, like a lot of people, very curious. And he tells her, the Bishop of Polignac supposedly said to it, orangutan, through the bars of its cage, speak and I will baptize you.
7:38 We're at the end of the 18th century. There had been Buffon, who had sketched out hypotheses of evolutionary theory. Soon there would be Darwin, a generation and a half, two generations at most. And already Diderot shows how little there is, isn't there, between the orangutan and man, but that little is enormous, since he actually decides to find out whether the animal in question is going to be baptized, in short, whether it belongs to the anthropological sphere or the zoological sphere.
8:23 As for me, if he stays in the zoological sphere, speak and I will baptize you. So you see, we're already into problems of language. I'm not talking about machine language, Christophe will speak about that much better than I can. But we're already dealing with something that has a relation to the ear, for now. Speak and I will baptize you. First anecdote. Second, I'm giving them to you like this, in apparent disorder, which brings us a bit closer to Mignon, because for now it's speak, not sing.
Kant's mechanical nightingale and the telephone in The Castle 8:53 And you see, Kant, in his Critique of Judgment, says, I'm walking in a park, all this is very 18th century, if you look closely, it's very gallant. I'm walking in a park and I feel delight at hearing the incomparable melody of the nightingale's song at the vesper hour. Isn't that so? And then my delight is spoiled by the owner of the place who comes up to me and says, you heard that, it's beautiful, but it's only a mechanism.
9:26 This is already the development of automata at the time. There was already a chess-playing automaton, one played chess with it, it wasn't a computer yet, but it was an automaton. You see, we're getting closer to it. And the man in question immediately says that his pleasure, yet nothing has been changed, not a single note, it's the same intervals. But the fact that it's a mechanism that imitates, and here I take up the word Christophe will use to introduce artificial intelligence, the set of processes by which a machine simulates, I say imitates, it's the same thing, an old Platonic term.
10:07 The back of the cave is a simulation. But let's not get ahead of ourselves. The fact remains that his pleasure is spoiled even though nothing has been changed in the music. And even the fact that we know it's not a someone singing, but a something programmed to do it, makes the walker express his total disappointment.
10:38 There you go. So, remember how I began the text on Mignon. Not even with a language, not even with a piece of music. I use the same word, our first text was the Kafka excerpt from The Castle. K, the novel's central character, right at the beginning, in the first pages, isn't that so? Since he's told, you're megalomaniac, what do you mean the castle has hired someone, a vagabond like you?
11:09 He says, fine, since that's how it is, let's phone and see. And he picks up the phone, we're finally going to hear the voice of God. The castle is transcendence, it's the Grail. Since the whole novel is a desperate quest to try to reach it, he will never reach the castle, far from it. Through his own efforts, he even deprives himself of staying in the village, so he's neither in the village like the ordinary sensible people, nor in the castle either, he's shuttling in between. That's not our subject, I remind you, though it's admirable as a structure, as a pitch.
11:40 He picks up the telephone, we're finally going to hear the voice of the absolute, the voice of God. It's the opposite of the machine. It's the voice, not of something, but of someone, what am I saying? Of the someone par excellence, of the only one who is. That's how he names himself to Moses, the Eternal. He says, we say "Yahweh" because vowels are needed, but otherwise it's unpronounceable. Because a man cannot say "I am." Only God is. We, we become.
12:11 I'm getting ahead of myself, we'll discuss all this, but it's interesting that Hebrew, like Russian, has no verb "to be" in the present indicative. In Russian you don't say "I am Russian," you say "ya russkiy," "I Russian." Because only God can speak of being in the present. Because as Pascal says, if he is eternal and once, he is always. I'm sowing a few seeds here, we're going to hear the voice of God.
12:42 I come back to the castle. And what does he hear? I told you it's neither a language nor a piece of music. Kafka's term is a crackling. He hears a crackling. Then this crackling seems like the voice of a multitude of childish voices, but distant, very distant, then merging into a single one, giving the impression of wanting to pierce something more resistant than a poor ear, to quote the admirable text.
13:12 There you go. So I just wanted to show you, I'm not solving any of the problems. I'm not answering any of the questions I've just raised a bit haphazardly. On the other hand, I've just reassured those who think we've jumped from one thing to a completely unrelated one. No, no, no, no. We are continuing our reflection on the relationship between the sciences that relate to man, the sciences of history, as I call them, you can call them the human sciences, the social sciences, it's all the same to me.
Building a bridge between the human sciences and the natural sciences 13:46 I've already told you that I remove myself from structuralist terrorism. I have nothing against structures. That would be anti-scientific, but structuralism is more than structures, it's an ideology. So call it what you like, the abyss, the unbridgeable abyss according to Kant that separates the sciences of history or human sciences from the sciences of nature or hard sciences, especially when you add mathematics to them, without which none of these sciences would have the hardness that makes their prestige — without which, if we say mathematics of the most real, creating a link between the two, throwing a bridge, however frail, over this unbridgeable abyss that Kant was trying to establish in his last Critique, which most inspired the post-Kantians, not the neo-Kantians, but Hegel and company.
14:43 You know it's going to lead to Marx after all that. So this cycle that I've titled with a question, how can historical sciences be real, and not merely possible, but real, you see, we are in continuity with the examples I've just given. Speak, I will baptize you. An engineer could say the same thing to someone who told him, here, you have a robot, you'll see, it's really a guy, there's no longer any difference, let's grant that.
15:21 The interlocutor would then be entitled to say to the robot, in order to test what his colleague is claiming, let it speak and I will baptize it. There you go. Yes, the problems are posed, they are not solved, especially not by me. The questions are posed. I'll give the floor to Christophe, who I think is concerned, first of all, out of courtesy toward our audience, with making a bit more public, a bit more commonly understood, what is meant by artificial intelligence, as I was saying earlier, by dwelling a little on what he has seen of it from the inside, as someone for whom it is an object of research, of practice, and of conceptualization — that's already a lot — and who moreover transmits his knowledge, since Christophe, as far as I know, has a great concern for pedagogical transmission.
16:21 That's not always the case among researchers, so I insist on this triple competence. And I hand him the floor so he can begin by telling us, well, from the inside, what should be understood, with as little nonsense as possible, about this phrase — I'll drop the quotation marks — artificial intelligence. Over to you, Christophe, if you don't mind. Hello. So, first of all, thank you all for being here tonight.
A definition without consensus: simulating human intelligence 16:53 So, I'm going to have the... So I'll also briefly introduce myself for those who don't know me, because there are many. So I'm Christophe Denis, I'm a lecturer-researcher at Pierre and Marie Curie University, which has been renamed Sorbonne University, and I carry out my work — I teach and carry out my research work — in the field of artificial intelligence.
17:23 And what interests me in particular is knowing what's inside the box of artificial intelligence, so to speak, but also in connection with scientific practice — that is, in what way machine learning, since we mentioned it right away, upends or doesn't upend the scientific discipline, in what way it can be a contribution, etc. And so I also teach, I'm really very interested in teaching too and in popularizing artificial intelligence, because we realize — I realize — that those who are kind enough to listen to me afterward ask me questions, and often they're far from stupid questions, and that plunges me back into research too.
18:25 So, what I propose is, well, I might start by disappointing you, because if you want a complete definition of artificial intelligence that has scientific consensus, that's not going to be the case, since it still exists — you see, for example, when there are meetings of people working around artificial intelligence to define what proper use of artificial intelligence is, generally in these working groups a large part is always, well, what is artificial intelligence — so you see, there isn't even a consensus among researchers. And today we're not going to present a whole theme on the ethics of artificial intelligence, because we hear a lot that artificial intelligence must be loyal, etc. In fact, that's a bit about the purpose of a tool, and sometimes, for me, it's a bit delicate
19:35 to attach somewhat exogenous or normative properties, as if you took a tool and said, this tool, or whatever, this saw must be loyal — I'm caricaturing a bit — because it must saw wood and not cut off a finger, or something like that. So I propose — I put a first definition in the slide, maybe I'll try to bring it back up, actually you can see a definition, well, Peter Norvig, he's the boss of artificial intelligence at Google, I took one among others, just so we can start having some fun together — so, here's a definition: artificial intelligence is the set of theories and techniques implemented
20:45 in order to create machines capable of simulating human intelligence. So it's a fairly recent definition, but it echoes a certain dream, a utopia of the so-called fathers of artificial intelligence. You have to understand that the term artificial intelligence came about because four researchers wanted to obtain funding from American universities to hold a summer seminar, and so they had described, well, we're going to propose a new technique, artificial intelligence, and that was the idea — and what's rather funny to see already is that, imagine, at the time, well, the computing resources available were really quite ridiculous, and as soon as those first calculating machines appeared, people projected themselves forward and said, well, could we, ultimately, from this mechanism, simulate human intelligence.
21:54 So, a lot happened around this, maybe we can come back to it if you have questions — there was, for example, the famous Turing imitation test to try to measure a machine's intelligence. Actually the idea was, you put a human being and a machine in two different rooms, and if the observer, by asking questions of one or the other, cannot tell them apart, one could say there's a certain intelligence in the machine.
22:29 Now, that was well criticized by another — well, Alan Turing is an important figure, someone important in computer science, so if you've seen, for example, there was a somewhat fictionalized film retracing his life, I believe it was The Imitation Game, which showed the role he played during World War II with the Enigma machines, but he also did theoretical work on computing before computers existed, and that's what's always interesting — computer science is constrained by the technological evolution of computers, choices are made because at a given moment industry, from an economic standpoint, develops a certain type of machine more, so people adapt, but there's a whole theoretical part that is independent, and moreover these theories of computability
23:41 for example predated the advent of computing machines, and it's for instance the Turing machine, by the way — I invite you, because there are quite fun things, for example some people have built Turing machines out of Lego, etc., so if you look, for example, in your browser or in your favorite Google, you'll find them. So, back to the idea — it was to simulate human intelligence, and there was — and I like these definitions because if you add a single letter, well, that reverses this definition a bit. For example if you add the letter T, I think that's also the role of this tool, so if I replace it, I also put artificial intelligence is the set
Stimulating rather than replacing; symbolic and connectionist approaches 24:51 of theories and techniques implemented in order to create machines also capable of stimulating human intelligence. I think — one shouldn't fall into certain fears that intelligence will replace human beings, human intelligence, etc., because I remain convinced that we have a tool at our disposal, and for the use of that tool not to be stupid, or really to be useful, we also need human genius, mathematicians too, to understand it, etc. So, on the other hand, we're not going to address this here — undeniably it raises societal problems, problems that need to be thought through regarding work too, but I believe that's not only linked to artificial intelligence, it's something that has been going on for a long time, from digitization,
26:01 from mechanization, etc., and it is not a rupture but rather fits into that continuation. So artificial intelligence was born in the 1960s, and often people talk about the revival — you see, the revival of artificial intelligence, for example, is about all the predictive capabilities, for example automatically recognizing in photos whether there's a dog or a cat present. So since there have been impressive predictive capabilities that dethroned a whole field — to give an anecdote about recognition activities — actually it's not even an anecdote, it's historically accurate — if we take, for example, in the context of image processing, before there was a whole scientific community working on algorithms, meaning on well-defined sequences of operations, to do
27:11 precisely image processing, facial recognition, etc., and there was a conference with challenges for researchers to test their algorithm, and a bunch — so to speak — came along with the machine, knowing image processing more or less, using these machine-learning techniques which we'll see right after, and they really dethroned the performance of all these algorithms, and that's why now, moreover, in the field of image processing, machine learning is used massively. So it's an upheaval too, in — I'll focus a lot more today on the scientific discipline, and it's also an upheaval from a generational standpoint too, because you know, in the field of research,
28:21 there are generally young researchers guided by the older ones, and with these new techniques it has somewhat reshuffled the deck. So, if I tell you a bit more, but I think it's also important to see that there were two main fields, two fields in artificial intelligence, and today we're in the field of machine learning entirely, but there were two fields — the symbolic approach, meaning that we explicitly define things from a formal point of view, for example we define high-level computer languages to describe concepts, a computer ontology, so everything is written down — and a connectionist approach, which we're currently at the peak of the wave of. So in the symbolic approach
29:31 what does that mean? It means we think of a set of rules — for example, it could be, not limited to this, but it could be the expert systems that existed in the 1980s, which were used a lot in — attempts were made to use them in the medical field but also in industry. It means we define a set of rules — like, for example, to detect the deterioration of a piece of equipment, if the temperature is above, I don't know, 10 degrees, whatever, you see, it's a set of rules, and so on. That's why it's called an expert system, because all the rules are described, and from the data — we input data — there's an engine to decide how all these rules will interlock, and we get a result. And on the other hand, one must also note that despite what's said, well, the revival of artificial intelligence,
30:41 of course, is linked this time to the fact that — I'm not going to explicitly write the rules this time, we say it's based on the data — for example, we'll measure this time, as I said earlier, the temperature of a room, the pressure, etc., but we won't define an explicit link — it'll be the correlations between the data that lead us to say, well, based on the correlations, statistically that produces this result, and what we hope for afterward is to deduce rules or models from it — that's what's called in the jargon explainability: can I actually deduce things from these models? And what should be noted is that in the life of AI, historically, there's been a succession here of phases where it's the connectionist approach that — and here the scale is, I believe, the number of publications, so it's very much linked
31:51 ultimately to research funding, so at a given moment something was more in vogue — you see, at the start, I don't have time to go into detail, but there was the connectionist approach, at one point the first neural networks that were developed — not only in the 2000s, actually the first neural networks date back to 1950, there was the famous perceptron, which people could understand fairly well how it worked but it was rather ineffective, so there was criticism of that, and people moved more toward expert systems, the symbolic approach, and at a certain point, linked also to a technological revolution in computers, we moved to connectionist approaches. So all this to say that artificial intelligence is far from being reduced to the approach we see now, that is, we're in what's called the big data approach,
33:01 we take the data, we put it all into an enormous black box which does its own adjustments, and we're left with a feeling, and we understand nothing at all. There was the symbolic approach, which had the problem of being poorly performing. There's a bit of a holy grail for people who work in artificial intelligence, and you can immediately see it would be to combine the two to get the advantages of both and avoid the drawbacks. So now I wanted to show you, because it's important, and — well, finally I'll show you rather quickly what a neural network is. You can imagine it as a set of circles like this, on which there are weightings, weights associated with them, and actually what's the principle — the principle, which is still a bit dumb, we'll see right away — it means that if we want to detect
Neural networks, dogs and huskies; correlation and causality 34:11 for example, in an image, a dog or a cat, we'll have a training set with lots of cats and lots and lots of dogs, and each time — this is what's called supervised learning — it means that each time there's an image of a cat, we'll attach a label to it, this is a cat, and through successive optimization algorithms, each time there'll be an error criterion that says, we're going to adjust this so that on average it makes as few mistakes as possible. And you see that if we compare this to intelligence, we might ask, but where's the intelligence, because imagine you're a teacher and you need to give a student, I don't know, 100,000 images of a cat and a dog for them to understand the difference between a dog and a cat — you wonder whether it's really intelligent, and whether it's not
35:21 just rote memorization to the power of 10,000, so to speak. Another thing is that if, say, we give it this time a photo of a cat that, due to some accident, is missing a paw, well maybe it won't be able to recognize it, and it'll say it's a dog, because that wasn't in the training set — whereas of course, for anyone who has children, when you show them the difference between a dog and a cat, you say it two or three times, you don't give them an astronomical number of examples. Another problem is that suppose this time you're not trying to detect a cat but you want to detect a dog versus a sled dog, for example — well, if you're not careful, I don't know if you realize, but the algorithm, for example, what can it do? Well, it can answer the wrong question, it might actually detect — because it's going to try to find differences between the two images — and it might conclude, well, I see that in the vast majority of dog images there's no snow, and in the images
36:31 of sled dogs, well, there's snow — so actually your method might end up detecting whether or not there's snow in the photo, and that's a problem, because if you take a sled dog, a husky, that happens to be in the city, it'll get it wrong. So that's what's meant by supervised machine learning — this has been made possible because we now have more and more data and because we have very significant computing power. So, coming back — I've almost finished, like this, we'll be able to — I can't get it to go further, there — what's — right, before, before giving the floor back to Dominique — so what has this produced? There's a somewhat illusionist side to it — and here too I want to say, this is because, when I was talking to you about human genius, etc., well, people, researchers too, have developed algorithms
37:41 a bit more sophisticated, a use a bit less crude than that, there's a whole research activity that will actually contradict what I wanted to present here, so Chris Anderson in his popularizing article in 2006, so he's an American who is very - you see - very tech, very with-it, saying that in the end machine learning and big data are going to save the world, and especially the scientists, that we need to be brought up to date a little, and well, now the machine is going to tell the truth, in quotation marks, that's why he said, well, the avalanche of data is going to make the scientific method obsolete. Why? Because he says now the models are imperfect, so it's true that models are imperfect and we have something that's going to be truly perfect, we have petabytes
38:52 that is, a large amount of data - what he doesn't mention either is that there's also a huge amount of computing power behind it and a very significant electricity consumption, so well, we could use energy in a more reasoned way. He says, well, correlation is enough, we can stop looking for models, we can analyze the data without any hypothesis about what they might show, and we have the result. Now, from a practical standpoint, as a practitioner, I tell you right away this is false, since the data is in fact always pre-processed - by whom? By people who pre-process it, for example to replace missing values, because if you're working with an industrial sensor system, you might have a faulty sensor giving you aberrant values, so there's pre-processing involved, etc. So I won't go into detail, but the main point is that we actually need models, since what is the model of machine learning, how can it be embedded, and one thing
40:02 that's also important is to really distinguish between correlation and causality. It's not because - and here you see, for example, it's always a bit of a recurring theme in the press - they tell you, we did statistical analyses, that by drinking a glass of wine you might decrease or increase your chances of getting cancer, and then another study comes out saying more or less the opposite. That's because when you have correlations, if you have a correlation between two phenomena, you don't necessarily have causality, there might be a hidden variable that explains the phenomenon. There are two examples for this - there's actually a website called Spurious Correlations that I use often, and that we often use with our students to explain this. It shows that you can construct correlations between phenomena that have absolutely nothing to do with each other
41:13 for example, margarine consumption in Maine and the divorce rate - maybe there's a connection but I don't see it right away - and on correlation with a hidden explanatory variable, for example you can say this tanning cream really does increase the rate of skin cancer in people, but we forget that there's something hidden, namely that people who use tanning cream also expose themselves to the sun. That's what I wanted to tell you in this introduction on artificial intelligence and machine learning. Of course there would be more to say, but I think it's important that we can now discuss things with Dominique. Yes, thank you very much Christophe, because this is an opening that we needed, and I don't know, I have a feeling the public is only asking to continue this elucidation that you've just begun. As for me, I
Hume, Kant, and causality; the myth of the artificial man 42:24 I'm going to insist mainly on - once again - the continuity that your remarks bring to issues I've been trying to bring out for quite some time now. I've already proposed a correlation - it's an example of the correlation-causality issue that makes me stumble onto a dichotomy between the something and the someone. This matters so that the machine doesn't become the one in the famous film - isn't it - 2001: A Space Odyssey, where the machine is put to death outright. By the way, I'll pick up my thread later on the song of Mignon, an obsessed quarchot - that's not me, because Kubrick himself, when he has it put to death, the robot's last expression is nonetheless a song, and even a nursery rhyme, which brings us even closer to the character of Mignon who, as Goethe specifies, is between two sexes and two ages. Well, but more seriously
43:34 you spoke about - you recalled the very important correlation in scientific rigor between correlation and causality. I remind those listening, and it's exam and competitive-exam season, I'm addressing mainly the philosophers here, that for decades the bulk of Kant's theoretical effort - I say theoretical effort, meaning his major corpus, the Critique of Pure Reason - was to grapple with the concept of causality, which for Hume, wasn't it, was something illusory. I remind - sorry Christophe, I'm subtitling what you say for certain ears that need translation - I remind you that when Kant says 'Hume awoke me from my dogmatic slumber,' an important phrase, that means up until then I was a Newtonian, a bit like everyone else, that is, if I see
44:44 a billiard ball heading toward another, I anticipate, I say that as soon as they collide, the other one will start moving, and that movement is caused by the movement of the ball I just sent toward it. And here comes Hume saying, yes, but that doesn't mean there's a law - let's be rigorous now - an objective link between the two phenomena, but a subjective link, namely that I have always seen one, I as subject, follow the other. And at that point you have to be very careful - and here I'll make the audience laugh a bit to lighten things up - Hegel used to mock those who confuse causality and correlation by having someone who observes their mother taking in the laundry say: it's not that when the sky turns threatening [and] in the garden one takes the laundry off the line - the kid says: the cause of the rain is that my mother takes in her laundry, because I've noticed that every time she takes in her laundry, five minutes later the rain starts falling. I think that makes clear the abyss between cause and correlation. So let's sum up rigorously
45:54 Rimbaud, in A Season in Hell, poses this question, or rather expresses it in the form of an alternative affirmation, which is precisely our problem. He says this strange phrase: there is no one here and there is someone. You could say that of a machine - there is no one and there is someone. It's a bit like my artificial bird earlier, you see, we always come back to this margin of uncertainty. Before I give the floor back to Christophe, let me mention an agrégation topic that was proposed about fifteen years ago, not for the philosophy agrégation, still less for the mathematics or hard-sciences agrégation, but for the literature agrégation, because on the syllabus, as always in the agrégation, there's what's called a comparative literature syllabus, a given theme - first encounters in a novel - and you compare how the first encounter is treated in different works, in The Princess of Clèves, in Sentimental Education, in Swann's Way, whatever. But here the theme was the artificial man, how it appears
47:04 because there's a whole literature - Frankenstein, Villiers de l'Isle-Adam's The Future Eve, a kind of fabrication like that at the end of the 19th century, isn't it, the homunculus in Goethe, isn't it, made in a laboratory already - that means the little man, you see. And the topic was formulated in a strong, intelligent way, because it consisted in saying this: in all these works, whether masterpieces or less successful works - that's another problem, purely literary - all these works actually always affirm the same implicit judgment. And you'll see how this applies to the discussions on artificial intelligence, especially in public opinion, namely that something has been transgressed - an order has been transgressed, an order sometimes considered human, sometimes natural, sometimes both. I'm thinking of Spinoza who says God, that is, nature - man has transgressed an order by creating
48:14 Frankenstein for example, and that of course creates disorder - he transgressed, you see. I say this to show that one shouldn't be too geeky and up to date either, like Christophe showed us earlier that some people like to show off a bit; but I mean one shouldn't advance the chemical too much either, because these are very old problems, contrary to what - well, what I'd like to point out is that what's new, of course, is the way of treating them, the way of experimenting with them, the way of putting our hypotheses to the test of facts, and in that respect the machine is an extraordinary field, isn't it, but in its dialogue with the one who manipulates it, of course. So I'd like to point out that these are old problems - the artificial man has always fascinated theologians, writers, filmmakers - I'll illustrate with Kubrick later, isn't it, it's not the first nor the last time we'll do this - but I assume your discussion digs a little into this difference, my dear Christophe? Yes, but I wanted to come back also to another example on
ELIZA: being listened to rather than understood 49:24 one always tends, in any case human beings tend, to think that the artificial intelligence system is more complicated than it actually is. So let me give you an example - it was, I believe, in the years - I hope I remember the date correctly but it's not very important - I think it was in 1967, there's a computer scientist named Joseph Weizenbaum who developed a program called ELIZA, maybe some of you know it. So he was a bit mischievous, actually - he wanted to simulate the interview between a patient and a Rogerian psychotherapist, which is where the mischief lies, since it's the practice, the branch of psycho- psychoanalysis based on empathy - so you see that's a bit of a wink, since here it wasn't empathetic at all. And so it was in a programming language called Lisp
50:34 but that's not very - not very important, except to mention one of the fathers of AI whom I find, with a beautiful career, quite inspiring, in any case for a computer scientist, who was John McCarthy, who was at once an engineer, computer scientist, mathematician, and who developed things we still use now, such as, for example, the possibility of sharing machines - he had thought of cloud computing, meaning computation etc. So I close the parenthesis, but that's part of it - sometimes when you're a researcher in computer science, and it's the same in philosophy, there are people you sometimes hold onto. Parenthesis closed, let's get back to ELIZA. So ELIZA, so it was a simulator, so the user poses to ELIZA as - so it actually had quite a crude interface, you can imagine, like the old
51:45 old screens, a bit monochrome, without images, a sentence gets typed, you answer, etc., so there wasn't all that side that you now have with chatbots or maybe things a bit in 3D. And yet, well, when a user - so he said 'I'm doing well,' no maybe not, otherwise there's nothing - if the user, the patient, typed 'I'm not doing well,' the machine would answer 'why aren't you doing well?' He'd say 'because of my parents' - 'what's going on with your parents?' etc. So we already had the first approach, the user's feeling that there was some intelligence behind it. And in fact if you look at the program, it was really, from a computer-science point of view, that you detected patterns in the question - for example when someone said 'I'm doing well,' if you recognized the pattern 'I'm,' you'd rephrase
52:55 with a question, you'd take it up again - I take 'I'm doing badly,' so it detected 'I'm,' it took 'badly' and put it back into the question, 'why are you doing badly?' You see, so it was a bit of a trick, and one might think - so what's interesting is that psychologists actually detected that there were people who became addicted, there was addiction to this thing which was, moreover, really not visually appealing at all. And that's what was called at the time the 'Eliza effect,' which has been studied, and there's an anecdote, for example, it was Joseph Weizenbaum's secretary - she wanted to use this program but alone in her office, with the door locked. So Christophe, just so we understand well, because your example is fascinating - if I've understood correctly, Eliza ended up exercising, let's dare the word, a seduction - exactly
54:06 on a patient, it seems to me - and well, well before the film, if you've seen it, the film Her, where in fact someone falls in love with a voice - and there, you see, there was nothing sexy about it at all. And that means we shouldn't fall into certain traps either, because there can be an illusionist side to it sometimes - it's a bit like, and this ties in, if you look at the techniques, we've always used things sometimes for fun, but you still have to know you're having fun - for example we use electricity to do I don't know what, but there we shouldn't fall into certain traps. And where it gets complicated is that sometimes we need these illusions, for example from a medical standpoint, for certain people - there are studies that have shown that the empathy of certain robots contributed to healing. So there you go, but in any case
55:17 we could produce an act, that's it - so that was to come back to the point about being, how you put it, the artificial being, etc., and so what's rather funny is that, for instance, this is what I gave my students who take the introduction to AI, they programmed it, they saw, well yes, there was no intelligence in it at all, but it can be seductive too - afterward it can create problems, since not knowing whether you're facing a real human person can also create rather serious psychological issues, because in the end we were locking people in who preferred to be listened to rather than understood, in a way. Ah, that's a nice distinction, I'll remember it. There you go, thank you very much. Well then, precisely, so we've stirred up
Adam, the demiurge potter, and man made in the image of God 56:28 material of such richness, at the same time of such topicality, that it allows us perhaps to come back - I say this to reassure some members of our audience who sometimes feel they're mixing things up - on the contrary, to reassure them, we're coming back to fundamental, simple things that everyone can understand and whose reference practically everyone knows. Let me take an example from the Bible - whether one is an atheist or a believer doesn't matter at all - I'll take an example from the Bible, but I could say it's because it's an extremely well-known cultural reference, regarding - you spoke about the artificial man, that was the initial phrase - what is man in the Old Testament if not an artificial creature? Well, still, I don't - I don't like interpretation, but my listeners know I only do analysis - I analyze the text. The text says God takes a clod of earth, silt from the Nile, silt is called 'Adam,' isn't it, in Hebrew, Adam means the mud, the mud of the Nile, the clay to which the potter
57:38 can give whatever shape he wants, isn't it - and it's interesting to see, moreover, since I'm in complicity with a logician, mathematician, computer scientist this evening, to see that data is mythical here, namely Adam as the mud, the matter to which any shape can be given. And in Plato, by contrast, traditionally, apparently quite independent, the demiurge in Plato who creates all the things we see around us - at the start he's a potter, you know, Plato is the greatest stage director in the history of thought - he's a potter who has his potter's wheel there under his foot, looking at the models of the noetic cosmos, he sees models - the idea of the dog, we were talking about dogs and cats earlier - and with his wheel he makes copies, simulations of what he sees of the idea in the noetic cosmos
58:48 except here's the thing, that's where you see the depth and the humor are always linked in Plato - he was a bit tipsy at the moment of creation, which means that when he sees something round he makes an oval, you see, it's never quite the same. But let me get back to the biblical reference - so God creates an artificial being called man. Of course, there's the material side, he kneads the mud and all that, but let's be clear and honest, to give it a soul you also need the breath - you know, I remind you that breath, in German, is the old Indo-European root that gives, in Sanskrit, 'atman,' you see. You know, in the New Testament, when they ask Christ what the spirit is, he says: 'you hear the wind, you know neither where it comes from nor where it goes, so it is with things of the spirit' - the wind blows where it wills. So there you go, it's almost a banality in the symbolic order - I mean, I'm not diminishing in any way the majesty of comparing the soul or the spirit
59:58 to something as unstable, isn't it, and as hard to grasp as a breath. But where I want to get to is that it is said now - let's leave speculation, let's get back to textual analysis, because we're also doing experiments here with the text - it is written 'God created man in his image,' and we find simulation again - that doesn't mean God created God, he created something that gives the image of God, meaning there are as many links between man and God as there are differences, for it is, after all, only an image. So let's try to dig into that, and then I'll give the floor back to Christophe - but you see the philosophical stakes underlying these issues of artificial intelligence. What does that mean - what makes man be in the image of God? Does that mean God has hair in his nostrils? Well, that would be an insult to him, so it's not an image in the visual sense, you see clearly, it's not that God has a look like that with two eyes, no, no, it's maybe
1:01:09 precisely - and I come back to the bishop who said 'the orangutan speaks, and I baptize you,' and that in the beginning there is the Word - maybe it's that man speaks, and it's insofar as he speaks that he doesn't merely, like dolphins and bees, or even machines, he doesn't merely transmit information, because animals know very well how to do that, an animal's mode of communication is very sophisticated, where the collective matters a lot, like with bees - it's very sophisticated, I recall von Frisch's work, isn't it, where a scout bee can come and report to its community, to the hive, that there's food, at what distance and in what direction - that's not bad at all. But that's not language. Why? Because there isn't the double level of articulation - I'll ask Christophe, I could even ask him the question before - before I finish my little bit, he'll tell me what he thinks of the famous double level
1:02:19 of articulation which the most rigorous of the human sciences, linguistics, one of the most formalized, let's say, has brought to light - that in every human language there is first of all - human language is divided into languages, which is not the case for bees, bees in Japan, California, etc. have roughly the same information system, whereas there's no relation at all between 'lion' in French and 'diara' in Bambara in Mali - to say 'lion' they say 'diara,' for the rest it has nothing to do with it. We saw last time, so I won't insist, you no longer confuse signal and sign - there we've moved on to the sign, that's the double level of articulation. The first we share with animals - the first level of articulation is the set of sounds a human throat can produce depending on its physiological, physical configuration, that's nature, there we're still in the physico-chemical, biological world, whatever you like, we're not yet in the human sciences, not yet in anthropology. And on top of this first level of articulation, a second level
1:03:31 of articulation, this time conventional, transmitted from human to human, allows access to the sign, a combinatorics that is this time symbolic. So there, I recall here that if man can somehow be considered an image of something he transcends, it is by the fact that he speaks, and when I speak, I am not like a machine in univocity - it's very important to understand that a machine is always in univocity, what do you call it, that's called a mapping in mathematics, isn't it, and even bijective, that is, to each element of the starting set corresponds one and only one element in the arrival set, a bit like in the highway code, if you like, which is very different from a language - in the highway code, a given signal must imperatively have this as its meaning, otherwise you fail your driving test. Whereas in language as actually practiced daily, you notice that it's just as important to conceal as to say - which is not what
1:04:41 mechanical or electrical or computer communication does, isn't it, because language's function is not, in the first place, to make us know - it can be to conceal, it can be to seduce, it can be to command, it can be to grant grace or, on the contrary, condemnation - all these are functions of language at the outset, it's not for saying A and B, or A and not-A. There, I'll mention in passing that - I know that all those who work on artificial intelligence, I know nothing about artificial intelligence compared to someone like Christophe, but what I know from the exchanges I have with those who work on it is that one of the thresholds they're trying to cross is how to get the machine to speak the way you and I are speaking right now. There, Christophe, correct me if I'm saying nonsense, don't hesitate. I'll say - no, no, but it was in relation to algorithms, I think - yes, that's what you were saying at one point about the somewhat biblical passage, it reminded me of a quote from Saint-Exupéry at the end of Wind, Sand and Stars, which I had, which said
Road signs and unicorns: the machine cannot say "I don't know" 1:05:51 only the spirit, if it breathes upon the clay, can create man, which comes back to what you were saying. And I think, yes, actually, on the signal-sign approach, how can we illustrate it, for example - well, let's take a concrete application, for example self-driving vehicles - self-driving vehicles that need to recognize road signs, so there's the whole mechanism, so there's the acquisition of signals by cameras, etc., with techniques - then there's really the optical technique behind it - and the system will say, well, that's a no-entry sign, that's a turn-right, that's a turn-left. And there have been examples - we'll see other examples later - but if you put, for instance, small pieces of tape on a no-entry sign, well, maybe
1:07:01 the system will say it's a turn-right or an accelerate sign, whereas of course a human being will immediately see that it's still a no-entry sign, perhaps a bit damaged. So on this sign-by-sign approach, and to come back to what's difficult, as you said, is everything about the non-ah - for example the non-ah, let's - let's take our example from earlier again - we wanted to recognize animals, imagine we have a database of elephants and of, what, horses, there you go, we have a database, and at a certain point, so with the whole mechanism I told you about earlier, we put in an elephant, so it has learned by heart, and it's going to see with the signals - precisely at the input of the neural network that I presented earlier - it's going to say it's an elephant. But now imagine this time we put in another - a horse
1:08:11 it's going to be a horse, whatever its size etc., but this time imagine we put in a unicorn - it will say it's a horse, whereas even a child would say 'but wait, that's still a rather peculiar horse.' So there isn't this possibility of saying, well, you gave me horses or elephants as training data, so I'll tell you it's a horse or an elephant - in this approach there's no side of saying, 'well, I don't know.' So there's a choice, a choice that's actually somewhat hidden, since it's often expressed in terms of probability - what's hidden from us is that when earlier I told you, well, it recognizes - I myself used these terms - it's going to say that's a horse
1:09:21 or a - what was it, a dog or a cat - well it's more a matter of probability, for example it's going to say I have a 30% probability that it's a cat, or a 70% probability, so there's something that's going to say, well, I'll take the higher probability, I'll say it's a dog. So what can we say about doubt? Well, it's more up to the person, or the people, who will post-process the result to say whether 40%, for instance, is something acceptable or not. Hence also the interest of having humans in the loop, since with automated processing, for example on road signs, afterward there was work done to improve stability etc., to come back also to
Galileo, the beauty that leads physics astray, and the passage from signal to sign 1:10:34 the use of artificial intelligence also in the scientific field - there was a saying about models, that all models are wrong but some are useful. It's a bit the same in the field of simulation too. Now, if I may share another slide with you - I don't know if this is the moment to talk about it, but at one point, to come back to why machine learning is used in the scientific field, one could say we've had centuries of mathematization of physical phenomena. So for example, there's - Galileo in The Assayer said
1:11:45 that philosophy is written in this vast book which is continually open before our eyes, but we can understand it - to go back to what you said about language, Dominique - only if the language and the characters are written in mathematical language, and the characters are triangles, circles, and other geometrical figures, without which it is humanly impossible to grasp a single word of it, and without these means one risks wandering in a dark labyrinth. And I find that the word 'humanly' is also interesting, really very interesting, worth focusing on. So there was already a side to it where Galileo was also trying not to draw the wrath of the religious authorities by saying, well, we put this, but to say there's still God, it's God who governs everything, after all. If I may interrupt for a second - 'humanly' means 'if I set aside'
1:12:55 the effects of grace or of the supernatural. There you go, exactly. But I draw another reflection from it too, which is also - is it, in the end, about the positioning of mathematics, actually it's a somewhat metaphysical angle, but on which I think we won't be able to give an answer, but it also helps situate artificial intelligence - is it, in the end, that mathematics is the language of phenomena, that a phenomenon is its language, and so the human being needs to know this language in order to understand nature - or is it an intermediary between the natural phenomenon and the human, actually it's the human who needs to formalize, etc.
1:13:52 so why am I talking about this - because we could imagine that machine learning is also, when you brought up earlier the example of Plato's demiurge, we could say that machine learning is a bit like the link between the world of laws, the intelligible world, and the sensible world. So why, you might ask, why all this, we already have beautiful equations, beautiful theories - it's that in certain disciplines we see certain limits. For example, I encourage you, there's a book that's well written, really excellent, a popularization book by Sabine Hossenfelder, who's a researcher - Sabine, in her country they call her that, yes - and in fact, this is already quite outside of AI
1:15:02 it's that she also shows conflicts within the scientific field - you see, because we have the impression that with Covid we realize there are conflicts among scientists, and well, what we see there is really quite harsh, I encourage you to read the book, you'll see exchanges between well-known physicists where it's sometimes almost a fight, a bit of a squabble, so to speak, but well, after that, what she does is - you shouldn't be surprised, scientists are human beings like anyone else, but what drives research forward is that at a certain point, consensus, dialogue, etc., confrontation - since there is confrontation - is needed to arrive at concepts. So let me get back on track, because I'm digressing, even though the title is 'Lost in Math: How Beauty Leads Physics Astray.' Why do I say that - well, you know, there were great advances, mathematization did allow us to
1:16:12 discover phenomena, and we even had mathematical results and later found that, from a physics standpoint, they corresponded to real things. For example, the best example is the Higgs boson, where actually there was a problem in the mathematical theory, and this small particle was introduced, which was later rediscovered, but with particle accelerators, etc. And then there's also, for example, all of Einstein's work, on, for instance, the orbit of Mercury - yes, thank you - of Mercury, which couldn't be explained by the Newtonian approach, and with the relativistic approach it could be. On the other hand, what Sabine says is that maybe we've now gone a bit too far - to put it briefly - is that now, with all this work
1:17:22 of course most physicists - so here I'm quoting - see formal beauty as the royal road to scientific discovery. Blinded by mathematical elegance, they speculate about black holes, develop stunning theories, propose dozens of new particles, decree grand unification models, but - and this is important - almost none of these ideas has been confirmed by observation; in fact many of them are simply unverifiable. And you see, that poses quite a dilemma regarding scientific practice, where is the experimental side, etc. One could argue that, for instance, with the Higgs boson, it started off - it required a lot of investment - so to get other particles it would require much more money too, and so on. But what Sabine says is that we also need to go back to a certain practice, and for instance that can be machine learning, meaning we have a lot of data, and well
1:18:32 with that data one can try to use experimental data to bridge these equations in terms of validation and explainability. There you go. So, just to be a bit concrete, the queen - what's really very formal, sort of the queen of disciplines in terms of formalism - is mathematics, and if you recall or not, it doesn't matter much, there are equations - and actually you have types of equations, partial differential equations that govern phenomena - when I was talking earlier, for example, about mathematization, that means that, for instance, if you take the flow of a river, well, you can model it using equations, you take a certain type of equations, you don't have to know their exact form, for example equations of the Navier-Stokes type, which allow modeling the flow of a river. Personally, for example, when I started
1:19:42 doing numerical simulation, I was quite amazed at first, astonished that with equations you can reproduce physical phenomena, so that's what pushed me to continue - and there, you see, that's here, and that's why - I wanted to present this slide to illustrate the signal-sign approach concretely. What are we trying to do here, you see - we have data, data, so here for example it's the '1,' it's data we measure - you see, they're signals - and what we'd like to arrive at, in the work, is to arrive at a rather sign-formal approach, an equation, for example here a mathematical equation, and what plays the role here of 'explore candidate equation,' so exploring candidate equations, is going to be machine-learning techniques, for example it could be pattern-recognition techniques, and so we arrive like that
1:20:52 we can arrive at moving from a sign approach to a rather signal-based approach, using machine learning techniques, and very concretely, there's work by physicists at the CEA that I took part in - there was a doctoral student whose thesis I reported on, who had actually used machine-learning techniques to help them understand, to differentiate particles in detectors, etc. And there was also the economic side to it, meaning you can either further improve the devices and make them more complex, or help us process the data. So that was another explanation of the signal-sign approach, and what's interesting is to know how you get there
1:22:02 precisely, what is the process that, within machine learning, allows this signal-sign approach to be carried out, in order to try to understand, explain, validate. Given how late it has gotten - it's not far from 10 pm - I wonder if it isn't time to give the floor to those who have questions for us, and I - so, there are two last ones that ask a comment: isn't AI simply to algorithms,
Does machine learning fall under the category of algorithm? 1:23:13 what the universal machine is to mechanical automata? Yes, Daniel Arias is a loyal companion and we thank him for his loyalty. Yes, I think that's another analogy. What I find important too, and it's a whole area of work we're doing, for example with people who are more into the philosophy of logic, is also to redefine the notion of algorithm.
1:23:53 What is an algorithm, etc. Since in machine learning techniques, are the techniques we use really algorithms or not? Why? Since we can use a subclass of mathematics for optimization. When I say subclass, it's of course not pejorative, because a lot of the work involves fairly complex optimization methods.
1:24:24 But is it really an algorithm? Since a particularity of algorithms is precisely that you can specify them. That is, here's my algorithm, it will perform this processing. If I give it this input data, I'll get this output data. And AI, in any case with deep learning, that's actually part of the difficulty we have now. We can't really understand how it works inside, from a mathematical point of view, etc.
1:25:00 So, yes, on some, I would say, maybe, if we distinguish between symbolic AI and connectionist AI. In any case, I think that symbolic AI really does fulfill that. For connectionist or statistical AI, I'd say it depends, but it really does depend, it depends on the type of method.
1:25:31 So, let's imagine we use a linear regression method. Linear regression, if you have a scatter of points and you draw a line that's closest to all the points, to minimize the error between the points, well then in this approach, we could say it is an algorithm, since I know exactly the process, I know exactly how to determine it.
1:26:01 There you go, but for neural networks, do we really talk about algorithms? Well, not yet, I think, and that's precisely the difficulty, and that's why it's complex to validate. That's quite clear. So there might be others, because sometimes I can't get through, because there were earlier questions. There aren't only questions, there are also some rather gratifying things, let me tell you about that for just a few seconds.
1:26:35 I'm thinking of Charles Michaud who tells us, "Good evening, delighted to join you for my first live session, fascinating workshop." Ah, thank you. That's encouraging, it's... Because I question myself, it's a confidence I rarely share, I question every evening what I thought that morning. So, in a way, that encourages me. Doesn't it, to see somehow that not only are we not working with a violin, or even a double bass, but that there are people who find something worthwhile in it.
1:27:08 And somehow that means there's really something circulating, a common interest between us. Thank you for this... Thank you very much Charles. To Charles, there you go. I'll hand the floor back to you, Christophe. Ah yes, so... From QL, I hope that's how it's pronounced. In any case, thank you for the question. So, I'll re-read it. The one-way street there? Yes, but in the case of the one-way street, doubt comes into play in man.
Can the machine doubt? Autonomous vehicles and Kasparov 1:27:40 Now, the machine has no room for doubt. Well, that's a very good remark. I think so too. Really, it's really about the place of doubt. Already, can a machine doubt or not? That's the question. And there's also a side... You see? So, how should I put it? It's that even if we... Is it... If we don't talk about doubt, but about making mistakes, there you go.
1:28:10 There you go. Does doubt... Does man also have the right to be wrong, and the machine doesn't have the right to be wrong? That actually echoes what you were saying about 2001: A Space Odyssey. In fact, it's the machine that makes a mistake once. We tell ourselves it really has to be unplugged because if it made a mistake once, it's going to make more mistakes and we no longer trust it at all. And it's the same thing too, you see, from a psychological angle too, you see, for example, with self-driving vehicles, when there have been accidents, which is tragic for the person, but when there was an accident, it also caused a whole resonance.
1:28:54 Whereas... Why? Because, in the end, unfortunately, if we say there was an accident because the driver was drunk, etc., in a way, that's reassuring. I say that in quotes. That is, there you go, it's the human, I have my explanation. And that's a bit of a... A human failure, as they say. There you go, it's a human failure. And on... For example, you could say that, you see, if we take self-driving vehicles and say, "Would we accept living in a society where we reduce, say, the number of fatal accidents by a factor of 4 by only having self-driving vehicles?"
1:29:41 I mean, there you go, we reduced it by a factor of 4, but there are only self-driving vehicles and they kill 4 times fewer, 10 times fewer, I don't know, 10 times fewer human beings. People would say, "Well no, because we'll manage to get to zero and already, we wouldn't understand why we can't get there with mechanisms." So, back to doubt, there was another interesting example too about failures.
1:30:17 I don't know if I have time to talk about it, but I think this ties back a bit to the ELIZA angle and the representations we form of the machine. It was, I believe, in the famous chess tournaments between Kasparov and a machine. There you go. And what was quite... So it was a machine that, at the start, there had been a lot of progress, but it used fairly rudimentary algorithms, algorithms developed in the 1970s.
1:30:54 But where the machine's strength lay was that it had a lot of data and very significant computing power, which allowed it, for those who play chess, to have a very deep look-ahead compared to Kasparov. But, without really meaning to, over the course of the tournament, Kasparov realized, even without knowledge of this algorithm, he told himself it had rather simplistic behaviors and deduced a certain algorithm.
1:31:35 And, at a certain point... He spotted recurring patterns. Exactly, there you go. And, by the end of a game, well, by the end of a game, all of a sudden, Kasparov loses his composure and loses. Because in one move, the machine made a move that was quite discontinuous with the whole sequence it had followed before. And he told himself, well, this time, the machine is becoming intelligent and it's going to be tough for me.
1:32:07 And afterward, I believe he lost or he resigned. And what was behind the story? Well, it turned out... So afterward, Kasparov said, I still want to have the record of my moves, not the machine's moves, to have transparency, how it played, because it's not fair, because the machine had recorded everything, or rather the engineers behind the machine. And what we learned afterward was that there had been a bug in the machine. At a certain point, when making the move, there was a bug in the algorithm, and instead of getting the algorithm's result, it produced something rather random.
1:32:43 And so, Kasparov told himself, ah well, that changes things radically. This ties back a bit to approaches around signs, signals and signs. That is, at a certain moment, there's a discontinuity in signs that he sensed. So, on doubt, there's also work on the role of chance in decision-making.
1:33:16 That means, well, these are things I'm not necessarily in agreement with, although, there are also people who argue precisely for not having transparency, and precisely for taking a perhaps interesting approach, saying, if we go back to the earlier example, on detecting, what was it, elephants and horses.
1:33:48 And for example, if it's a unicorn, we could say that if the probability is a bit, a bit low, etc., well, that it shouldn't be the machine that makes the decision, but that it tells the human behind it, well then, I'll tell you, I'm telling you I'm picking something at random, and here's the result, because it's up to you, after all, to sort it out and make the decision here. There you go. And, it's Alexei Grinbaum, at the CEA, who developed this theory, also using the example of a biblical passage, I believe, involving Joshua, when, well, I don't remember exactly the title, but it's when, so, God, God had said, well there you go, you can go into this territory, you can conquer it, but above all, you must not, steal, I believe, you must not steal, there was a theft that was, that was committed, and then, well, let me get back on track, there was a condition,
Joshua's dice; weak AI and strong AI 1:34:58 there you go, there was a condition, in fact, the thing was, they had to win, and they lost, and in fact, well, it was this condition that had not been fulfilled, since there was, someone who had, who had not, who had broken a divine rule, so, there you go, that's to set the context, and so Joshua said to God, well, tell me, tell me who is guilty, and God told him, well, take the dice, in fact he told him, well, take the dice, you cast them, and you'll have the result, and of course, well, God knew who it was, but in fact there was an aspect, he didn't want to create an asymmetry, in fact, in the role of the accuser, and that it should be man, who, by casting dice, finds the guilty party, and well, the analogy that Alexei Grinbaum draws is the same thing, in fact, shifted, between the machine and man, it's rather, this time, there's one who says, well there you go, well I cast the dice, and it's up to you to make your decision, so it's, it's something, it's something that,
1:36:08 that also allows for an element of chance, that's another theory, so on doubt, to come back to doubt, there you go, in fact, it's very difficult to finish, because I wanted to speak up on doubt, yes, of course, I think I know, perhaps, what, but doubt is a bit complicated, because, in order to doubt, I think you need to have a certain consciousness, in order to doubt, and that raises, that raises questions, can the machine, perhaps, have consciousness, be conscious, etc., and, there you go, and, I'll just finish, then I'll hand you the floor, is that, also, I encourage you to read a book, written by the person who runs the research team I work in,
1:37:18 called Jean-Gabriel Ganascia, who thinks about ethical problems, and who has defined, and who has written a book, on the myth of the singularity, that we're sold, by high-tech companies, etc., in Silicon Valley, that at some point, we're going to have a technological rupture, and, we're going to have, well, machines, that will be endowed with consciousness, that will perhaps replace humans, etc., and, where it becomes a bit problematic, is saying, well, if we don't want to be replaced, well, we have to go in for transhumanism, etc., there you go, but, above all, this is very, very, very far from being true, why is it far from being true, because, still, machine systems, artificial systems, are very strong, but for doing, really fixed, well-delimited operations, recognizing a dog, a cat, etc., and things that are, fairly easy for a human to do, well, it's very difficult
1:38:28 for the human—for the machine, to do it, and, that's, that's the distinction I wanted to introduce, because, sometimes, even in popular science magazines, between what's called weak AI, well, that's today's AI, meaning that, in the end, it's a bit of an extension of mechanization, because if you take a doctor's stethoscope, well, that's also, a certain form of, not AI, but, in any case, it increases his perceptual capacities, etc., so it goes in that direction, and strong AI, well, it's really not for tomorrow, not for tomorrow at all, also because of technological problems, of computing power, we'd need, well we're also reaching limitations, but really physical ones, on increasing processor speed, from a physical standpoint, because we can't manage to reduce what's called the manufacturing process size any further, etc., and also from an algorithmic standpoint,
1:39:38 and I also encourage you, if you want to go further, there are also, good videos, by Yann LeCun, who is, sort of, the father of the revival of AI, in any case, on machine learning, and who has good ones, that show that, in the end, well, before, the machine, is endowed, not even with consciousness, but can do things, that seem very simple to us, well, there's still a lot of work, so before the machine doubts, right, there's still a lot of work, if it's even possible, if that's even possible, there's still a chasm, even if it's perhaps not unbridgeable, there you go, you took the words, I could well have found my own words, thank you for giving me, so, precisely, I'm very glad, about the question, from, I hope I'm not mangling his name, from, whoever, Huiel, about doubt, because, well, not only does that not contradict everything you've just said, but, it complements it in an almost orthogonal way,
Doubt, a minus that is a plus 1:40:48 well, let's not, let's not be confused, I'll try to be clear, man doubts, the machine does not doubt, for now, that's where we stand, I don't, I don't confuse doubting with glitching, which is quite different, doubt, as you rightly said, presupposes a consciousness, I arrive at a crossroads, should I go left or right, there, I have a doubt as to whether I'm not about to mess up, it's, it's very different, from the machine which sets off, without hesitating, in a direction that's not the right one, for the operation, but, that's not my point, my point is, in every case, we say, weakness of man who doubts, and strength of the machine, I don't know, that doesn't doubt, those who, I was about to say, who share my hair color, who are my age, but, certainly, many young people know it too, I'm thinking of the adventures of Blake and Mortimer, it's not, Edgar Jacobs, I believe that, it's not confined to my generation, I think, of The Mystery of the Great Pyramid, The Secret of the Swordfish, and above all, masterpiece of masterpieces, that's a value judgment, I'll allow myself,
1:41:58 The Yellow "M", now, in The Yellow "M", Professor Septimus invents the tele-cephaloscope, doesn't he, which allows, his own Frankenstein, the man he put the tele-cephaloscope in the brain of, to have no inhibition, no moral judgment, no superego, we'd say, in analytic terms, which means he doubts nothing, and is fabulously efficient, in my opinion, that's a misreading, right there, I come to what I believe is the essential point, and there, I'm grateful to Huiel for having raised the question of doubt, it's that this minus in man, the fact of doubting, which is a minus compared to the mechanical perfection, so to speak, of the machine, for me, that's a plus, I'm drawing a disturbing parallel, that some of my listeners, who are starting to be familiar, some of my listeners, between the fact that the great secret of the contemporary economy, since the capitalist mode of production,
1:43:08 to speak in Marxian terms, appeared, the great secret of the extraction of surplus value, according to Marx, comes from a minus, not a plus, that is, from unpaid labor, from a part of the working day, or of working time, that is not paid, that is siphoned off, that is subtracted, that's one, two, the myth of Prometheus, the superiority of man over other animal species, comes from a minus, let me recall, let me recall the pitch, as told in Greek mythology, Zeus had entrusted Epimetheus, Prometheus's brother, they're all Titans, Epimetheus in Greek means the one whose mind is not quite sharp, precisely, who isn't very bright, basically, he told him, you're to distribute qualities to the animal species, so that each has its chance, to the elephant, you give strength, but without agility, to the bird, you give speed, but without strength, there you go, that way, everyone has a chance, but the other one made a bit of a mistake in distributing the qualities, when he got to man, there was nothing left, you know the story well, man is naked, the naked ape, as a famous book puts it,
1:44:18 isn't that so, man is naked, he doesn't have really effective claws, compared to the badger that digs its burrow, he has no claws, he has no fur, he has no scales, he has no, you see, and so, he was in a position of weakness, there's my second minus, after surplus value, after unpaid labor, now, from this minus, when he goes to ask his brother Prometheus for advice, how am I going to manage, Prometheus tells him, don't worry, I'll take care of it, and Prometheus goes to fetch fire from the gods, that is, intelligence, but not artificial intelligence, fire, that is to say, judgment, consciousness, and this makes it so that this minus becomes master of the other creatures, and it's said in the Old Testament that God gave Adam the privilege of naming the creatures, of giving a name, and not merely an algorithm, that's interesting, a name, the notion of name is complicated, this is an ox, this is a microphone, this is a keyboard, that establishes a discontinuity in being, like that, which isn't obvious,
1:45:28 at the tip of a dog's nose, for example, which is a kind of continuum of impressions between outside and inside, so let's not complicate things, third example, well, my robot from earlier, the one from 2001: A Space Odyssey, I alluded to its song, everyone, everyone who has seen the film, says themselves that it's pathetic, that moment, of the agony, in what way can the agony of a machine be pathetic, after all, if it's a machine, but precisely, the robot never seems so human as when it is on the verge of death.
1:46:07 And Dominique, if I may, there's also a line it says at one point, it says "I'm afraid," I believe, if I remember correctly. No, but that's exactly why, "I'm afraid," that means it senses an end coming, a limit, and there, I'll conclude my point, so to speak, with a quotation, it's not the first time I've given it, but each time I bring it back out, it takes on even more meaning, from the poet Hölderlin, isn't that so, Hegel's great fellow student, who says "for not all things can the gods, but mortals, that is you and I, before them reach the abyss."
1:46:48 There's the limit, the end, so that with them, the echo changes, which is a nonsense in terms of an inequality, in mathematical terms. The echo can at most give back the sound it was generated by, but no more, isn't that so? At most equal to, but no more. And there Hölderlin says, "but with man, precisely because he knows he will die, because before them, before the gods, mortals reach the abyss, thus with them the echo changes."
1:47:24 So, it's no longer just stimulus-response that works as the model, the echo changes, for long is the time, that of history, no doubt, but the true comes to pass. That's how, provisionally, I'm not resolving, how could I resolve, that would be megalomaniac on my part, this kind of problem, but that's how I'm orienting a range of answers to QL's very inspiring question.
Understanding and reason: the machine has no purpose 1:47:54 There's a remark, ah yes, a good one, that's often also a joke we make among AI colleagues, we often talk about "natural stupidity" as well as... Yes, natural stupidity... That's not far from what I just said, meaning that... Exactly. Man's superiority is his stupidity. Yes.
1:48:24 Compared to the machine, which has none. I'll pick up too on the point from the nameless idiot. The nameless idiot, but not without malice, I don't mean malignity, he says, doesn't the machine possess a pure reason in Spinoza's sense, a knowledge of the third kind? That is, the supreme knowledge, Spinoza, the highest stage of knowledge, that is, the knowledge that a dog or a fly has just as well, that is, sensory knowledge, a sensation, a stimulus.
1:49:01 The knowledge of the second kind is understanding, concepts, if... then. And then the third kind is, that is, which is much more all-encompassing and synthesizes the two, which is as immediate as sensation and as thoroughly mediated as that of the second kind. But there, I'll correct, the nameless idiot, if I may, on one point, it knows without doubting, it's a dead intelligence. Well, precisely, that's not reason. That's what's called understanding in philosophy. Understanding, that's if... then.
1:49:34 That, the machine can access this form of intelligence. Right? I'm not saying anything bad about understanding, reducing it. And I refer here those interested in this problem, especially in the year Hegel is on the agrégation syllabus. Let me recall that no one was more critical of reason reduced to understanding than Hegel. Let's say understanding is a bit like mechanical reason, if you like, it's not true reason.
1:50:07 But conversely, Hegel rehabilitated understanding against the Romantics, who would have liked to dispense with what nevertheless made possible the development of human intelligence in the universe. So, once again, let's not confuse reason with understanding. Now, so as not to be frustrating, I'll state the difference. Understanding is if... then. And the text that, that Christophe proposed, from Book VI of Plato's Republic, is very clear on this point.
1:50:42 Geometers, geometry being typically a discipline of understanding, proceed, says Plato, from hypotheses toward a conclusion. Instead of working back up from the hypotheses, in the opposite direction, toward the unhypothetical first principle, anhypotheton, which is what reason must do. Reason, and I'll give Kant's answer, is the faculty of ends. A machine has no purpose. It has a function given to it by another, but it itself has no desire of its own of which it is the first source.
1:51:15 Isn't that so? There is no teleology of the machine. There is only teleology on the part of the one who assigned it a function, like the car that parks itself without me being at the wheel, or that reads traffic signs or the rules of the road so as not to hit anyone. Those are functions that a being who is not a machine has assigned to it. There you go. Please, go ahead. I find what you're saying, yes, quite...
The tool and the purpose of its designers 1:51:47 I think we also need to insist on this point about ends, because we tend to talk about the purpose of the machine in order to hide the purpose of the machine's designers, which is a bit... There you go. And that's why we also talk, at some point, about... Debates about the machine's responsibility, etc. If we go back to self-driving vehicles, on...
1:52:19 It's a way of letting the designers off the hook, after all. We mustn't... And of having more insurance-type systems for that sort of thing. But that's why I always find it a bit hard each time when... As I said earlier, there's a lot of... There's work on algorithm fairness/loyalty, etc. Putting purposes onto algorithms. And of course, we can say by transitivity, we're talking about the designer, but not necessarily.
1:52:56 So you have to understand… That was a book about the saw earlier. Right. You have to look at the function of the machine and what purpose lies behind that machine. Personally, I'm a bit blunt on this point. When I see people putting technology on trial in general, in the Heideggerian style, for example, I say, if I… I'll speak slang, if I stab someone with a knife, it's not the knife that's going to be sent to jail, for heaven's sake.
1:53:29 But if you do… Imagine that… You know, you take the knife and you take a killer robot, what's called a killer robot, drones, etc. Who knows whether we'll also send the operator to prison if he makes a mistake, etc. So that's also the question, this question of tools. I think we still need to be careful about all the possible drifts, not to consider these systems as tools that have no purpose of their own.
1:54:09 Because after that, we can still end up with a monstrous drift. Of course. If I may, at that point, what's undeniable is that the more powerful the tool, the more it increases our power. Power to do evil as much as power to do good. But it's not up to the tool to decide whether that power will serve good or evil. Yes, quite right. That is the privilege of the subject and of the collective subject of the general will, to put it in republican terms.
1:54:42 Right. And if I may pick up on what we were saying, for example, about chatbot systems, or in French, conversational agents. I think it's also important to look, if we come back to language, at the fact that the AI object is never the subject. If we take, for example, the characterizations of the Palo Alto school, for those listening, they tried to model, in a language, that there is a subject, there is the one, there is the channel through which the message passes, there is the receiver, etc.
1:55:26 and what's always a bit disorienting is that the conversational agent does not carry the purpose of the message either, it is not the subject of the message, and that can lead to drifts in terms of— we can become completely lost, imagining there's a subject behind it, and it also prevents us from seeing what the purpose of the designer of that conversational agent actually is.
1:56:01 Well, I think that, thanks to your questions—thank you—we've covered a lot; we won't keep you from sleeping either, I mean, because what Christophe is developing with his research topic is so fascinating we could spend the whole night on it, but I know some of you have young children. Yes, go ahead, please. Perhaps, just to answer fairly quickly to Abderrahim, Driss, who raised several things I'll respond to.
Complementarity, data, the Kantian schema, and labor 1:56:38 He's the one I owe a vendetta to, because he says Corsicans are genetically tireless. He knows us well, bravo Abderrahim. So, I agree completely on several points. There was this first point: AI and human intelligence cannot be compared, they're rather complementary. I think that's quite true, and that's exactly, for example, if we take a radiologist, well, there will be AI systems that increase his capacity to perceive a tumor he might not see with the naked eye.
1:57:24 Right. But the diagnosis still remains within the radiologist's domain, since at the outset people said, that's it, with these AI systems all radiologists are going to be out of a job. There have been studies, actually, in the United States, showing that the task of various professions—the radiologist's, for example—doesn't just come down to looking at an X-ray.
1:57:56 There are all sorts of other activities that can't be delegated to an AI machine. So, to sum up, there's a complementarity, and also, from a somewhat less scientific point of view, I would find it so, so sad if we reproduced an intelligence that was just like ours. That would be a bit… a bit Doubtful. Right. And I'll finish there.
1:58:26 Yes, go ahead, go ahead. And if there were another—right, thank you, thank you very much, because I think this remark is really, really judicious. Between representation and reality, there exists an entire universe of data. So, in two senses: data is not reality, we manipulate the data—we haven't talked about it, but there's also the existence of biases in the data, etc.
1:59:01 But actually, we manipulate data, and depending on how the data is manipulated, we can get different results. And another point is about representation—between representation and reality there can also be a difference, since, how to put it, the algorithms used in deep neural networks tend to cut the link we have with the physical phenomenon.
1:59:39 Why? Because with the data, we actually use mathematical optimization techniques that— you see, when you have a deep neural network, when you have layers of neurons like that, it's to shift the data into other dimensions. So, we shift into other dimensions, and the data also tends to become more regular.
2:00:09 For example, if you have a phenomenon that's somewhat discontinuous, a physical phenomenon, I don't know which one, but for the optimization algorithms inside to work, certain mathematical properties are required, and that can cut this link between the physical phenomenon we wanted to reproduce and the result we get at the end. So there's this double thing, this interface between representation and reality, and that's why the figure, like the demiurge, can also be a first approximation.
2:00:55 Yes, so, very good example, we find again this good old potter, this demiurge-potter, and what you just said about the interface played here by formalization between the data and the subject makes me think of a fundamental Kantian concept, not the simplest but one of the most fruitful, that of the schema.
2:01:25 What is a schema for Kant? It's something that precisely allows for an airlock between two entities that are normally never in contact, that are opposed, dichotomous, and which allows them to be put in contact—unity, plurality, the notion of number is a schema for Kant that allows them to be brought into contact, for example. Right. So, well, we're not going to—as I said earlier, I feel ready to spend the whole night on this, but I don't want to impose on our listeners.
2:01:55 I believe, and Christophe and I agreed on this while preparing tonight's session, that the material he's brought us is so rich that it deserves—I don't just believe, I'm sure the majority of you will agree—it deserves a second session, it deserves a follow-up which I'll leave to you, Christophe, if you're up for it. Yes, absolutely, though we'll have to stop soon, but there were still really interesting questions and remarks, notably those that were posted, but also Daniel's, who talks about how with digital power we simultaneously lose complexity and mastery, and I think that still shows the purpose we want to give the tool.
2:02:46 Right, that's— but that goes much further, it's also about the kind of world we want to live in, do we want to use—and we'll come back to this—what do we do with the productivity gains we get from artificial intelligence? Is it to make work even more mechanical, or is it, for certain repetitive tasks, to let people focus—we mentioned innovation perhaps—so they can free up time to do something other than work, since if you look at history, productivity gains have also allowed people to work less and do other things alongside, which are very good for society.
2:03:43 And so, to come back to digital power, that we simultaneously lose complexity and mastery—it's true that, well, I understand the remark about IT people, having lived through it myself. I worked in a group that did numerical simulation, and, as we saw—let's go back to what Chris Anderson said at one point—in departments like that where people had built up expertise on models, etc., people showed up who said, well, I'm going to take a machine learning model, everything you did before, you can throw it—I'm exaggerating a bit—in the trash, and I'll come along with my machine learning algorithm and get a high quality of prediction.
2:04:45 But, but, but what we forget is that all these things, at least from a computing standpoint, etc.—we've developed a lot of proofs of concept, but they aren't always used operationally, since, at some point, imagine we design an automated control system for, say, a hydraulic power plant, I'm just saying that at random—well, we can develop concepts, but when we go see the operator or the person in charge, he might hesitate twice, he'll say, show me how it works, how can it be validated before we put it in place.
2:05:30 But on the other hand, it's true there are quite a few societal problems, even in terms of work, that are pretty problematic. But the question we might ask ourselves is whether AI isn't, in a way, a fig leaf, I'd say, covering all the efforts, all the drives toward digitization, mechanization, breaking down people's work into little tasks, etc.—it plays into that.
2:06:04 That's a question that makes an excellent provisional conclusion, I don't know, but really, to put it that way, this fig-leaf idea, I love the metaphor, it's so fitting. Exactly. It shows what's irreplaceable—for instance, computing power, it's clear that there the computer easily dazzles us, I mean, computing power—I don't see how an individual's consciousness could just as quickly perform a billion operations.
2:06:35 And then, very quickly, still on the question you just answered, Daniel gives the example of bank employees. Now, here we're talking about purpose, it's quite clear there's currently a serious social crisis among employees of the banking system because their employers no longer want to pay rent. And so they're saying, stay home, that's the world of tomorrow. We can clearly see an example here of what Daniel is saying, can't we, stay home and you'll come in 20% of your working time to the office, but otherwise we won't need to rent office space in Paris anymore, which costs our financial directors an arm and a leg.
2:07:19 And there you go, stay home, with your kids under your feet. There you go, I… I… Daniel hit the nail on the head. Oh yes, I think that would be… Daniel knows the world of the capital-labor relationship well. So, right, listen, I'll give you an appointment for the continuation of this fascinating series. I'm hooked by Christophe's subject and I want to see what comes next. And otherwise, I'll subscribe, this will be my conclusion, to what the nameless idiot says, who apparently has gone to bed like others, like Abderrahim who wishes you all good evening.
2:08:01 The nameless idiot says yes, may there long be fishing in the morning, hunting in the afternoon, poetry in the evening, and love all night long. Listen, how can one disagree? I think that's a beautiful conclusion. Thank you very much and see you next time. Thanks to everyone, and thanks to Pierre-Alain and the technical crew who allow us to meet in such good conditions. Thank you, Christophe. Bravo. Goodbye. Good night. Thank you.
