Alexandra Ebert: Welcome to the 23rd episode of the Data Democratization Podcast. I'm Alexandra Ebert, MOSTLY AI's Chief Trust Officer. Our last guest for this year is Denis Rothman, a brilliant AI expert. Denis studied computer science and linguistics at Sorbonne University. He has a truly amazing resume from basically all over the world. He's a prolific AI practitioner as well as researcher with tons of important publications in the AI, explainable AI, and natural language processing fields. His ideas and his patents have influenced many people around the globe.
In today's conversation, I wanted to find out what Denis thinks about two of the topics we particularly enjoy talking about here on the Data Democratization podcast. One of them being ethical AI, and today we're also focused quite a lot on explainability. I'm sure you will enjoy this conversation. Let's go and meet Denis.
Alexandra: Welcome to the show, Denis. It's really great to have you here today. Before we jump into the first questions, can you briefly share with your audience a little bit about your background and what made you interested in explainable AI?
Denis Rothman: Okay. Well, thank you for inviting me, Alexandra. I've been in languages for a long, long time. I was born in Germany, so I learned German and English. Then, I lived in France where I learned French. Then, I traveled a lot, I studied in Europe and United States. Languages were always something I liked, and at the Sorbonne in the 1980s, this is a long time ago, we started playing around with computers to see how we could parse words and use expert system to analyze languages. I registered a patent on word embedding and then I registered a patent on chatbots and they both worked immediately.
I found corporations that liked it, and they funded me from there all the way to the end of my career, companies like LVMH at the luxury companies, aerospace, Airbus, companies like that.
Alexandra: Quite impressive list.
Denis: Yes, so just a long adventure. Then, I taught a bit at the Sorbonne, and then I was always self-employed and I didn't experience an AI winter. I never even noticed there was one. Then, in 2015, when Google came and made it mainstream, well, I was a happy person because I didn't experience an AI winter, but it was lonely. There were only a few people talking about it and now everyone. Now it's better because now I can share what I learned, what I did. In a nutshell, that's it. Then, I wrote a few books, and I'm an instructor. Let's say I spent my whole life in artificial intelligence.
Alexandra: Sounds like that. I can also imagine why you didn't experience any winter because with all of the activities I'm sure that you were busy. I specifically want to talk about explainable AI today since you also authored a book which is titled Hands-On Explainable AI with Python if I recall correctly.
Alexandra: One of my first questions would be, why do we need explainable AI in the first place? What's the reason for it?
Denis: Well, it's nice that we do this right after my short biography because when I started for an aerospace company, which is now Airbus, and I was in a defense project with the military, there was no way they would accept a result if you couldn't explain it. It's in the contract. If you're going to produce a result that's going to have strategic consequences, you have to explain it or they won't accept your system, so there are domains where it's critical. You can't go to a medical facility or a hospital and tell a surgeon, "This is the way and you just trust it." No, no, no you have to explain it. Explainable AI comes from the roots of necessity in corporations where you have to explain.
There's no way they will even pay you if you don't explain. Then, after I went into supply chains where I was managing warehouses, transportation, boats, ships, all over the world, there's not a single company that's going to say, "Oh, we like your schedule. You just scheduled $70 million of supplies. We don't want to know why, we trust you." No, you have to explain every single result and you have financial auditors that will come in and say, "Why did you purchase that there?" I would say it's only recent that it seems that people are discovering that, but in real life if you go to Blue Origin or SpaceX, there's not a single thing they do that they can't explain.
You imagine that, "Okay, the rocket's going off. Oh yes, let's trust this--" No, you have to explain everything. What changed thing is social media. This is why we're having this conversation because social media is random. People are anonymous. They can say anything they want and then you have billions and billions of posts going on social media, and you have these poor CEOs that know nothing. They don't know nothing about anything. People think, "Zuckerberg, he's doing this." No, Zuckerberg is surfing. He's surfing. He's in his yacht. Jeff Bezos has one of the biggest yachts in the world. Then, we go say, "Why did you do that?" He'd say, "I don't do that, I have thousands of employees."
Social media is very random because we type things in all sorts of ways that are with poor grammar, poor spelling, we're not polite, everything is biased, no one really checks any fact, everyone's sure about something. Then, that's where my book comes in saying, well, we do need some way to explain these more difficult things and maybe some judicial decisions like what are the judicial decisions based on because judicial systems are based on the opinions of a juror, of a judge, so how is this done? My book is full of methods already. You have Shapley, LIME, anchors.
We both live in Europe where we have the RGPD or GRPD, which you have several articles saying, "If you cannot explain an automatic calculation, no matter what it is, we don't care if it's AI or not, we can give you a fine of up to €20 million, maybe €500 million, and we can take 5% of your annual revenue if you can't explain why you did that." This is appearing in the United States, too. Many [inaudible 00:07:50] [crosstalk]--
Alexandra: You're referencing the GDPR now, the data protection legislation, or the upcoming AI Act because of the [crosstalk]--
Denis: [crosstalk] [inaudible 00:07:56] GRPD or GPD, yes, contains these articles. I think if you go to Article 9, you have to explain it. You can't make racial references in Europe. You can make them in United States, but less and less. Recent events showed that they're going straight towards European law. They don't want to hear about a race anymore. They're trying to get rid of it. It might take longer, but anyway, I would say it comes more from social media today than in corporations where it's really structured and it's been structured ever since computer science got there.
Alexandra: That's quite interesting. What makes social media so difficult to explain and how could you achieve explanation that is satisfactory to those who are the target of the explanation?
Denis: I won't go into the clichés we're all talking about all day, social media, this, and that. I'll go straight to what I think, okay. It's my opinion. I think it should be forbidden for a person to go on social media without revealing his identity. I think there should be an ID card, ID which is a problem in the United States because many people vote without IDs in the United States, [chuckles] so we won't understand that in Europe. I vote in the United States and I vote here, but I always give ID, but anyway. What I'm saying is we should have ID and it should be scanned and it should be verified like with a bank, a double-check, with our phone to be sure it's us.
Once your identity is there, I'm sure that 80% of the problems will disappear.
Alexandra: You're really sure? I know so many people that with their full name opposing so many things that arguably are not true.
Denis: Sure, because true and not true is another problem. Then, you need a second layer, which is accountability, local accountability, that means that if someone in I'd say, South America, insults someone that lives in Germany, then it goes into German law. I would consider the physical place of the person be insulted to be the law. Now, you insulted a German person because of his name, and you went back and gave them bad references, you insulted that person, you're responsible under German law.
Now, we don't care who you are, and now it's out of our hands, it's not up for you, and need to decide. We have the identity. Someone filed a complaint, in Germany, it's very easy because you can go on websites now for German law, and you can enter a complaint there, and it's up to a judge to decide, it's not up to us. It can't be the president of Twitter, no, it has to be a judge with the law of that country.
Alexandra: I agree that a judge has to decide, but I'm not fully following how does one person insulting another person correlate or has to do with explainable AI and the algorithms that are deployed in social media?
Denis: Because when you have 1 billion posts in one day at Facebook, you can control nothing. Manually, it's over. You can't control it with SQL. You can't control it with classical things, so you have to perform sentiment analysis. You're going to have to run sentiment analysis on the post, on all of these posts automatically. Then, you're going to have to put it through something like LIME, and you're going to have to have an HTML page, which says, I'm highlighting why this went to negative.
If someone thinks that that negative thing is something he doesn't like, he can make a complaint just like he can do in everyday life, and it's up to a judge to say, "Okay, I have your LIME thing, I used the algorithm. I don't find these words insulting, I can understand you felt insulted, but under German law, this is not insulting, or under French law, but maybe under Austrian law it's insulting." That's why I say you need the identity and the local law because you'll never find an international agreement on anything, it doesn't exist.
Alexandra: That's true, but one question I would have in that regard is, of course, it makes sense if a judge decides what's crossing the line and what isn't, but there are many studies out there showing that sharing information that, for example, is not true or even insulting to a specific individual or a group of individuals, and then deleting that content at a later point in time, or even posting some statements of apologies, doesn't undo the damage if the content already went viral, or at least reached a significant amount of people. Shouldn't there also some mechanisms be implemented by the platforms to prevent, for example, hate speech or some of these things happening?
Denis: For the moment, let's keep the platforms as the last resort. I can give you precise examples I know here in France. I know a person that is the head of a high school that got insulted, and very badly. Then, exactly what you said happened, the person had taken it out, but he had made screenshots. He went to the police, and the police can go to the platform, and they have to have the backups of all this. It's a legal thing. This person was a student, and the student was banned from the high school. In law, it's not because you steal something and then you got out of the house before the police got there that you can do something. If you have the proof, screenshots, or something, then you can do it.
I would say first, we need the identity, and then we need local law to apply. Now you're saying what we're saying, the platforms. Now suppose we're both the head of a platform you and I. We're sitting here this afternoon and saying, "Denis, I'm not so happy about our platform, we have 500 million people, and I see things I don't like," and I say, "What don't you like because what don't you like, where does that start, and where does that end?" For example, if I'm an Iranian woman, maybe I don't like what I see European women writing and I find that that's very negative and it's very destructive and the way they speak about me and the way they speak about Iran maybe I don't like it."
"Maybe I'll answer all you Europeans." What I'm saying is, the cultural differences are extreme in our world, our world has stretched to a point where we still have people with beliefs that can go back, I'm not giving any judgment what people believe. I mean, people believe what they want. They can live in Iran, United States, China, I don't care. That's what they believe. How am I going to judge? Suppose I'm talking to a Chinese person because people always take as Chinese, Russian, Ukrainians, they're the devils and we're God, okay. Let's take 1.4 billion Chinese people, you're not going to tell me that they're all bad. I mean, they're all like us.
Every day, they wake up, they want to go to work, then they come home, then they want to do some leisure, and this person says, "It's not my fault, I like my country, I'm a nationalist. I like my country. No one's bothered me, I live 600 kilometers from Beijing, I never even saw a soldier, a government official. I live in a little village, I have a little internet connection. What did I do to you? I like my country." The other person, "No, your country is dictated." "What did I do to you?"
Alexandra: I see in which direction you're getting, so definitely it's really the [crosstalk]--
Denis: Where do we start and where do we begin, right? It's very tough.
Alexandra: Definitely [inaudible 00:16:30] [crosstalk] people have different values and that it shouldn't be up to the platforms to decide what's right or wrong. We now also see or have seen the examples where news, fake news, and other information published by these platforms [crosstalk]--
Denis: But what is fake news? Yes, finish, please.
Alexandra: Published by these platforms, even shifted elections and influenced people. I think, currently, we see these platforms being one of the main source of media for many people without having the same obligations in carefully curating and editing their content as traditional media has. There, I think, only waiting for some instances where after weeks or months a judge makes a decision potentially helps with some problems we see with these platforms, but not the immediate effects that could happen if some of these materials go to a large audience. Then, in this point in time, for example, influenced one election.
Denis: Yes, so what you're speaking about now is the United States because I don't see problems like this in France. I don't see problems like that in Germany or Italy or Spain, I don't hear about that. I don't hear about presidents being banned in Europe. I don't hear about that because in Europe, we've gone through 1000 years of terrible wars, and we're a bit tired of all that. In Europe right now, it's more of social unrest. It's not left and right. It's more bottom and top. I'm not earning enough money. The gas is too expensive. COVID I don't understand what the government is saying. It's more a bottom-to-top thing than left to right.
That's the feeling I get in Europe because I'm a full European. My mother was born in Italy, I was born in Germany, my father was born in Russia, I live in France. Okay, so I'm really European and an American at the same time, but the United States is something different. We're talking about big tech, we're talking about the United States. Now, big tech, in the United States only inherits what I've seen since I've been a child in America [unintelligible 00:18:52]. The country is divided into huge blocks of Democrats and Republicans. We don't have that in Europe. In France, for the presidential elections, we have about I would say five or six candidates and none exceeds 25%.
In Germany, you can go nowhere without a coalition, it doesn't exist.
Alexandra: If I recall correctly since you said we see only these problems in the United States, also with Brexit, it was highlighted that behavior of voters was influenced by advertising and information displayed via Facebook. Also, this is a problem that I think needs to be tackled, and it's not an individual insulting another individual, but on a broader societal level.
Denis: Okay, so take Brexit because I like the English so much that I watched the discussions in their parliament, all of them, I watched all of the [inaudible 00:19:55] [crosstalk]. Every day I was hooked on it because I just love the way they talk, I like their accent, and I like the way they're so polite with each other. Even when they get mad, they're so noble and they speak so well, there's no vulgarity. Even when they're extremely mad, they're still polite. If you see Boris Johnson getting mad, he's still polite. Of course, I'm pro-European and I'm not for Brexit.
It saddened me to see Brexit, but I can say what we're saying here is not true because I watched all of this and there was a vote. There was a vote influenced with buses and advertisements and all that. That came to the British parliament, and you had the opposition because I watched all of these. I watched all of them and the opposition was saying, "You influenced, you brought us into a situation, and it's going to be a disaster." Of course, I don't speak as well as they do. They don't speak like that, "Of course, you went into a disaster, and they don't speak like we do. They're so polite. It's so beautiful to see English democracy.
At one point, you have Boris Johnson saying, "Okay, if it's like this, why don't we go back and vote again?" What maybe Europeans didn't see is that Boris Johnson say, "I dare you that I'm going to now call for elections, and we're going to renew the parliament." It was sometime in October, I don't know which [inaudible 00:21:25] year and he said, "Well, let's go back." The opposition said, "No, no, no, let's not go back," because in fact, in general, the English population felt that Brexit would be good for them. The fishermen thought it would be good for them, the farmers thought it would be good. The population, whether they're right or wrong, they thought that.
Boris Johnson came along and he organized the election again. The election brought his majority back to the parliament. In the end, what is the real deep influence of media? If you take the American elections, I would say the Americans right now are just a little more Republican than they are Democrats. That's why the elections were so close. That's why it wasn't a landslide for Biden, it wasn't a landslide, it was so close. It's just that many Americans just wanted to get of Trump, but not that many, just maybe a few percent, so it brought it to an equal level but if you go back in time, you will see that in 2016, Hillary Clinton said that the elections were all rigged by the Russians.
You can go back to Al Gore in 2002 that said the Republicans rigged the elections in Florida. You can look at every single election. There was only one election where they agreed but every time, no, because it's not a problem of social media, it's a problem of organization. In France, mail ballot was banned in 1973, '74 because of these problems. It was banned because everyone was piling stuff in the mailboxes. If there was a system just where you present your ID, and then you can only vote for one person, you can go get 500 people. I would say the media only amplifies because I was a lot in the media, in fact, in the 80s.
I was very often in the media, on television, and all that and I was always asking the question you're asking to the journalists and say, "We only amplify what's going on at a lower level. We don't create, we amplify and we follow the movements, and if the movement doesn't and then we follow another movement."
Alexandra: That's true but still, from everything I know from journalistic practices, they have to follow more rigorous approach of selecting whom they're amplifying and how they're amplifying the content, fact-checking, and I think this is something that's not necessarily happening with social media but I have the feeling if we continue on this topic, we won't cover explainable AI.
Denis: I can give you an example [inaudible 00:24:13] [crosstalk]--
Alexandra: I actually want to move to explainable AI.
Denis: Yes, move on because [crosstalk]--
Alexandra: What I would be curious is to get your perspective on what actually makes a good explanation and also for which audience or how should an explanation be different for a developer model versus the end-user affected by it? What are your thoughts on that?
Denis: Okay. As I say, these are only my opinions and by now you know that I believe in all opinions [chuckles] and I'm not shocked by any opinion because the French is [unintelligible 00:24:49] where I'm not manipulated by anything, so I don't mind. My opinion is first, model agnostic, ignore the model because no one understands the model. That's what I think. Now we have Microsoft just produced the Megatron Turning Transformer with 500 billion parameters.
Alexandra: That's a lot.
Denis: Right now, I'm working on the research and the math of a trillion model transformer that can do more than a million tasks. We're talking about [crosstalk]--
Alexandra: Trillion parameter model or?
Denis: Yes, trillion parameter model with 1 trillion parameters.
Denis: Okay, and it's going to come out as Google Pathways very soon, but we already have the Turning Megatron with 5 billion parameters and we also have GPT-3 with 175 billion parameters and we have Google BERT that's somewhere in this size, and Google Search. Google Search is powered by Google BERT. Now, just to prove my point in my opinion, someone can come along and disagree and I would be happy to listen and I wouldn't get into any kind of [inaudible 00:26:16] [crosstalk]--
Alexandra: No worries. Today's is about your opinion, so, please [inaudible 00:26:19] [crosstalk].
Denis: I don't care. I can listen to another opinion without getting all worked up. I can listen to that perspective and add it to mine, but what I'm saying is, okay, find a way, you have 100 layers, you have a model even with GPT-3, we have 96 layers, 1, 2, 3, 4, 5, 6, and each layer does something else. The first one just looks at the definition of the word. Then, you go up a little more and now it's looking at the word and the context. Now it goes higher and it's looking for long-term dependencies, maybe a word at the end of the paragraph where you just get the name of the person after someone said something for like five minutes. Then, when you get to the top, you get a huge relation.
I would say it's impossible. Now, of course, people can believe that they're going to debug it, but they're going to have to find how because they were trained with supercomputers. You need a supercomputer to debug it and find what's going on. I believe in model agnostic explainable AI, I don't care about the model. I just want to know what I said and how it's interpreted. For example, I'm saying social media is good because I think it [unintelligible 00:27:42] communication between scientific communities. Then, I get something social media positive on the other side.
I get someone says, what is this person's say, I just enter that, and the other side, sentiment analysis, it's a good opinion, but then I'm thinking, why would someone think that social media is good? What I'm saying, someone that thinks the opposite of what I'm saying says, "How can that be possible? Who's that person that thinks--" Let me try LIME. Let me enter LIME which is in chapter eight of my book. It's local interpretation, that's the L, which means it's local, means you're just going to that input and you're just going to that output and then you get this nice little page and you get the highlight where you see, "I think social media," good. Now, it's highlighted because it helps scientific community.
He likes social media because helps the scientific community but then what about the second sentence? Second sentence comes in. It says, "I hate social media because people are always writing bias and fake news and everything and it's horrible." Then, the answer comes out. Prison says, "Oh, what's going on here? Schizophrenic or something because I just got the system saying it was good, now it's bad but why did it say that?" Ping, I click on LIME, ah, because now he's speaking about bias, fake news, but that's not exactly like the scientific one.
Then, you can have classical AI, no AI, classical, take that pack there, and you can build a rule-based system which will count which ones are good, which ones are bad and then you can make a little summary. A little list with the good ones on the right, the bad ones on the left, and you can parse the words to make a little dictionary or you can use GPT-3 if you really know how to use it and it will do text classification automatically, put them in two categories and give you the keywords automatically. It can do that. You take the text, it'll give you the keyword. Then, you can build your dictionary that says, this is why people like it and don't
Denis: like it.
I didn't talk about the model here. Talking about the input and the output.
Alexandra: If you say dictionaries and classifying some words as positive, others as negative, how [crosstalk]--
Denis: In the context.
Alexandra: In the context. How would it be, for example, with all of these models that learn, for example, I'm thinking just of this example of men relates to doctor, woman relates to nurse. You would never be in a position where you could say, "Okay, doctor is a bad word or nurse is a bad word," or something like that. How would it work for these more nuanced examples where it's not like a wide [inaudible 00:30:39] [crosstalk]--
Denis: That's interesting because I was having this conversation with my wife only two hours ago because my wife's a linguist too. She studied five languages. Sometimes I just like to have these kind of conversation. I was working exactly on this topic with her. We talked about it for 20 minutes while we were walking outside because I say the whole gender analysis is changing. It explains better to you why I don't care about opinion. Linguists just are scientists. We just see what's going on and we write it down. If a word is used a lot in the language, we put it in a dictionary. It's not up to us to decide if this is good or bad or gender, non-gender, multi-gender, it's not our problem.
Our problem is to see how people are using vocabulary to describe something and giving proper definition. There's a corpus called [unintelligible 00:31:48] gender. You have in natural language processing, we usually use corpuses that are handmade by people and they write sentences, and the machines learn from these corpuses. It's not only raw data.
You have these corpuses that are made by a lot of people, either professionals or open source. Then, they write sentences and they say, "Is this right or wrong?" There is one that I was talking about while I had this conversation, and I said, "You know I was looking--" because you're beginning to know me and I go into detail.
This morning, I opened all these corpuses and I was looking at everything that was in it. I was looking at problems like you just said. I went into this [unintelligible 00:32:38] gender corpus, and I say, "Oh-oh, they're going to have a problem." Because in French we don't have that problem. We have [French language]. We have the male nurse and we have the female. We don't have that problem. I was saying in English to my wife, we have problem [unintelligible 00:33:03] nurse, we don't know who it is. We're saying, "The nurse just entered the room. She went to see the patient." It's not exactly the sentence, but it's in the [unintelligible 00:33:14]. It said, "Is she the nurse?"
I told my wife, "We can't answer that question anymore because why would the nurse not be a man? What if this nurse is transgender? What if this nurse doesn't want to be qualified as man or woman?" Because I don't like the word in English, female and male. It sounds like we're animals, I prefer man and woman, or whatever people want to call themselves.
I say it is a big problem and they're lagging. The corpus is a bit lagging. It's a bit old. Yesterday evening, since you're beginning now to know me even better, I spent a lot of time on Stanford University on their website analyzing something that history has always taught us. I found an extremely good article because I really admire Stanford University.
For me, it's like the Sorbonne or any beautiful university. I was analyzing one of their articles and I read it word by word. Sometimes I spend hours doing this, and I was saying, "What year does this look like to me?" I said, "This is old." It was 2014, and that's before transformers. I say, "Whoa, that's a problem." Then, I go to another article where they have a center for transformers, and I read the 200-page article weeks ago, word by word. I printed it out and I analyzed it. It's funny because there's not a single reference to an algorithm. When I go deeper, I can see they don't have access to the algorithm because GPT-3 is not communicating it. Now they're trying to build their own algorithm to do things.
I say, "They're lagging there too." It's lagging. The truth is history is stretched. That's my belief. History is stretched. You have people like me today, I'm going to be very careful when I'm talking to someone about gender because I have an excellent friend who is a transgender, a person that was a man that became a woman. The woman's a fantastic woman. We have conversations, of course, about data science and artificial intelligence, [laughs] so I'm very careful about that. Then, you have people that come from former generations who have former state of mind, but not be ready yet to speak like that, and there are tensions.
Tensions may grow, but that's history. When horses were replaced by cars, I don't know if you remember this, but in 1914-
Alexandra: No, I was not present there.
Denis: Yes, but you can remember this in history. In 1914, the French still had the dragoons on horses, and they thought that they were going to go into the World War I trenches with their horses and do something. They were all cut off by machine guns. Because what I'm saying is it takes time. There is a serious problem with gender. It's a very serious linguistic problem because it's stretched. Some people are accepting, for example, in France, the word il elle is now in the dictionary. You have il which means he and elle, which means she, and now a new word is created is il elle.
It's like saying, I don't know, a he/she, that doesn't really know yet who he or she is, or may never know because we have to accept that a person that's a teenager doesn't know.
Denis: Maybe it might take a lifetime to find out, and many people know and all of this creates-- I'd say the problem is not solved at all. What I know is on GPT-3 OpenAI, if you start this, you'll get a filter saying you are not following our rules and you might have a problem. This problem has not been solved until humans are more tolerant because you can do the same thing with color. For example, you're White, but suppose you marry a Black person, you have a child. Is that person White or is that person Black? You have 50% genes. It's not because that person has a skin color that's brown that that person's Black has 50% White.
The whole discussion, for me, makes no sense at all because everyone on this planet is a mixture of something. That problem hasn't been solved either in the United States. I don't know what proportion it's taking, but these problems are going to take maybe centuries to solve.
Alexandra: That's interesting. [chuckles] So many questions that came up, so what to ask next? Initially, you started out when it was about what makes a good explanation, and you gave an example of language input. Can you give us also more abstract example of non-language-related models, especially how the model-agnostic explanation would be helpful for, I don't know, for example, for a bank in the financial context or insurance provider?
Denis: Let me go from what I know, to something I know less which is banking and insurance. What I know is supply chains, and I know supply chains in banks and insurance as well because they are supply chains. They have to have servers, they have to have documents. In this world, when you say, for example, let me speak about banking in the supply chain thing, I have to transfer this money from this bank to that bank. I'm not talking about customer transfer. I'm talking about billions of euros. I have a supply chain and I have to feed €5 billion into this branch and that country in that way.
That's a supply chain decision, but the problem is I only have €4 billion right now in here and I have to distribute it to several locations. I don't have all that right now, so I want to find a way to figure out how much each location needs for the next two weeks. I'll give you 500 million, I'll give you 1 billion. Then, I've made these type of calculations because I wrote the supply chain engine. I wrote this. Believe me, you better have an explanation, and it's certainly not about the algorithm. What I generally do is it's in subject matter expert explanation.
You have to produce, for example, if we take this example, you have to show at the beginning, like in a bank account form, you had five billion, but then in this order, you had to give out that much money. You had five locations, and you can't give five billion to each location, but then when we did the forecast for this week, that was the need. I scheduled the need and then I said, "Now I did an automatic distribution for one week." They can see exactly, on a Gantt, for example, exactly where that money went and why, and they say, "Okay, I see, but I don't agree." They say, "If you don't agree, change it. You can click, it can change the amount, then you can run it again to see what happens."
If you go into aerospace, you have sensors everywhere. There's no way anything is going to be done without sensors. Even welding, a little thing like that you want to weld, or a little screw that you've put somewhere, that is traced in detail, and often with images. At that point, if I say, "Okay, there were--" No, I shouldn't say this, but it's true. Sometimes you have, I won't say which airplanes, you have 1000 screws and you can decide that some of these screws fell off during bad maintenance in one country, which I won't mention either, and then these screws fell off. You can say, "Well, the plane can still fly. There's 200 screws missing, let's do another flight and see that after." Or, "It's going to cost a lot of money."
You didn't hear me say that.
Alexandra: I didn't.
Denis: [chuckles] I didn't say the airline, I didn't say anything.
Alexandra: All good.
Denis: They do that every day, everywhere. That's why you say when you get in the airplane, when the door is shut, it's shut. Then, you do that calculation, you say, "Okay, you only gave me 300 screws, and I have two airplanes and there's 500 screws missing on both of them." Now, you have the critical locations one by one, and then you have to show it and you put it in a visual explanation as well. You can say, "Okay, I agree, but someone has to sign that. There's no way it's going to go on its own." Sometimes you have 10 people that have to sign, "Okay, we're only going to put 200 screws." You need three signatures, sometimes 10.
I would say in industry, banking, insurance, I don't really see the problem because everything is so structured, and everything is so analyzed. The New York Stock Exchange is managed 50% for a lot of equity transitions by machine learning. It's fed back into the stop-loss systems, into classical systems, and it's refiltered and controlled. It's not running on its own. It's just because it can go faster. They can go see what's going on on online shops, they can go see what's going on in competition. They can say, "Okay, the wheat prices are going up in this country, so the wheat prices are going to go up there. That's good, maybe I can sell my shares in two weeks."
Machine learning, but then it goes into the legal process of stop-loss regulations and the contracts you sign. I'd say in banking and insurance, I'm not worried because no one does anything without regulations. I'm talking about the big ones in the big countries like United States and Europe. I'm not speaking about any other countries I don't know, but in the main countries, I know it's extremely controlled.
Alexandra: When you say you are not worried from a customer perspective of one of these banks, you're not worried from the standpoint of the bank having some models that can't be explained or where people [crosstalk]--
Denis: No, no, they are explained. A bank won't accept a model that's not explained. No bank is going to say, "Okay, I like your machine learning thing. Tell me if in the south of France, I should insure that house or not." Because as you may know, insurance companies are stopping to insure houses that are in dangerous climate change regions, like in California, they don't. In the south of France, it's beginning as well. There are too many floods and the insurance companies just can't insure you because if they do that, then they're going to have to go into their legal reserves, which is forbidden under European law. European law since 2008 says you have to have that much.
If you have a machine learning system that says, "We have 400 people you want to insure, but we calculated that with fire risks and the weather risks that you shouldn't insure that person," it goes back to a human. There's not a single insurance broker I know that's going to say, "Okay, the machine said that. Goodbye." No, it's going to go back into a process where the person is going to analyze what's going on in that region physically, then they're going to talk to management, management's going to say, "Okay, in this area, that's it, we're not insuring houses anymore." It's going to go through humans. Machines will suggest, but insurance companies and banks are very careful, especially since 2009.
I'm not speaking at all about investments, I'm not speaking about the way they create stock. I'm just speaking about every day, I would say, linear stuff like billions that you transport, customer-- I'm not speaking about investments, because that's a very different field.
Alexandra: Makes sense. Earlier, you said that you find model-agnostic explanation the way to go as opposed to specific explanations of one certain type of model. There are statements out there in the explainable AI discussion, for example, from Cynthia Rudin, who published his paper in Nature, I wrote it down, "Stop explaining black box ML mods for high stakes decisions and use interpretable models instead." Her approach or her suggestion was more to not ever have any post hoc explanation, but build models that are interpretable by design, so that you don't have this fundamental dilemma of having to trust one AI to explain the other AI. What's your position on this?
Do you think we need more interpretable machine learning, or is, especially with the complexity we are seeing in certain models, AI the only way to [inaudible 00:47:19] [crosstalk]?
Denis: The problem here is I'm always very careful because I'm very respectful of everybody. I think that the person that wrote that paper did it in goodwill, and the person really believes that we could build a model that you can interpret. Why not? The research shouldn't stop, you can try. I wouldn't stop that research. Suppose I was the research director of this university and I have my opinion that I don't believe in this because of what I said earlier about 1 trillion parameter system with maybe 200 layers, and I would say, for me, it's inconceivable. For me.
Now, if I were the director of research, this person comes up with the paper and I said, "What if you're right? Maybe you're right. I don't believe you, I don't believe in it, but we're in a scientific area, so I'm going to give you a nice share of our budget, I'm going to respect your thought, and I'm going to let you work on it for two years. Here is, I don't know, $500,000. Is that enough for you?" The person says, "Yes, that's enough." "Okay, and I'm also going to rent a machine on Google, which is domain-specific, so you'll have a supercomputer to work on it, which is not too expensive." Google now rents domain-specific supercomputers at a very low price, it only costs maybe €15,000 per year, $17,000.
"I'm giving you this supercomputer, giving you your time, I'm giving you five PhD students, is that enough?" The person is going to say, "That's more than enough." "Okay, but if in two years you come up with no results, I won't continue the third year. I'm going to listen to your path, although I don't believe in it myself, but I'll give you equal resources to what everyone is doing." That's my way of thinking because I'm working on this trillion parameter model, which is based on sparse modeling. Now, sparse modeling is a very specific thing where you have an Excel sheet and a lot of zeros in it, and you just ignore many parameters and you go through it.
Maybe this person is right. People thought that one day airplanes won't exist, but then airplanes existed.
Alexandra: With all the lacking screws and they're still flying.
Denis: Yes, so I would say I would go for model agnostic, but I would respect anybody that writes that paper and works on it and hope for it.
Alexandra: Makes sense. Another question I had when I was looking into this topic of explainable AI was also, can you think of areas where we absolutely shouldn't have explainable AI?
Denis: That leads me to the question are there areas where we shouldn't have AI because we should have explainable AI everywhere there is AI. There are areas of AI that should not exist. Yes, of course.
Alexandra: I agree with that, but my question was more or less the direction of models being used in some security applications where explainability and too much transparency could introduce tweaking the system behavior by malicious actors.
Denis: You have a hacker mine.
Alexandra: I just did a lot of research.
Denis: Let me tell you this, in the 1980s, one of my passions was hacking. I was the ultimate hacker, but I would tell the company before. I was an ultimate hacker up to around 2008, 2009.
Alexandra: Let me briefly ask, was this the time period where you appeared on TV and in the media quite often?
Denis: Neither, and then speak about it publicly. It was something because maybe you know I'm a bit ironic sometimes. I remember this one example where I had a Tunisian engineer working with me and we went together to a very major aerospace site in France full of security. There was no explainable AI, full of security. The guard would take our laptop, open it, turn it on, control us, and everything, and then they put us in a meeting room. My Tunisian friend and I were always having fun hacking stuff anyway. We were saying this software, we can hack anything.
We were just sitting there talking, and then the manager overheard our conversation. He said, "What are you talking about?" I said, "We could hack your whole accounting system in five minutes." He says, "How can you do that?" I said, "Can't you see that there's an ethernet plug here on the wall? What are you doing in a meeting room with strangers with an internet plugin there?" He said, "You don't know the passwords." I said, "Are you kidding? You're using Cisco. You're using a 25 or 26 password thing, we can crack that in two minutes." He says, "If you do it." I said, "You sit here because we're going to show you it's possible, but we don't care.
It's just to show you that you got on our nerves with all your security here, but there is no security in fact." In five minutes, we cracked the system, we entered their accounting system, and we showed them everything that was in there, but I won't tell you how here officially.
Alexandra: We don't want to inspire somebody too.
Denis: What I'm saying is, a hacker doesn't need explainable AI. I remember another corporation where it was very serious security and they came to my offices and in my offices, they found nothing because in my company, you could not develop locally, you had to develop on a server. On that server, only two of us had passwords. Even developers didn't have the passwords. They were changing all the time and there were passwords on the door and no paper was allowed in the office. When there was paper, it had to be destroyed at the end of the day with a special machine and not put in public trash cans.
They came to came to see us, they say, "Wow, your security is higher than ours." I say, "Yes, I'll tell you what I think of your security when I go to your company." I went to this corporation and I said, "I bet I can find private information in one minute." He says, "That's not possible, but I won't let you touch a computer." "No, I just emptied the trash cans on his office." Took the trash cans, emptied. I said, "Look at this, here you have this, that, that, that, that." What I'm saying is, a hacker can enter anytime, anywhere, anything.
Alexandra: We don't have to be concerned of explainability when it comes to hackers.
Alexandra: Do you see any other risk that people could then exploit the system if they know too much about the inner workings?
Denis: Now you're speaking about industrial secrets.
Alexandra: Industrial secrets could also be, I don't know, the person applying for loans.
Denis: I'm model agnostic, so no one can know what's going on in the algorithm because I'm not giving it to anyone. I'm not building a system with interpretable AI in it. No one knows what's going on. It remains a black box, which means that anyone that has access to the input and to the output can find out what's going on. In my security system, which I
had in my company and you just block any access. For example, there was no Wi-Fi in my office. No Wi-Fi.
Alexandra: For explainable AI, I might want to know which features are taken into account when a certain decision is made.
Denis: Yes, but anyone can use explainable AI the way I'm saying. I just have to put an input into the system, take the output, store it in a text file one by one, even manually, and after a month, I can run explainable AI without the model and find out.
Alexandra: Sorry, I'm not following. If I'm, for example, the person affected by a decision of an algorithm whether I should get a loan or not, I might want to know factors were taken into account, which features were used to come to this decision. This potentially could introduce some change of behavior or change of supposed facts to come to different decision at another point in time. Or, for example, one of example I've read in one book was that the sales agent of a bank found out that the algorithm who had to give clearance for certain types of, I'm not sure if it was mortgages or loans or whatever, knew that the algorithm classified differently depending on, I think it was up until €500.
It was category one from €500 to €1,000. t was the next category. When they filled out the loan application or whatever application form with the clients, they rounded up everything so that it was classified in a way that the outcome was in favor of the person applying for the loan and was in favor of them getting their bonus at the end of the year which was some type of behavior that was only possible because you knew about this inner working.
Denis: Yes, so that's an excellent example. Let's say that we're in that office together, and you have decided when you say, "Denis, you know what? You have access to the input because you're writing the forms and you have access to the output."
I'm giving you 5,000 inputs and outputs, so you can't do it by hand. Your little hacker mind won't work. I don't mind, I'll create a table with the inputs, I'll create a table with the outputs, and then I'll run an artificial intelligence algorithm. I can even use K-means clustering. I don't even need deep learning machine learning and I will create clusters of the inputs and the outputs and how they fit together. I can even use a cosine similarity, a tiny little trigonometric equation and I'll find it. You just give me the inputs and the outputs. Everyone has access. He fills the forms, you get an answer. After a certain time, this reminds me of a documentary I was watching the other day. I fell asleep watching it.
Alexandra: It wasn't a good one?
Denis: It was excellent. Before I fell asleep, I saw crows and they wanted some seeds in a glass of water, but the seeds were at the bottom and there was just that much water and they couldn't get it, so the crows went and got some rocks and they put the rocks in the glass until the water came up and they ate all the seeds. What I'm saying is, a real hacker always finds a way, always. He'll hack the data or hack your Wi-Fi, hack your network. Someone who wants to hack hacks. There's nothing that will stop it. You'll create a countermeasure, then the countermeasure of the countermeasure, there's nothing.
Because, in fact, in many companies, you may know this, all you have to do is talk to the people. You just invite them for lunch, have them drink a couple of beers, they love to talk. People love to talk.
Alexandra: Now I see your hacker mind again.
Denis: "It's so fascinating what you do, how is that system? Don't tell me how the system works, but I think it works like that." People just love to talk. To make my conclusion, I don't see explainable AI as a threat. If it doesn't exist for a hacker, hacker won't find any way into the system. In fact, I even did ethical hacking at one point and I even updated my knowledge a few months ago.
Alexandra: Help me, what is ethical hacking?
Denis: Ethical hacking is you have a hacker in a big corporation and the person is hacking all day long-
Alexandra: To help them.
Denis: - trying to find the security. Yes. I can't say the company, but I know a defense aerospace company in France who did that and five hackers, and boy, the security now is so high. I can tell you that now it's impossible to hack that corporation. It's impossible. It is literally impossible because now they're monitoring every single computer in real-time on huge screens just like the police. Anything that is unusual, it doesn't fit another day, pops up with an alarm and you have security guards running down the hall to that place. It's pretty tight. Big corporations are very tight.
Alexandra: One thing I wanted to get back to, you mentioned earlier, I think you said you would just look in the input and output and the model itself would remain a black box. Did I understand you correctly? How does this then sufficiently satisfy the requirements and needs for explainability in AI where so many people are complaining about AI being this black box? Don't we have to get rid of this black box to make everyone happy?
Denis: You're going back to the article you were speaking about earlier, but the thing is, I think people are projecting emotions on artificial intelligence, but aren't they questioning before artificial intelligence? Before artificial intelligence, people were always complaining to the bank because a banker comes and he says, "I didn't refuse your loan, I just entered your data in the system and the system says you can't have the loan." People were already complaining, "Why does your system say that?" Because the banker would not tell the customer face to face that he didn't want him to have the loan, so the system did that.
They say, "The system said, it's over a third of your income to get that loan, so you can't get it. It won't go through. See? I try to type it, it won't go through." Then you can go back to ATM machines. Some people they go to the machine and they enter their credit card and they can't get the money so they begin to hit the machine because it won't give them money. We're just projecting things on artificial intelligence because it's just a tool. Believe me, if I go and I take you to, I won't say which corporation, where you build rockets, I can tell you their plays, there's no artificial intelligence.
I show you an algorithm of how you calculate the angle where you're going into space and you're coming back, you can't explain that to anybody. No one understands. You're talking to the manager and he says, "It has to be in that angle when it comes backward or it'll bounce off the atmosphere." I can explain why and the manager say, "Hey, don't talk to me, I don't know anything about physics." If you go to see a surgeon and the surgeon says, "I have this complex operation we have to do on you because I can't go into the details," and the person says, "I want to know. Explain, I want to be operated." "Okay, you want me to explain?" and he's going to be using words we don't understand.
Say, "Explain in simple words." You can't. Or you get a blood analysis from a specialist which can be 15 pages and you go to your general practitioner and you ask and he says, "I don't understand this. I'm limited to this part. You have to see the specialist." You go see the specialist, the specialist begins to talk to you, and you, after five minutes, say, "Hey, I didn't study all this. I don't even understand what she's talking about" or he is talking about. Every day we don't understand anything, but all of a sudden, we want to understand AI which is just like everything else. You have to take what's going in.
You take your politicians in Austria or in France, let's take your present chancellor. He said things before he was elected, that's the input. Now he's doing things after he's elected. Now you put them both together and say what's going on in this country. Can you explain what's going on in his mind? You don't know. Why would we try to explain something so complex when we can't explain anything, in fact? Suppose I invite you to dinner and I cook you a couscous. A couscous is an Arab dish and I learned how to cook this with my Tunisian friend when I went to their place in Tunisia.
Alexandra: Yes, you already told me last time, so definitely she heard our listeners speak.
Denis: You know the couscous. Now, I give you the couscous and you say it's good or bad or I like it, I don't like it, "Give me all the ingredients and tell me how you cooked it." Is that going to help you? It won't even help you to know how it worked. You say, "I just don't like it. Denis, I don't like your couscous." "Yes, but I'm going to explain all the ingredients and everything that's in it." "I don't care what's in there and I don't want to know, I don't like it. There is not enough salt." "Yes, but I put salt in there." "Next time put more." What you're doing is you're just taking the output, you're not even taking the input.
You say, "I don't like it and I know why I don't like it" because a lot of people don't need artificial intelligence to say why they don't like a message on social media or they don't like the result in the bank. They don't even need that, they're just angry and the lawyer is the one that's going to say, "You better explain why you did that." I'm still into my model agnostic thing and I believe that people are projecting emotions. It's just an automatic algorithm and I don't understand. Do you know what's in your smartphone?
Alexandra: Not exactly, but I'm just wondering because last time you told me the couscous example, I think you referenced it in the context of that explainability is not that complex because you just have to look into the input and the output,-
Denis: That's right.
Alexandra: - and the recipe list of the couscous and if you find a spice mix there that you don't like, then you could tweak the input in a way until you like the output.
Denis: Okay, You can write the next book.
Alexandra: Maybe we can co-write it.
Denis: You just gave the model an agnostic explanation.
Alexandra: Wonderful. Did I understand you correctly that you say, okay, even though explainability is important and we can achieve it in a model agnostic way and should only look into the input and the output, it's not necessary that we understand each of these models until the last detail because [crosstalk].
Denis: That's another subject.
Alexandra: That's another subject? Which one?
Denis: Because I think it is necessary for us, the ones that build these models to understand what's going on. I work a lot on explainable AI for the systems themselves, but it's like surgeon discussions. I use sparse modeling, I use linear sparse modeling. Ian Luka from Facebook came up with this sparse modeling to explain how it works, but he only limited to a few layers. Of course, I spend my time trying to figure out how the algorithm reached that conclusion. I spent a huge amount of time doing that, explainable AI for an expert and I say, "Ah."
I'm testing the model itself, I'm entering the code, I'm looking at the math and I'm analyzing the math and I'm analyzing all these parameters and I'm taking one pile out of there and I'm looking at them because I can take one layer out of these 500 billion parameters and I can look at maybe a million parameters, but it's very easy to do with a billion parameters because you can create vectors and you can see directions it's taking and you can see the mistakes. Then from there, I can build back and give them inputs and it gives me bad outputs and change the system.
I say yes, for a surgeon, it's very important for that surgeon to have a video and do the operation again, but for the patient, the patient knows they're sick and now he knows there was an operation. You come to a person's house, the person tells you the ingredients he puts in it and what you like and you don't like and you say, "What's that taste in it?" The person tell, "You don't need to know exactly what happened," but for experts, it's definitely a must. That's where your article comes back again. Where that person might end up two years after saying, "Maybe for mainstream, people know, but see how it can be useful for us?" and I say, "See? I was right to give you all that budget."
Alexandra: Yes, absolutely. It's also what I read during my research that explainability on the one hand has its role for the end-user who's affected by the algorithm and using the algorithm to build trust in it, but yes, we potentially need different explanations than when it comes to the people developing the algorithm. The first place here can also see this interpretability being of benefit. I also found some resources on correlating this interpretability also with improved model quality and robustness.
Denis: For development, I would say it's a prerequisite. You just can't go on building models not knowing what's going on inside of them, but it's so complex it takes like college math to really go into all this math. Sometimes I spend hours, days, but I find it. I'm a detective, so I find what's going on wrong, but it takes time.
Alexandra: I can imagine. Although I've heard some people say during my research, I can't remember, it was also a university professor giving a talk on explainable AI. That one of his critique points was the majority of the research is currently steering in the direction of explaining AI to experts who are building AI systems, and that we're not good enough yet with creating great explanations for end users. Can you agree to that or do you have the feeling that end-user explanations are sufficiently researched and [crosstalk].
Denis: By now, you're beginning to know me through different angles, I'm not going to agree or disagree. I'm saying that I think that LIME, Shapley, and other tools and rule-based tools can already help people, but I would agree with that person. I never disagree in science because you have to be careful what you're saying. The way I see it, personally, I would have enough. I've developed for 40 years and I've always produced explainable AI.
Alexandra: Yes, I would also put you more in the box of the experts.
Denis: I would listen to this person. I say, "You're right, there's a lot of room for improvement. I did it that way, but maybe I could do it better, so what are your ideas?" I'm not going to exclude people just out of ideology. I have no ideology in science. I worked a lot with Germans and what I like about Germans is pragmatism and efficiency. A German, if you come and see a German and you say, "This works better," "Okay, let's do it." That's the way I think. I don't know how you think in Austria because each country is very different, but I liked working with Germans because of that. It's a very special culture with pragmatism.
Alexandra: That's true.
Denis: I would listen to that person, I say, "If you give me something better, I'm taking it."
Alexandra: From your experience, what would be the best or the necessary explanation for end-users? How would this vary?
Denis: Let's take my perspective away and let's introduce the two perspectives that I'm looking at but not fully adopting, which is the first article you mentioned where the person says, "Why don't we build better models?" and then the second person says, "There's not enough end-user AI," and I would say, "That's maybe a blind angle for me." Since I have a lot of experience implementing successfully, it doesn't mean that there's not room for improvement. I would take both of those critiques and I would try to find new ideas. I would plug it.
In fact, I would say, "We have 100 layers, why don't we plug in something that takes the result out of 10 layers and see what happens there? Then after 20, 30, 40, and see how the system is evolving through there with force modeling." If we can do that, why can't we put it into an HTML page? Since we're so smart, why can't we just explain it to end-users better words than talking our jargon saying, "At this point, the word cake was interpreted as something to eat, but at level 20, it was looking at other contexts, and at level 30, it says, 'I understand that this job is a piece of cake, it's easy to do.'"?
I would agree strongly with people who disagree with me because they're the ones that will make me evolve, so I would say we need to follow that person that wrote that article and this person you just mentioned to improve. There's nothing perfect in this world.
Alexandra: That's true. One thing that just comes to my mind, which I took away from our first conversation, you also shared this example of I think it was the Chinese warehouse where you developed an algorithm. Initially, you presented them something that would, I think it was predicting how much or what the best solution would be and also how much money they would save, but of course, the question was, how did the algorithm come to this conclusion? Then you have the second version of the algorithm to solve this problem, which made more step by step approach and was amplifying the human, making it faster.
Can you maybe quickly share this example with our listeners? Because I think this was also super nice example of taking the user with you and not presenting them with the end result resulting in higher acceptance and trust into the system.
Denis: In fact, that's an interesting one because I did one thing that you just mentioned. At first, I had this complex optimizer that optimized, really, I would say, large amounts of money. Then the corporation, the general manager says, "I would like to understand." In fact, I'm contradicting myself because I wrote a specific algorithm for him because there was a lot of money involved for me too, so I decided I wouldn't be so lazy. I wrote an agent. I wrote a specific agent on a visual interface that took the algorithm and represented the way it was thinking in the warehouse.
You would see the little agent going there, then you see it stopped, then you can see it's taking constraints, then it will move there, and the guy said, "I love that as an educational tool, not only for your model, but also for all the people in the corporation, so I'm going to buy it and want to use it as an educational tool." That's for your explainable AI and it goes in the direction of the paper you were mentioning earlier and the person that says we need more, but it wasn't enough because when I reached the point of implementing in real-life warehouse, none of my calculations could have been good. That's when I told you that I shelved the whole project.
I said I can't work that way because no one knows exactly when the trucks are going to get there because there can be snow, it can slow down the trucks, you can have people that are sick, you can have an automatic machine that breaks down. I would rather just present decisions automatically, and then the person will click and approve or disapprove them in real-time, and on the top of the screen, what I told you, so I had a KPI that showed how much money this person was helping the corporation earn for each decision. At the end of a period, that person got a bonus on his paycheck.
In fact, it turned into a huge video game thing where it's spread out to several warehouses where everyone wanted to use the system.
Denis: Yes, use the AI up to a certain point, make a decision, earn money. They were talking to everybody, it just stimulated people. I believe in cobots, collaboratory but with humans. You have a human and you have a machine and they're both together. It's like driving a car, you can go faster than walking.
Alexandra: Absolutely, so really, these amplifying human capabilities. I assume the model was also much more useful because the human could put in all the unforeseeable changes like, for example, a snowstorm. Plus, it was more understandable and relatable because the human was presented with the decisions at each step of the journey and not just with one final.
Denis: Exactly. In fact, it goes back to what you were saying for this lady that wrote the article, why don't we build models that are interpretable from the start? This is one.
Alexandra: Yes, definitely. That's a nice example. I think if I recall correctly, you also mentioned that this reduced the amount of planning each day from, I think, eight hours from different people or something.
Denis: Yes, you bring them down to one hour, and the rest of the time, you're trying to figure out how to get that truck in there faster to deliver everything faster and get their bonus.
Alexandra: Sounded like a win-win for everybody involved.
Denis: Yes, that's the way I presented it to the managers and I said it's the only way to motivate people, otherwise, why would they be motivated?
Alexandra: Absolutely, that makes sense. Then it was really an insightful conversation and I really liked the directions in which we moved. My last question to you would actually be how do you see the future of artificial intelligence and also the role of data scientists? Because I've seen some of the materials you posted. Also, how the job of the data scientists will change and end-users being in a position to use all these new supermodels we're going to have.
Denis: Two perspectives. The first perspective is technical. Transformers are progressing at such a level that end-users will take over data scientists and AI specialists functions, many, many of them.
Alexandra: How do you define a transformer for those of our listeners?
Denis: A transformer, the best definition is Stanford University, which we were talking about. They don't call them transformers, they call them foundation models because one model can do 100 tests. That's really the thing. They can do all kinds of tests; predictions, text classification. We can say a transformer is one model that can do a hundred tests. We can compare it to a Swiss knife. We have the Swiss knife do many things. Before, we needed one model for everything. The transformer foundation can do many things. Since it can do many things, then the end-user can use it like he was using Excel before to replace some of the data sciences work so that these people have to move up.
As they move up, they have to have a wider view of the infrastructure. For once in this conversation, I will be less technical. I would really get them to worry about the supply chain because there's a real resource shortage on this planet with semiconductors, energy, the carbon dioxide. We have less percent in the atmosphere. In the supply chain, I would include climate change. In fact, in the corporation I was talking about in the last projects I did, we included carbon emissions into every single calculation to bring them down. I would bring these people away from their traditional tests into bigger infrastructures in which they would include optimizing climate change through optimizing supply chains.
If your truck drives less, you're earning a lot of money, but you're saving carbon emissions. If you can make one ship instead of three, well, then you're saving a lot of carbon. What I'm saying to users, it's a win-win. For corporations, that's where they need to go. They need to go into higher levels of optimizing supply chains with data, not just processing it, and let the end users do what they were doing before. That would be my vision.
Alexandra: That sounds good. I think that's also the way to go because neglecting climate change, I think, is not going to move us in the right direction. Since you mentioned resource shortage with non-data scientists being in a better position according to your prediction to use machine learning--
Denis: It's only my opinion.
Alexandra: It's your opinion, of course, but how do you see the role of data scientists evolving? Also, since we currently have this extreme shortage of data science talent, do you see this go away in the future with more people [crosstalk]?
Denis: No, they'll not stay there, but they need to evolve. It's just like when the pocket calculator came along in the '67, they began to use it, and worked better than other things. It's the same thing, they just have to adapt, but it will take time like going from horses to cars.
Alexandra: Can you quickly elaborate on how the job of the data scientists will evolve according to your assumptions and predictions?
Denis: My vision is he needs to go into infrastructure higher. He needs to have an infrastructure where end users can do what he's doing today in a very standardized way, prepared just like for ERPs or any system. The data scientist and artificial, they need to move up to architecture and infrastructure and see how to fit the whole corporation together, and then prove even parts that are not artificial intelligence to make the whole system faster, more productive, optimized to consume less resources, which is good for financial results, and good for climate change. It's a fact that we can't ignore.
Alexandra: Yes, definitely not. Wonderful. Any last remarks from your side, Denis?
Denis: I was happy to talk to you. [laughs]
Alexandra: Happy to hear that. Thank you so much, again, for taking the time and for everything you shared today. It was a real pleasure.
Denis: Okay, so see you soon.
Alexandra: See you soon.
What a wonderful conversation with Denis. I really hope you enjoyed it as much as I did. Thanks again for tuning in for our last Data Democratization episode this year. Of course, we will be back next year with new episodes about synthetic data, privacy, ethical AI, regulatory developments, and as always, data best practices for business leaders, privacy professionals, tech experts, and anyone in between. Make sure to tune back in in January. Until then, goodbye and enjoy the rest of the year.