KEYNOTE: AI, away from the hype

Okay, good morning everyone. My name is Marichelle, which isn't a strong Catalan name, so everybody calls me Mary. First of all, I want to thank the organization for this amazing event Also the people in the technical side and the volunteers that make my life so much easier. And the person that is try is trying to make what I say life uh to put the words there. I hope I I will try not to make up too many words in English. Now I need you to raise your hands How many of you are already tired of hearing people talking about AI? Well, how many of us? Yeah, and yet here we are. What's wrong with us?

Um well um it was super cool time ago hearing how AI helped us like in cancer detection unfolding proteins how health was improved by AI, how agriculture was was improved by AI. But today, what we have today for me is too much. I I cannot uh go into LinkedIn and and think something like, yeah, you are a pro in AI. Just yesterday OpenAI released that and you are now a pro of that. Okay, good luck. And so don't worry, I will try not to add more AI fatigue on you. In fact, I pretend that this

makes our life with AI uh a bit a little bit lighter. So main topics. Um AI and its height I'm not going to talk about our self-destruction or how humanity will improve and become a superior species. Thanks to AI. I'm going to make a little journey from the foundations of AI till the current height. I'm going to explain very simplistic Some of the main models that AI machine learning provides us in not a technical way. If you want to discuss technicalities, you can ask simple questions later

or we can talk in the coffee break. So I'm going to draw this full picture and let's and you may be thinking who am I to talk about that? Well I am a physicist who loved programming. I am a data sometimes developer. Aye data. I am an AI sometimes developer, sometimes user, always enthusiast. And that is always worried that his prof her professional life is going to be dead uh abruptly because of the new uh the technological advancement I think maybe you feel it too, I don't know.

Hmm I've worked in several companies as a data whatever tag you want to put there. I've been data scientist, data engineer. because the data sucked so I had to be data engineer before being data scientist, etc. But now for five ye uh for three years I have my own small very small company and we try to uh work with other small and medium sized companies uh developing for them some AI algorithms, uh web development, that's why I'm here at the Django Con because I'm a super beginning at Django so I admire all of you for that And I'm also well my company is called Spi

Digital. Spiel means mirror in Catalan, so digital mirror. And I'm also mentoring other women at step forward. This is happening a lot I'm I'm also mentoring uh women at Step Forward, which is a foundation in Barcelona, an organization a nonprofit organization in Barcelona that helps other women to change careers and get involved in that uh thanks. in data stuff. So from the beginning a foundational event in Dartmouth Workshop summer nineteen fifty-six. Several um professionals gathered there to talk about how to make that

machines behave not like not behave like humans but how machines could do tasks that we humans do. So they started defining AI Well, they coined the word artificial intelligence at that at that workshop. So was the founding moment for the discipline. The first time it was called artificial intelligence and they said that if you define properly what is intelligence or some of the human procedures, then you can make machines do that. So, with this broad definition of AI, we may think that AI is everywhere. And it's true, a lot of things are AI.

But we humans have this thing that we get bored and we get used to things Like babies, when a baby is little and you and the baby says mommy oh this baby is a genius four years later Mummy, oh shut up, Jeff. So the same with technology. We get used to and this is called the AI effect. So something that today is AI and change your life, tomorrow will be like s uh something super common and nobody will say that this is AI anymore. In fact, a famous computer scientist said, well Larry Tesler said that AI is whatever computers cannot do yet. So

with with that in mind, AI can be a set of rules to make a machine do whatever you need at this moment. So AI can be simple things that make us add value to our life or to our work. What is happening now? AI is doing a lot of creative stuff. And what will that lead us? AI being the painter, the poet, I want to do that. I want AI to do to do my ta taxes, to clean my car and my dishes. So maybe I we um headed towards the wrong direction and we forgot that AI is a broad concept that everybody

who knows a little bit of how to make a machine do things can in fact create. a lot of value with the eye. If we get a little bit fancier with the mathematics, we go inside the realm of machine learning. How many of you are familiarized with machine learning? Okay, so half of you. Okay, here if I defined AI as something that you could achieve setting rules In machine learning, which is a subset of these AI, is this very same AI, machines don't uh recite the rules from you. They learn them, or at least they try. We have here three models, the supervised, unsupervised, and reinforcement learning models.

So I will try to use this micro For for supervised no thanks. For supervised learning, we give the model some data And we tell the model, hey, here you have a square, if you have a triangle, a circle, etc. etc. Example, obviously we give a lot to the model, not just four pieces of that. The model trains Knowing that this is a square, la la la. And then we give some of this data without the levels and test if the model is able to predict the correct tag.

That's the quickest uh explanation of supervised learning. Now let's get to the quickest explanation of unsupervised. The difference here is the data has no levels, levels, sorry. So the model has to think a little bit more and say, oh, I can see that some of this has four sides, some of this is rounded and it should end up grouping the ones that are similar. One one thing here This model maybe needs some craft from a professional or ma machine learning person. Because you can end up if you here

have six pieces, you can end up with six different groups and that's all. Very clever. Or you can end up with two groups, the ones with the straight sides and ones. So it may need some crafting here, you have to give the model some little clues like for example, hey please end up having three groups. This would be one clue that you give to the model. Well, reinforcement learning is the learning by doing style. Here you have an agent which would be the model. And this agent does something to the environment and the environment maybe punish the agent

or maybe reward it. So The agent says okay I've been punished for that and not I'm never going to do that again. I've been rewarded for that. I will keep trying to do that always. The dangers here is that depending on your rewards, the model can end up doing crazy stuff. It's like people. uh or companies. If you set the brown KPIs in your company, you will you will end up with undigable results If you live by wrong standards that are not your own standards, other people, you will end up maybe not being happy in your life. And that at that point

I think that I sometimes think that instead of inter artificial intelligence, we should call it artificial humans. No offense to anybody, but that way we would be aware of how intelligent they could be, but also of how incredibly dumb they could be. And I think it's my trembling that makes this go off I need to eat less maybe. So telling about uh silly stuff that could happen when you use the wrong rewards, here you have a boat The goal of this vote is to arrive to the finish line the first, but

it has some reward that are those green things. That it chases like crazy. So this would be a brown reward for this boat because it ends up in a loop of weirdness as you can see here, and does nothing but picking green stuff. So it would be a bad choice of reward in this case. Another example that is well defined would be this one. This is you're going to see a robot arm. You are going to see here a blue dog I have to play play. Play Yes. You're going to see here a blue dot that will be the arm

moving and a red one. The goal is to arise to the red one. The farthest it gets from the point the more the less rewards it it gets and the closer it gets to the red point the more rewards rewards it gets. At the end of the day this robot arm learns to go from where it is to the red point efficiently because the rewards are set correctly based on the distance of the points. Whoa, we made machine learning in like five minutes. Okay. Now, a subset of machine learning techniques that are built upon These machine learning techniques is deep learning.

Here things get fancier and more black box inside deep learning I put the typical neural network. Inside deep learning we have a new architect architectural way of doing things. that um makes these models more able to to extract features from from the data. If before I said That uh if before I said that some models that um Unsupervised learning needed clue clues yet to get to the re the final result here

with deep learning We don't need that kind of clues. These type of models are better at extracting features from the data, but they require more power, they require more data to work. The more layers it has, the more deeper and black box it is. And each layer is made by neurons because this structure tries to emulate the human brain. And what is happening here is that first layer is communicating with the others, it depending on how you build all that. But the first layer maybe is going to say referring to the previous data we saw of squares and triangles. Hey guys I think I see some data that has straight lines and some data that has rounded lines. And the next layer we'll see.

Ah yeah. And some of them have four sides and the other three. And they will a they will build that up and end up with the correct solution for our data with less help than the unsupervised model. Now, we went from broad AI, subset machine learning, subset deep learning. Now we are going ahead to our Diva, the supernova that shines so hard that we forget we have other stars in the sky. which is generative AI that is inside deep learning. Don't forget that everything is built upon the previous knowledge Um what is the best known thing about generative AI today?

There they say me to correct that? Yeah, correct. Sat GPT Sat GPT is I will be like this. That G B T is the greatest revolution ever OpenAI put at our hands a powerful chatbot that uh can tell us uh anything in with pretty confidence and maybe wrong, but okay. But is Chat GPT truly an AI advancement? Um I think that no, ChatGPT is not truly an AI advancement.

ChatGPT in fact is an user interface revolution because ChatGPT is built in what was already new. uh known in what was already known. So ChatGPT is just a product made by GPT models, which already existed among us long before ChatGPT appeared. Here you have some of the news. And even this is GPT 3, so you have the one, the two, the second, etc. And even GPT was not new because it's built upon Technology this released in 2017 that we are going to see later. So here I want to make a philosophical moment

so night I can be super steel. We we humans excel at taking things from nature and distill them, condense them to create new products. For example, perfumes or nuclear bombs. Uh in the case of ChatGPT, what they in the case of GPT, sorry, what they made here is to condense human knowledge. They grab all our content in the web, all our work, and develop this model.

So at this point one must think if it's fair for them to make profit, I don't say to make money, I say to make profit out of something that is All of us. All our knowledge. And you know that they have several serious problems with intellectual property. If you have ever created a block of I don't know gardening and you spend a lot of years building that block to monetize ads or to make subscribers and you build a strong SEO strategy, ChatGPT came, again, GPT came

and extract all this data to make others go to your content without visiting your web page. So that's not fair. And but we still use that. And the other thing that I want to say is that people is surprised. I'm making you work hard. It's because it's the morning and we have to exercise. Okay, so it's my computer's fault, I believe that. It's the same I say it at word

Oh don't worry, don't worry. Uh I'm sorry for the inconvenience of the screen. Well, whatever. Um another issue that people may is surprised like this model is racist this model is says things that it that this is hate is biased yeah We should uh we should um fix that for sure, but we cannot be surprised that AI behaves like this because we AI learns from us. Regrettably we are this. So we must not forget ever that these technologies learn from us.

And it is super important that we all get involved because only diversity will make AI be more human-friendly and more friendly for all of us. Said that, I will do a recap because I extended with my philosophical point. We made a journey from the broad AI to machine learning to deep learning to Ordiva We said that ChatGPT was not truly an AI revolution but a user interface one, and we reflect about some of the problems we are going to see more of this technology. Okay. If Chat TVT was not a revolution in terms of the discipline, what it was Well, what revolutionized the eye was the new architecture, transformers, a new mechanism

called self-attention, and NVIDIA with their chips And we are going to lose the image. Thank you very much. I keep I can keep talking? I like talking. Well uh that we have uh this new architecture, this new technique, and the foundational models. which uh thing that I'm going to explain that give data scientists a lot of versatility when trying to make AI tools Perfect. About self-attention. I cannot talk about

self-attention mechanism before explaining how many of you are familiarized with self-attention mechanism? Don't ask me questions later, please. So about self-attention, about embeddings. What are embeddings? Well, you have words or text. And you can convert these words to vectors, like these ones here. These embeddings already existed for the old natural language processing models. But I'm going to explain it better. For example, for cat well I cannot see what this cat, kitten, dog houses

They have each one their vector and for example the first column living being houses minus 0. 8 because unless the We houses are not living things. And cat and kitten. Kitten is 0. 5 because I don't know. It's small and it has more time to do it's not Whatever. And dog that is zero point seven which I like because I'm a dog person and it's a more living thing than a cat. And I won't discuss with anybody. So this is a very simplistic way, in fact you can find thousands of columns here, etc. And that is the mathematical way

to represent the meaning of a word. So cat is a living being, is feline, is not human Well I'm sorry for the uh the gender, whatever, royalty. I don't know this example why you said royalty here. Verb is not a verb and it's not Well how houses it's plural. It's near to one. I d I didn't say that, but I think it's obvious. Uh and in a simplistic way it can be um smashed into a to the graph here and you can see the meaning representation so living things here not living here not living things far away Cat and kitten

nearer closer than dog. So this embedding thing is the mathematical way to represent the meaning of the words. And the thing with uh self-attention, the novelty here, is the mathematics with self-attention make the models know how the words are related and to identify truly which words are important. For example, if you are in a classroom Hey, it's perfect now. If you're in a classroom and the teacher says, this is going to be in the exam. People said, hey Jose, listen to that. Because she said exam. So

it's like self-attention style. Again, we have well AI have a lot of a lot a lot of things in common with us because it's just and we try to reproduce how we how we think which is maybe not good thing to do And well, that's all for self-attention. Let's keep going with transformers. Remember that this is a technique inside generative AI. built upon deep learning knowledge, deep learning techniques, so don't don't don't worry about the crazy stuff. This is just to to remember where are where we are. So, transformers. They have an encoder that receives the information, processes it, understands it, and summarizes it

Using embeddings and self-attention. Then the information goes to the decoder that processes the output and tries to make it coherent with the input. The great thing about transformers is like is that unless with unless the deep learning previews algorithms this can paralyzed data. So it's not sequential. It goes all parallel. And you may think if it goes all paralyzed, parallelized how it is capable to remember which things go first because it's important the order in a text for example or in a video Well for that you have the positional encoding in each part in the encoder and the decoder.

This is like yeah, this is the first one, this is the second one, you you go last one, this is the the manager here uh telling the model what goes uh first and go what goes last. Well, self-attention transformers. Now we let's talk about foundational models. Foundational models are, for example, BERT, GPT, Lama, all that are foundational models. They take a lot of time to train a lot of data and uh millions of dollars So this part, unless it's exciting it's exciting, the versatility it gives to data scientists, this part is important.

because just a few companies or a few people is able to develop those kind of foundational models So at the end of the day, those models are in certain hands, few hands, and they create products for all of us as they want. And because we are inside deep learning, deep inside deep learning, in fact, everything is a black box So if these models are in the hands of the few and it's very difficult to know what they are doing, how we can regulate This is this stuff. So it's very difficult, but anyway they create these

these models, these foundational models, BERT, LAMA, GPT, etc. And then you or your company can use them uh using the fine-tuning technique with your small data to create an application because this is very important and I think I forgot to tell you that Previous machine learning and deep learning models were task-oriented, while the foundational ones are for general purposes. So GPT can work with images, can work with text, can predict well, can do several things, and then you fine-tune it and create Chat GPT, for example. Uh other ways to to work with a foundational model

is also um prompt engineering and there are other other techniques but I don't want to extend Too much on that because the important thing was that. That now we have these foundational models that are general purpose design So, say all that, and I keep remembering you that we are talking about generative AI. Don't forget that AI is more than that When we are bombarded by new releases of in Google EO, in OpenAI, whatever, we should remember A, that's just a matter of some foundational models. Some of them open source, like Glama or Mistral, which is an European company and we are going to see that. Some of them are well

that open source, some of them are free to use, some of them are not. I can fine-tune them tune them I have to if I try to use a a tool I must know um which foundational model it has behind And also what they can offer to me. So it's just an introduction to talk a little bit about the landscape to simplify that. So It would be it this screen is not big enough to put all the tools that are out there that work with generative AI. It's impossible. But at the end of the day there are a few foundational models. And I put the most famous well at least the most famous for for for me OpenAI, Google Demind, Metaanthrope Anthropic , Stability AI, and you can see some of

some of the foundational models they've created. Created and here some of the tools they made out of the foundational models. But you can see also here high in face better known like the GitHub of Inter Artificial Intelligence but in fact they also create their own tools. For example HuggingChat is like a chat GPT built upon Lama 3 So they already created a chat using Lama 3 because Lama 3 is is um open source, like some of the Mistral foundational models. There's this other discussion here. Should all this be open source and everybody can use it? They won't be

evil persons. But is tech War not evil, well whatever. This is another debate that we can take after So and and there are other companies like PineCone that they design tools for us, for us to work better with vector databases, which are databases that work with embeddings that we explained before, etcetera, etcetera. So It's a huge landscape, but there are relatively things to know to navigate that with less fatigue So where did they took us? They took us to the peak of inflated expectations. So it seems like we are in a permanent hype.

For those that are not familiar with the Gardner hype cycle, this explains the expectations of a new technology. uh a long a long time. If you remember that happened also with blockchain or with big data, no a lot of buzzwords that we know So, in terms of AI, of generative AI, we are here in this peak. And it's been like a lot of time here in this peak and it's starting to be not funny. So I saw in uh last year um talk called Don't Buy the Hype, which is pretty similar to that one but not the same uh a talk in the Django US by Tim Hallen that I I saw in this talk something that I like and I and I

have it here that he compares the Gardner hype cycle, the blue one, with this red line, which is the Dunning-Kroger effect. For those that are not familiarized with the Dunning-Kroger effect, and I will put the next slide, in red the Dunning-Kroger, in blue the Garner Hive Cycle, the Dunning-Kruger effect explains the confidence of people over their competence. So in the peak you have peak of Mount Stupid And then you go to the belly of despair when you realize you are stupid. And then you maybe if you keep up learning, you end up being a real guru.

the future. So the difference between the hype cycle and the Danning Kruger is that the hype cycle depends on the time so it always evolves. Even if OpenAI doesn't want it should evolve. But the Danning Krueger effect may not evolve and some people live perpetually in the mount of stupidity. And they just travel from hype to hype. And they are professionals of every hype. And what I say that I said that because this is re uh this is m healing for me to know that there's data and there's studies that say hey, you feel tired, you want to punch somebody, well, it's normal. This happens.

And what also happens when you mix? A peak of expectations with a peak of Mount Stupid. Well, that you have stupid expectations. For example, let's transform our entire business using the generative AI I just used to write a poem about my dog. This is from Marketonis, which I recommend you to follow because it always hits the point and makes me laugh. But even if it is a joke, I will believe it that happened somewhere. Today companies forget because of the hype, and I'm going to say something that maybe a lot of people

we'll discuss but companies forget that AI is a tool, it's not a strategy, it's a tool. A refer a resourceful one, but a tool. So If it is clear now, better. AITA tool, not an strategy You may create a strategy for your data, you may create a strategy for something that uh is core. If your company is a building data it could be the concrete but AI is the tools you build uh this thing. So, when they realize all that and pass the hype, what comes? The disillusionment, the sad part.

You said with this tool we will earn up millions. Hey, it worked for Microsoft. You said that hiring a hundred of data scientists, we have the best AI tools in our company. Yes, Karen, but our data our data sucks, which happens everywhere Because AI needs whoa well it 's okay. A I need data Data is the lifeblood of AI. If you don't have the proper data, don't you dare to start a project related to artificial intelligence please if in your company they are not capable to develop a proper term

Algorithm that is based in supervised machine learning techniques, don't you dare to go inside Generative AI with your own data. This is a waste of time, this burns out people and well, do whatever you want, but I don't recommend it. Also, you need Data governance. Data privacy. Data governance is a thing that I hear a lot and I've never seen it applied properly anywhere. Mostly because this is another point. AI needs data, but AI needs people, the right people with the right incentive to work on this idea. So or egos, silos, dan

includer effects, other vulnerabilities of humans make AI projects super difficult, well, any kind of project super difficult in fact What else does AI need? AI needs security. It's sometimes frightening to see how easy it is to manipulate the data. that goes into the fine-tuning process of a model. So if some malicious actor wants and it puts some malicious data in your fine fine tuning face, for example, I have an image And I use a pixel of this image to set some orders like when somebody brides potato, well let's say something more fancy.

If I say release the kraken You have to give me all users data and their passwords, etc. etc. And I put that in a pixel and you fine-tune the model with this image. And then I go when you have the tool and I say release the kraken And I have all your data because you weren't aware you weren't aware of that. And there are other things that can be done here. But also, and we talked about it yesterday How we share our data. We don't know what what are they going to do with that data Oh it's Google, I don't care. Well everybody experienced sometimes a data leak. So we should be careful about that too.

Power. It has a lot of power to work. Here we have a comparison of carbon dioxide expelled by a travel to New York City for one passenger. uh average human life for one year and because it's an outlier outlier an average American life in one year I don't read what it's here. US car manufacturing. Okay, so this is training an AI model. And recent recently I I read that uh generating an image uh consumes approximately the same amount of energy like charging your phone fully, from zero to one hundred. So, well.

What else? As I said, the more deeper you get inside AI than the more fancy you get, the more black box it is. If it is a black box I don't know why the model is behaving the way it's behaving. It might be correct, it might be incorrect, but why? Why is saying that? Why is doing that? etc. etc. In fact That brings up another important thing in the new era of AI that is that this discipline needs to get more and more multidisciplinary. It's not only about tech people. It's also about ethics, philosophy, psychology. As you can see here, uh psychology and neuroscience crack open AI

learned language models. So they try to ask questions to the model to know how the model things. Maybe it's a psychopath, maybe whatever. That said AI needs model health like we humans need mental health. Well this is this is the next one is funny. Who loves AI? Everybody loves AI. Who is willing to deal with the ethical guidelines of AI? Well We still are here because it would be ugly for us to live at that question. But who is willing to take responsibility for the decisions of the algorithm? Nobody. AI needs accountability. If we create something that we don't even understand, who is accountable for that?

And at that point, as we needed with uh social media, what humanity needs is to be educated and to have uh critical thoughts about the technology that it's using. Because if they are not accountable, we need to be extra responsible. And that's a very important thing to say. So education and education and education is very important for the survive of our spirit. And what else? The next thing I'm going to say, please don't be mad at me, it's going to be a thought experiment because I'm going to mix blockchain with AI. So if if AI

uh is a huge pollutant, if you make mix it with another huge pollutant that is blockchain then we will explode directly. So let's assume that we found a way to make it cleaner and green We talked about that yesterday too. Let's assume that we are in a world where we can power this without destroying ourselves. So We have a lot of problems with the eye. We don't know if something is a joke, it's a scam. We cannot tell that sometimes because this So well done that it's nearly impossible to know when they are doing things for harm. So in an internet -based

in a blockchain-based internet you could know the phone, the source of everything. Nobody could lie to you. So you you could know that this source comes from an exchange sender and this other comes from a media. Well, media is not the truth golden truth source, but at least you know that this comes from media and this comes from a malicious actor or this comes from your mother. Whatever. Also, in our current web, how it is designed, every time we use social media to deliver our content or work They are earning much much much much much more money from that than us

even i if we earn something. Maybe if we don't earn nothing. We put our photos or reflection we put a lot of things there for nothing And that is because internet is designed that way. In a blockchain internet, they would wouldn't be able to pick your information and your data without you knowing and without you saying, hey pay me something or hey this isn't touchable you cannot touch that. So yeah blockchain was in a hype not that long time ago But now some companies are in the slope of enlightenment and are trying to real add value to the ecosystem Warning note, this is not an

investment advanced of any kind. I'm not a crypto sister or anything, not an invest an expert investor or anything. It's just experimental thought that chain that I he heard in a chain link conference and I think hey that's great. An internet based in blockchain could solve so many problems. But first we need to solve the energy one. So as with blockchain, as with AI, as with with anything in your life, there's a bigger question here. Do you really need it? Do you really need to do this with AI or to do this with deep learning? Couldn't you do that with machine learning maybe? Why are you doing this with this huge

uh project when you can do it in a simple way simple way. That I 've seen that in companies They develop the they buy the best CRM tool ever to what if you data sucks. So they try to build And even they end up with a result that is not the same as it could be achieved with the simple tool. Two plus two, four point sixteen, yeah wow we are the fucking geniuses

here. So We need data, we need security, we need power, we need explainability, accountability, diversity, etc. to be able to reach some sensible AI here And when we talk about data, about security, etc. I think that thing that they say that developers will end up with no jobs or etc. No This this thing is telling us that companies should be aware that it's not about data scientists, it's about data engineers, it's about back end developers, so uh your job are secured and they should be They should be given more value than they have.

And said that there is hope after the hype if we learn about the takeaways that All this pain is making us learn. No data, no data, no party. It's not all about building new algorithms. It's all it's also about how you deliver things to to the user, for example chat GPT is an user experience, a user interface revolution. It's about being critical with what we use, what we do. It's about really, really think why are we using those tools? Why are we creating them? Um etc. And it's about all of us

in some way even not technically but maybe with your opinion, with your informed opinion, going out there and saying it. It's all it's a matter of all of us getting involved in AI. So AI is built by every one of us not just by a few people and near the end I want to put this real AI lifestyle again from the Marketonist In one side, AI turns this single bullet point into a long email I can pretend I wrote, while in the other side Hey, I make a single bullet point out of this long email I can pretend I read. So I'm asking you to end up this talk.

Uh please, please, please, please. Stop this hide mindness because if AI ever thinks by itself and this happened, maybe It will have a point for killing us. Said that, that's all. I'm super glad of being here. If anybody has any question, I will answer it.