<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Sandro Mikautadze on Medium]]></title>
        <description><![CDATA[Stories by Sandro Mikautadze on Medium]]></description>
        <link>https://medium.com/@sandromikautadze?source=rss-2630bbcfbc44------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*FC_koRNan1E1S3Y3s34Otg.jpeg</url>
            <title>Stories by Sandro Mikautadze on Medium</title>
            <link>https://medium.com/@sandromikautadze?source=rss-2630bbcfbc44------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 21:25:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@sandromikautadze/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Is Bocconi’s AI Degree (BAI) Worth It?]]></title>
            <link>https://medium.com/@sandromikautadze/is-bocconis-ai-degree-bai-worth-it-03809bc68ef8?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/03809bc68ef8</guid>
            <category><![CDATA[university]]></category>
            <category><![CDATA[bocconi]]></category>
            <category><![CDATA[bai]]></category>
            <category><![CDATA[review]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Sun, 14 Sep 2025 19:05:25 GMT</pubDate>
            <atom:updated>2025-09-14T19:32:06.336Z</atom:updated>
            <content:encoded><![CDATA[<h4>The Largest Public Student Review of the BSc in Mathematical and Computing Sciences for Artificial Intelligence (2025)</h4><figure><img alt="Bocconi University’s campus in Milan, featuring modern, curvilinear architecture illuminated at dusk." src="https://cdn-images-1.medium.com/max/1024/0*J5CNJ5VWxuvIOyRk.png" /><figcaption>A section of Bocconi’s campus</figcaption></figure><p>In 2020, Bocconi University launched a Bachelor&#39;s program about AI — <strong>Mathematical and Computing Sciences for Artificial Intelligence</strong> — internally known as <strong>BAI</strong>. If you have heard about it and are looking for programs to apply to, you may be wondering:</p><blockquote>Is it any good?</blockquote><blockquote>Is it worth choosing this path over more traditional computer science degrees at technical universities?</blockquote><p>My plan for this article is to <strong>provide a general review of the program</strong> for those considering joining. <strong>I have all the information to do it in detail </strong>because I enrolled in the second-ever cohort of BAI in 2021<strong> </strong>and graduated in 2024. To make the information as unbiased as possible, I also surveyed <strong>more than 100 current and former BAI students</strong> who have voluntarily shared their thoughts in the interest of truth. <strong>I encourage people to comment and give more opinions down below if necessary.</strong></p><p>Note, I don’t aim to give a review of Bocconi or about the university experience, but just about BAI. As of today, and to the best of my knowledge, <strong>this is the largest publicly available BAI review made by students. </strong>And it’s much better and complete than random comments on Reddit.</p><blockquote><strong><em>DISCLAIMER: </em></strong>This is an <strong>independent review</strong>. It is not sponsored by Bocconi and doesn’t necessarily reflect its official views and opinions.</blockquote><h4>A Note on the Survey</h4><p>The survey was conducted in June 2025, at the end of the academic year. 107 people responded. Thank you to <em>Aaqib, Agnese Adorante, Alessandro Pranzo, Andrea Fabbricatore, Batuhan, Beatrice Citterio, Cesare Bergossi, Edoardo Annoni, George Morris, Hristo Kukov, Irene Colombo, Jean Duboscq, Luca Martucci, Marián Gloser, Marko, Mattia Barbiere, Mattia Forzo MV, Riccardo Scibetta, Samuele Mancini, Sofia Lambro, Tommaso Ferracina, Tommaso Giacomello, Valentina Zampa</em>, and all other anonymous survey respondents. Out of the 107 participants: <strong>13.1%</strong> were from BAI ’23 (the first-ever cohort, already graduated at the time of the survey), <strong>32.7%</strong> from BAI ’24 (my cohort, also graduated at the time of the survey), <strong>18.7%</strong> from BAI ’25 (third-year students, one month away from graduation at the time), <strong>22.4%</strong> from BAI ’26 (second-year students), and <strong>13.1%</strong> from BAI ’27 (first-year students). The data is <strong>not </strong>public.</p><p><strong><em>· </em></strong><a href="#5efc"><strong><em>Why Does Bocconi Even Have a Degree in AI?</em></strong></a><strong><em><br>· </em></strong><a href="#7199"><strong><em>The Name and The Philosophy</em></strong></a><strong><em><br>· </em></strong><a href="#e594"><strong><em>Curriculum and Workload </em></strong></a><strong><em><br>· </em></strong><a href="#d2c4"><strong><em>Faculty &amp; Teaching Quality</em></strong></a><strong><em><br>· </em></strong><a href="#eb44"><strong><em>Class &amp; Student Life</em></strong></a><strong><em><br>· </em></strong><a href="#6c65"><strong><em>Career Outcomes &amp; Opportunities</em></strong></a><strong><em><br>· </em></strong><a href="#525d"><strong><em>Extra Personal Comments</em></strong></a><strong><em><br> </em></strong><em>∘ </em><a href="#0ead"><em>Random Things</em></a><em><br> ∘ </em><a href="#9521"><em>BAI vs Engineering</em></a><em><br> ∘ </em><a href="#1a3d"><em>The Elephant In The Room</em></a><strong><em><br>· </em></strong><a href="#8a50"><strong><em>So???</em></strong></a></p><h3>Why Does Bocconi Even Have a Degree in AI?</h3><p>First of all, <strong>Bocconi is not a school that decided to do AI because it was trendy; no.</strong> It identified an educational gap more than a decade ago and was one of the <strong>first universities in the world to introduce a BSc in AI</strong>, with a clear long-term roadmap ahead.</p><p><em>If you don’t believe me, go find anything close to BAI dating </em><strong><em>back to 2020</em></strong><em> — I’ll wait.</em></p><p>The short story goes as follows: in <strong>2016</strong>, Bocconi took its first steps into <em>computing </em>by launching the BSc in Economics, Management, and Computer Science (BEMACS) and establishing the Bocconi Institute for Data Science and Analytics (BIDSA). A year later, in <strong>2017</strong>, it introduced the MSc in Data Science and Business Analytics. This led to an influx of several physicists, mathematicians, and computer scientists who laid the groundwork for more ambitious projects. In late <strong>2019</strong>, Bocconi announced the Bachelor in Mathematical and Computing Sciences for Artificial Intelligence (BAI), with the first class starting in the <strong>2020–2021</strong> academic year under the direction of physicist Riccardo Zecchina. <em>This is the program under this article’s scrutiny.</em></p><p>For completeness, things didn’t stop there: in <strong>2020</strong>, Bocconi launched a new PhD in Statistics and Computer Science; in<strong> 2023</strong>, it established a full-fledged Department of Computing Sciences; in<strong> 2024</strong>, it introduced a new MSc in Artificial Intelligence, as a natural continuation of the BAI program; in <strong>2025</strong>, it created a new MSc in Data Analytics and Artificial Intelligence in Health Sciences, in collaboration with Humanitas, a top research hospital.</p><p>Do I have to go on to convince it was not just because of trends?!</p><h3>The Name and The Philosophy</h3><p>Let me give you some preliminary information about BAI: it is a <strong>degree <em>for</em> AI, not <em>in</em> AI</strong>. It’s right there in the official name, and that’s not just a semantic nuance.</p><p>From a bureaucratic standpoint, BAI is classified as a <strong>math degree</strong>, but in practice, the program focuses on building a <strong>strong theoretical and computational foundation </strong>of math, computer science, and scientific modeling with physics, economics, finance, and neuroscience. These are all subjects useful <em>for </em>AI, which is the long-term application horizon of the degree. In a sense, the philosophy behind the program is</p><blockquote>Teach what doesn’t change.</blockquote><p>Don’t get me wrong, you will have AI exams, but the <strong>real goal </strong>is to teach you everything you need to <em>understand AI easily</em>, which includes math and computing sciences (<em>the name is now explained</em>). The nice thing is that those subjects are also useful in other disciplines. So if you end up liking other fields, you’ll have all the tools to go there, too.</p><p><strong>Why am I saying this now?</strong> Because this motivates the degree curriculum, and because currently enrolled and graduated students didn’t have this very clear in their minds either.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6egT_rvmSGr0ItERi41XLA.png" /></figure><p>I think it also convinces some people that the name is not <em>just marketing</em>. Sure, it has an effect, but in reality, it&#39;s just the most representative name of what you’ll be studying. So let’s go see what that is.</p><h3>Curriculum and Workload</h3><p>As I said before, the curriculum is mostly math and computer science. If we analyze year by year:</p><ul><li><strong>Year 1</strong> focuses on the <strong>foundations</strong>: mathematics (analysis, linear algebra, probability), computer science (introduction to programming, theory, algorithms), and the basics of economics and physics. <strong>This year is intentionally heavy in quantitative foundations.</strong></li><li><strong>Year 2</strong> builds on this base with more <strong>advanced </strong>mathematics and statistics (statistics, decision theory, advanced analysis), deeper computer science and programming (advanced programming, optimization algorithms), and dedicated machine learning courses.</li><li><strong>Year 3</strong> shifts toward <strong>specialization and applications</strong>. You’ll encounter advanced scientific courses such as quantum physics, game theory,<em> </em>and domain-specific modelling for finance and neuroscience. Of course, you’ll do AI courses, focused on deep learning. You also get the chance to do an exchange semester.</li></ul><blockquote><strong>PS</strong>: for the most up-to-date information visit <a href="https://www.unibocconi.it/en/programs/bachelor-science/mathematical-and-computing-sciences-artificial-intelligence">Bocconi’s BAI website</a>.</blockquote><p>In a sense, the first year is quite common to standard STEM programs, and then the second and third years go deeper into the AI-related aspects.</p><p>One important thing: it’s <strong>very heavy on math and rigorous. </strong>This means that, especially at the beginning, things will be very theoretical: lots of proofs, proofs, and sometimes proofs again.<strong> I am warning you because if you like AI but hate math, this program is not for you.</strong> The only way in which you’d fit is if you’re willing to put in the work.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fD-G8_JBhWUxIUY32kAp5g.png" /></figure><p>This means that <strong>you have to study</strong>. It’s not a degree in the science of snacks; it’s a degree in math. That’s why the workload is generally considered high (<em>don’t worry, Bocconi does support students in case you really, really go crazy</em>).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8ddN-_6Zo1vAVyLWbNRsDg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*oqbofFlR1OEvbbSFe625Fw.png" /></figure><p>Even the content of the AI exams is <strong>mostly theoretical</strong>. In my experience, to really understand AI, it’s better to do theory now than when you’re done with the BSc, because the chances of you picking a book and having plenty of time to study it are low after you finish. Again, <strong>if you just want to call an API to build a cool product without caring about what is going on in the background, this is not the degree for you.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UhB3rkpywXB4QO1PX2uCkg.png" /></figure><p>While there is always a lot of math, from the second year, <strong>you start doing practical things and projects</strong>. So you get your hands dirty, starting with statistics, where you learn how to work with data, then with ML and AI Lab, and finally go on to a few projects per semester in the last year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Opt5BSCyj2u00YN4P4ri8w.png" /></figure><p>If you’re wondering whether there is too much theory or practice, <strong>I generally felt that it was more theoretical than practical</strong>. I am more of an applied person, so I sometimes suffer from studying too many theoretical things, but in full honesty and in hindsight, it was not that bad. I think doing more theory than practice, especially early on, was beneficial to me. I also asked that in the survey, and people seemed generally ok with the balance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IQXR6jdpcChvk0HvfHSQzA.png" /></figure><p>I talked a lot about math, but that’s not the end of the story. <strong>You do a lot of computer science and scientific modeling, too</strong>, because AI is a multi-disciplinary field. The computer science courses end up being advanced towards the end, and they make you quite knowledgeable on tiny little things. The same goes for scientific disciplines: you do physics, from classical to quantum, to neuroscience, and economics, from basics to finance. And let me add that knowing about these things, especially physics, gives you a<strong> real edge</strong> compared to other people who study only AI, because lots of AI ideas are inspired by physics and related sciences.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XSHw9m7gsJQXECjaPGhQag.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l77ywTskkWqXb8DN1sZMRg.png" /></figure><p>Does this explain the name well enough? Is there anyone still thinking it’s <em>just marketing?</em></p><h3>Faculty &amp; Teaching Quality</h3><p>Now that you’ve seen what you’ll be studying, let’s talk about who will teach you and how.</p><p>Bocconi is a business school, but as I said before, they hired several physicists, mathematicians, and computer scientists over the years. And not random ones, but high-quality. So <strong>you’ll be, most often, in good hands.</strong></p><p>In my experience, I always felt quite satisfied with the faculty. Of course, like anywhere in life, I liked certain professors more than others, but the overall quality is very good.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ISZ49N2lCc2-1iKHPfl2TQ.png" /></figure><p>The good thing is that most of them teach topics very close to their research, so whenever you have questions, they’ll most likely have an answer. And if you want to go beyond class, they are not a***oles: they’ll try to help you… <em>Apart from one that I still hate to this day for a grade he put… I am sure some close friends of mine can guess who I am referring to :)</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ABVFZkKqFq6nGRuJGTCIWw.png" /></figure><h3>Class &amp; Student Life</h3><p>Class sizes usually range around 100 people in the first weeks. While I don’t have official data at hand, some people drop out quite early, so usually you end up with <strong>roughly 70/80 people </strong>by the end of the first year. Attendance is not mandatory, so in class, physically, you are usually half of that, and I can’t stress enough how nice this is. There’s nothing worse than going to a room of hundreds of students where you need a telescope to see the blackboard and cannot even hear the professor. This is not happening here and generally at Bocconi. <strong>You’ll know everyone in your class</strong> and won’t need a telescope either.</p><p>Community-wise, <strong>we’re all aware of being on the same boat</strong>. From my experience, we are not like the sharks studying finance that do anything to penalize each other (<em>I heard some crazy stories about them, by the way</em>). We do help each other. Of course, it may vary based on the year, but I guess this is a general sentiment.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1amhsV3Q6ffgfYcPq_Kr5Q.png" /></figure><p>In terms of diversity, <strong>Bocconi itself is very international, and the same applies to the BAI class</strong>, which is another amazing thing. You’ll meet so many people from different countries, both at BAI and at Bocconi in general, and you will be more than happy to get drunk with their local alcohol. Also, generally, Milan is a very university-friendly city, so you’ll be able to have lots of fun. Although in truth, this degree doesn’t allow you to have as much fun as peers studying economics, purely because of the complexity of the subjects. So you’ll have to consider a trade-off between academic success and external activities.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0QRh8lsLr8mmHo6feTr3pQ.png" /></figure><h3>Career Outcomes &amp; Opportunities</h3><p>This is what matters the most in the end. It’s from here that you can see whether a program is good or not. The general feeling from students is that <strong>the program prepares you well</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bmFXv2b9Ivh85zkSunsJ7g.png" /></figure><p>But to really understand if that’s true, we need to understand what students actually do after. First of all, <strong>most people continue their studies</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*viTJN_ndXd15j879Lr7odQ.png" /></figure><p>Quite a few people end up in schools around the world: Stanford, Harvard, Berkeley, NYU, Oxford, Cambridge, Imperial, Ecole Polytechnique, ETH, EPFL, Tsinghua, you name it. Some stay at Bocconi. The <strong>subjects have a wide range</strong>: AI in some specialized fields, finance, pure math, engineering, physics, etc. As I said before, the curriculum is so multidisciplinary that it gives you all the knowledge and instruments to apply at multiple degrees.</p><p>Workwise, although the market is tough, <strong>job placement is not that hard</strong>. Roles range from pure software engineering and data science to AI research and quantitative finance. The companies are really top-notch, and a simple LinkedIn search could convince you that this is no baloney.</p><p>As a personal comment, I’ve met really smart people in those three years at BAI, so <strong>I won’t be surprised if in the future I hear some names again with really successful careers</strong>, whether in corporate, startups, research, or whatever else. The talent density is really high.</p><h3>Extra Personal Comments</h3><p>I am not here to make suggestions about university or on how to live it at its best (<em>I’d have to write a different article for that</em>). All here is related to BAI.</p><h4>Random Things</h4><ul><li><strong>If you want to succeed academically, you need to struggle.</strong> It’s a harsh truth that you need to be aware of. It’s not a degree in the science of snacks; it’s a degree for AI, one of the most revolutionary technologies in decades, and it’s a complex subject. <strong>Parent yourself, be disciplined, have fun, and you’ll have an amazing time</strong>. Once done, you’ll feel very accomplished and will be able to stand on the shoulders of giants and look further ahead. <em>Philosophical, but true.</em></li><li><strong>You’re in a business school. This must not be a reason for you not to join.</strong> You’re much more likely to know about AI and related fields than other people in other universities, because of your very quantitative math background, of the smaller classes, and of the laser-focused exams you have to take. <em>Sounds a bit cocky, I know, but it’s kinda true.</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LQxRklURJxwTaA1XvNs-0Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p-h3YsEnsbGSdA62ttenEQ.png" /></figure><ul><li><strong>You’re in a business school, and since you study math and AI, everyone will call you a genius.</strong> Indeed, ninety percent of the people at Bocconi study economics, so whenever they hear you do BAI, they’ll say things like “<em>Oh, so you’re very intelligent!</em>”, “<em>Damn, you’re smart!</em>” and stuff like that. If you meet those people, go talk to them. It boosts your ego, hihihi.</li><li><strong>You’re in a business school, so you’ll feel like a fish out of water</strong>. This was at least when <em>I </em>was a student, and the program was new. There were not as many connections to companies and exchange universities that satisfied the BAI needs, since most of them were business-related. The same for university associations. However, from what I can see today, <strong>things are improving a lot</strong>. First of all, now there is a set of former students who can provide their experience (<em>hey, here I am</em>), several industry connections, and many more student associations that do a lot more in the AI field. So <strong>the environment is much more adapted to BAI students than it was before, and this creates even more career opportunities during and after the bachelor’s.</strong></li><li><strong>You’ll be surrounded by really smart people.</strong> Rumor has it that BAI is one of the hardest programs to get into at Bocconi, so the selection is quite tough. To put it in context, I entered with a 1490 SAT and roughly an 85/90% high school average (<em>which is good</em>), but I am not sure I would have entered in successive cohorts, as the quality of students increased a lot, especially because of more people willing to study AI after ChatGPT’s release. Therefore, you’ll be surrounded by talented people who understand and grasp very advanced concepts very easily, and if you’re a normal person who <em>needs more time to understand</em>, just like me, <strong>impostor syndrome will kick in</strong>. Don’t let it discourage you, and use it as a boost to diffuse motivation and improve even more. If you feel impostor syndrome, it means that you’re in the right environment that is pushing you outside your comfort zone.</li><li>On this note, one thing I learned while studying linear algebra in my first year is that<strong> there are no hard things in math (<em>and life, too</em>); there are just things that require more time to become easy.</strong> Keep this in mind, and you’ll be happy to struggle a bit more than your brain wants to.</li><li><strong>Not all exams will be useful</strong>,<strong> </strong>because the curriculum is quite diverse. For example, I couldn’t care less about economics and finance exams, cause I simply didn’t care, but for others they were their favorites, and those people are now doing quant or studying finance. Or you’ll face moments in your path where you don’t want to study a useless proof that you’re sure you’ll never use in your life again. But the truth is that you’ll have to do useless things in life every time, so <strong>the sooner you get used to it, the better you’ll be able to deal with this feeling later on, both at BAI and in life</strong>.</li><li><strong>This degree does not teach you how to be a software engineer.</strong> It’s teaching you how to reason about problems in a mathematically sound way. You will not be an expert leet coder once you exit. That’s a downside in my opinion, cause a lot more computer science things could be useful, but <strong>the program is more oriented towards math than computer science</strong>. By the way, you’ll still be able to code very advanced things even outside AI, don’t get me wrong. Also, you’ll have the tools to know everything you want because you’ll have studied the fundamentals, the things that don’t change.</li></ul><h4>BAI vs Engineering</h4><p>Please, don’t compare BAI to engineering. <strong>They’re different things</strong>. Engineering is giving you a specific <em>forma mentis</em>, mostly applied. An average engineer is taking a set of precompiled notions and applying them with <em>ingegno </em>(i.e. ingenuity) to a problem. At BAI, you learn to do this, yes, but you do something more: you learn how to craft those notions. I am not marketing, but I challenge you to find engineering students who publish papers one year after and submit to top-tier conferences. Probably more engineers do it just because there are numerically more engineers around the world than BAI students, but if you consider the proportions, I think you’ll see what I mean. So maybe from this side, we can also confirm that BAI is more of an academic degree than an applied one. Again, <strong>not saying that one is better than the other. I am just trying to say that they are different things. So don’t compare them with the same weight scale.</strong> If you’re undecided between BAI and some type of engineering, remember that they’re different things.</p><h4>The Elephant In The Room</h4><p>The degree is a few years old, so it is not as well structured as other historic courses. For example, in my experience, some courses overlapped the content with other subjects, certain syllabi were different from the ones we’d actually do, and some professors did pretty much whatever they wanted. At times, I felt that some topics were a bit detached from the core focus of the degree. These are common complaints that I totally agree with, but the truth is that it happens in every uni, and surely less at Bocconi than elsewhere. Still… it happens, so don’t be amazed if it happens to you as well.</p><p>Moreover, consider that the degree is not that personalizable. Most of the program is fixed, and you can only choose two electives in the last year, from a very limited pool. The good thing is that the faculty knows about it, and the professors listen to us. The bad thing is that they’re slow at changing things. I don’t know why, <em>probably Italian bureaucracy…</em></p><h3>So???</h3><p>I hope you have a much clearer idea of what BAI is and is not.</p><p>I repeat that this is a personal review. Some people may disagree with me, but I surveyed five different classes for the precise reason of not making it about me only. <strong>If others want to have their say, please comment below or write your own article.</strong></p><blockquote><strong>It will boost the name of the program and give more awareness to external people.</strong></blockquote><p><strong>Overall, I did enjoy my experience at BAI. </strong>I learned a lot, technically and humanly, met crazy good people, and had fun. But since it’s not just about me… generally, people seem satisfied too!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r7BjDIHfeZuf2wW0AzTMzw.png" /></figure><p><strong>Would I choose it again? Absolutely yes.</strong> I am more than happy with the directions my life is taking, thanks to the doors that BAI opened for me. Would others choose it again? Most say yes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9tevN_HYQ5Bi5oU3kRjmpQ.png" /></figure><blockquote>So, is it worth it?</blockquote><p>If you like math, want a rigorous technical foundation, and see yourself in research, AI, or some other advanced technical discipline, then yes, <strong>BAI is probably one of the best degrees you can pick right now</strong>. If you hate theory, just want to code, or whatever else, you’ll suffer. I hope that the previous sections can give an understanding of why that is the case.</p><p>But who am I to tell you what it is like with so much certainty? After all, I am just a guy pressing buttons on a laptop, so the only way to know if this is true is to experience it. I do sound like I am encouraging you to apply, and that’s probably what I am doing: <strong>I strongly recommend this program.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A4D-LHjItlyIeuAjfz-SBg.png" /></figure><p>Let me add final comments coming from people who responded to the survey:</p><blockquote>I want to tell all the future BAI students one thing: if you decide to enroll in BAI, please do not study only for pass exams or for the GPA. You must put your focus into what you learn and have a broad view about the importance of each concept in a bigger scheme (like AI in general), but more important is to improve yourself and your learning method. You will see how understanding a concept will satisfy you more than a mark, and how your ability to understand will improve exponentially course after course. Do not compare yourself to other students. Everyone has its own interests and hobbies. A lot of things are exaggerated, like LinkedIn profiles, so focus on how to improve yourself, not how to reach the others.</blockquote><blockquote>Amazing course, just be aware of 2 things: 1. However smart you are, this is not high school, and you need to put in a lot of work if you want to do well 2. Do practical things outside of the course and start early!</blockquote><blockquote>As Zecchina once said, BAI is hard. It takes away the freedom of choosing what to study and have free time (let’s be honest, if you want a high GPA in bai you will not have much free time). In return, it gives you so much freedom of choosing what to do afterwards. I think this tradeoff is worth it.</blockquote><blockquote>Please inform yourself better before making any decisions. Talk to people who are doing BAI or have graduated and try to get as much information as you can from different sources. Because the name sounds appealing but make sure you know what you’re getting yourself into.</blockquote><blockquote>BAI is a big commitment. You will work harder than your peers at Bocconi and have a larger workload than many friends at other universities. But it is definitely manageable. As you learn to deal with the workload, you also have lots of time for fun in Milan, which is a great city.</blockquote><blockquote>Don’t enroll if you think this is just another CS course, where people tell you “you don’t need math”. There are very heavy math subjects, so if you just want to pursue software development or frontend development, BAI is not a place for you.</blockquote><blockquote>This is a perfect bachelor&#39;s to pursue AI</blockquote><p>Oh, let me add one more thing. While BAI is mathematically challenging, you don’t need to be the next Terence Tao or Albert Einstein to come to BAI! You’ll likely meet a few geniuses, and maybe you’re one of them, but to be honest</p><blockquote>determination and awareness of what you’re getting into are all you need.</blockquote><p>To give you an idea, I am not mathematically gifted at all (<em>I could prove it to you easily</em>); I just work hard, like what I study, and I am here writing this with a BAI degree and quite a lot of knowledge about AI and related fields.</p><p>So,<strong> don’t reject yourself before you have even applied.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03809bc68ef8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My First Hackathon]]></title>
            <link>https://medium.com/@sandromikautadze/my-first-hackathon-508ebcab7529?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/508ebcab7529</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gaming]]></category>
            <category><![CDATA[hackathons]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Tue, 04 Feb 2025 19:16:45 GMT</pubDate>
            <atom:updated>2025-02-04T21:33:43.325Z</atom:updated>
            <content:encoded><![CDATA[<h4>The Story of How We Reached the Finals and Secured 2nd Place in the Community Award</h4><figure><img alt="people presenting" src="https://cdn-images-1.medium.com/max/1003/1*gLAgEhxOyKKqAyo631anbQ.png" /><figcaption>The M1X team during the final pitch. From the left: Alessandro, Mattia, Felix, me, Emiliano, and Michael</figcaption></figure><p>This is a detailed personal summary. Some events are exaggerated for the sake of humor. <strong>Enjoy</strong>!</p><p>· <a href="#c4a0"><strong>The Choice of Participating</strong></a><strong><br>· </strong><a href="#4ec8"><strong>Day 1 — January 24, Friday</strong></a><br> ∘ <a href="#191e"><em>The Setup</em></a><em><br> ∘ </em><a href="#723b"><em>Locking In</em></a><em><br> ∘ </em><a href="#bd63"><em>The (Sleepless) Night</em></a><br><strong>· </strong><a href="#617f"><strong>Day 2 — January 25, Saturday</strong></a><strong><br></strong> <em>∘ </em><a href="#6442"><em>The Rise and The Fall</em></a><em><br> ∘ </em><a href="#f512"><em>The Pivot</em></a><em><br> ∘ </em><a href="#7b20"><em>The Night and The End</em></a><br><strong>· </strong><a href="#ff4c"><strong>Day 3 — January 26, Sunday</strong></a><strong><br></strong> <em>∘ </em><a href="#ca36"><em>Pre-Judging</em></a><em><br> ∘ </em><a href="#1f45"><em>The Announcement</em></a><em><br> ∘ </em><a href="#2fcb"><em>The Final Pitch and The Winner</em></a><br><strong>· </strong><a href="#cc36"><strong>The Community Award</strong></a><strong><br>· </strong><a href="#2543"><strong>I’m Done</strong></a></p><h3><strong>The Choice of Participating</strong></h3><p>Paris is constantly full of cool events, and hackathons belong to this category. For those of you who do not know what a hackathon is, Google defines it as “<em>an event in which a large number of people meet to engage in collaborative computer programming.</em>” Some of my university friends and I came across a hackathon organized by <a href="https://www.joinef.com/">Entrepreneur First</a> in collaboration with <a href="https://mistral.ai/">Mistral AI</a> and hosted from January 24 to January 26. Hoping to connect with people from big companies — and maybe, just maybe, beg for an internship — we registered. Little did we know that registration was not enough to get accepted, we also had to pass a highly selective screening process. This made things more complicated and stressful. Indeed, the hackathon was called <strong>Mistral AI Game Jam</strong> and we had applied just focusing on the first two words, not the last two. We had neglected the fact that the hackathon was mostly about creating a game. The stress was not coming from the fact that we did not know anything about gaming (“<em>You can learn it fast, right?</em>”), but because they were looking for people with some game development experience, something most of us didn’t have.</p><p>Yet, one week before the start, to our great surprise, we got accepted to the hackathon amongst hundreds of applicants. We were extremely happy, but we could not foresee what lay ahead of us starting from January 24.</p><h3>Day 1 — January 24, Friday</h3><h4>The Setup</h4><p>The hackathon was hosted at the beautiful Paris offices of Photoroom. We arrived there at 9.30 am to get our free breakfast (all food was free, by the way) and at around 10.30 am Alessandro, Mattia, Emiliano, Michael, and I (aka the M1X team) sat down at our carefully chosen desk on the upper floor to start throwing out ideas about what game to make. The only instruction that we were given was to generate the game around the following theme:</p><blockquote><strong>“You don’t control the character.”</strong></blockquote><p>This opened so many interesting possibilities, but overwhelming at the same time, especially for a team with just AI, and no gaming, experience.</p><p>At around 11 am a tall stranger approached us at our table, looking for a team to join. He said he had AI and software engineering skills. After 10 minutes of weighing whether he’d be a good fit (while he wandered between teams), the stranger — Felix — officially joined M1X. With our team assembled — like the Avengers — we dove into brainstorming game ideas. We developed two, let’s call them <strong>Plan A</strong> and <strong>Plan B</strong>.<strong> </strong>We decided to work on plan A and keep plan B as a last resort. At 12.30 pm we went to get our free lunch.</p><h4>Locking In</h4><p>Then came time to lock in: we needed to start hacking — the term used in hackathons for “working.” Based on our original plan A we needed three main teams to operate in parallel: one to work on the textual part using Mistral’s models, one for the reinforcement learning, and one to work on the game development.</p><figure><img alt="a whiteboard" src="https://cdn-images-1.medium.com/max/519/1*prxmS7W0rtFellXzwJFrag.png" /><figcaption>Our initial plan. It’s ok if you don’t get it; we didn’t either at the beginning</figcaption></figure><p>I was assigned the game dev task. Or rather, I assigned it myself. Yes, because knowing that we would have needed to make a game, a few days before the event I started learning Unity and I created some little funny games. So, I proudly said to the team:</p><blockquote><em>“I am your savior; I will be your game dev!”</em></blockquote><p>Ten minutes later, I wanted to kill myself.</p><p>In fact, you cannot learn how to make a game fast, as I initially thought. Following a copy-paste tutorial — where everything is done for you — is one thing. Building a game from scratch? That’s a whole different challenge, and I felt completely lost. <strong>Challenges are for the brave</strong>, so I started grinding — working on graphics, game mechanics, and anything else that might be useful — with Felix helping along the way. Most of the game dev work was done on my computer and Felix was helping me out and searching for secondary things on his laptop. Occasionally, I would fully lock in and live in my world of Unity struggles, leaving Felix a bit taskless (sorry, I did not mean it); yet in the afternoon, he came to me at random and said: “<em>Hey, look!</em>”, I glanced at his MacBook and there was a little demo of our plan B. “<em>Cool,</em>” I said, “<em>You have made a nice working MVP; it looks great!</em>”. And that was it. After a few seconds, we had forgotten about it.</p><figure><img alt="two people working on their laptops" src="https://cdn-images-1.medium.com/max/780/1*9D0iRJbw0TpL0NBd3Im2iQ.png" /><figcaption>Locked in, as you can tell</figcaption></figure><p>We kept grinding, and refining plan A continuously. The beauty of a hackathon is that you start big, and slowly realize that certain ideas are not feasible in two days, while others are well within your reach.</p><blockquote>The hard part is figuring out what belongs in which category.</blockquote><p>I think that we spent most of our early afternoon trying to classify the complexity of certain tasks, leaving us a bit far behind compared to other teams that already had fully functional game mechanics. But in the end, after a few back-and-forths and with the huge help of the mentors, we ended up with a very polished project idea, that would have showcased our AI knowledge and would have been possible to implement given our limited gaming experience.</p><p>After a few hours of work, we realized that we were aiming too high. We ended up removing the reinforcement learning part: there was not enough time. By 7:30 PM, our stomachs demanded attention, so we took a quick break for yet another free meal.</p><p>With full stomachs, we were ready to continue again and we split tasks to optimize the time.</p><figure><img alt="notes" src="https://cdn-images-1.medium.com/max/439/1*IOttRDK5tiJSpRgDcRBWog.png" /><figcaption>8.26 pm</figcaption></figure><p>On my side nothing had changed: I was still a game developer. After implementing basic game movement mechanics in the afternoon, I started creating from scratch pixel art of the tools that we would have needed during the game. Quite a tedious work, yet necessary. I even made a hoodie with Mistral’s logo — such a pity that nobody got to admire that piece of art in real life!</p><figure><img alt="a pixel art hoodie" src="https://cdn-images-1.medium.com/max/497/1*Sp7tT0VAcRhfHflRPE4wXQ.png" /><figcaption>Yes, I did it myself. Even the logo!</figcaption></figure><p>Overall, this defined my evening. The others were working on the LLM part, trying to get the best results from specific prompts and understanding how to connect the Mistral API to Unity. And if I remember correctly Felix was occasionally playing around with the MVP for the plan B, just for fun. I yet have to understand what he saw in that idea, to pursue it even as we were smoothly doing plan A. He probably understood something that the original M1X team could not.</p><figure><img alt="people working" src="https://cdn-images-1.medium.com/max/704/1*ZirQXEXI47lyz3oTCAZH7Q.png" /><figcaption>11.04 pm</figcaption></figure><h4>The (Sleepless) Night</h4><p>A hackathon is not a real hackathon unless you do an all-nighter — or so I thought. To my surprise, many teams started heading out around 10 or 11 pm. As I watched them cross the threshold that would have not let them back in the offices until 7 am in the morning (aka the door), the office grew quieter. Some people were sleeping at their desks; others were looking at their screens like zombies. I was one of them. I sat at our desk, staring blankly at my laptop, and mentally realizing that I would have needed to pass an entire night awake. The project required a lot of work.</p><figure><img alt="fridge" src="https://cdn-images-1.medium.com/max/579/0*KvsSfy6vT-E1WSgH" /><figcaption>Sources of energy. I’m referring to the orange juice — I don’t drink Red Bulls</figcaption></figure><p>Before I tell you about our nocturnal adventures, I should probably tell you a bit more about the infamous plan A. This is an overview:</p><blockquote>You are a parent about to leave on a business trip, and you need to leave your toddler alone for the weekend. You can only leave a note with a set of instructions to help them survive. The instructions are by you and thorugh an LLM, and the goal is to guide the toddler toward survival objects (e.g., an apple to eat) and away from dangerous ones (e.g., a knife) using as few commands as possible — all while trying to reach a specific location. The game consists of multiple levels with increasing complexity.</blockquote><p>We decided to call it <strong>REMOTE PARENTING</strong>. (<em>PS: yes, it is very similar to another game of the hackathon (Parental Control), and we were told that someone else had a similar idea, but we had no interaction with them until the pitch day</em>). Now that you understand the plan, let me tell you this: our goal was never to win — we made that clear from the start. We wanted this project to be useful for us, to sharpen our AI skills and apply them in game development. But to do that, we had to put in a lot of effort, which is why we stayed up all night.</p><p>I’ll be honest with you — I don’t remember much of what happened that night. Not because I was drunk, but because I was too tired. I can recall three things:</p><ul><li>We took turns to sleep.</li><li>We didn’t do any Unity.</li><li>We spent the entire night researching which Mistral model to use to guide the player’s movements.</li></ul><p>For the last point, we could have blindly chosen one model, but we didn’t want our characters running around like headless chickens — we wanted them to actually make sense (revolutionary concept, I know). So during the night, we developed a bunch of evaluations to test how different models would handle game situations and decision-making. Let me add that Diogo’s mentorship here was crucial in making us understand how to structure our evals. If you’re reading this, thank you from everyone!</p><p>At roughly 4.30 am my brain essentially stopped functioning and I went to sleep. You may be asking yourself where: at a hotel? Hostel? Or maybe at a friend’s place? The answer is simple: at the office. Trust me if I tell you that those offices were crazy good, with resting areas, and even showers. I crashed for roughly three hours and at 9 am I was back at the desk, slightly groggy but ready to start a new day at the hackathon.</p><h3>Day 2 — January 25, Saturday</h3><h4>The Rise and The Fall</h4><p>I woke up to find a paper neatly written out. We had started the draft at night, but as I was sleeping, Micheal, Ale, and Emiliano had finished it. You can see a draft at <a href="https://msheroubi.github.io/mmrs_20250127.pdf">Michael’s website</a> [last visited 04/02/2025].</p><p>At 9.30 am we went to grab our free breakfast and we were ready to start. Now, we knew what model to use and what instructions to give, but remember how we did zero game development during the night? As the deadline was approaching, we were still missing key elements. From that moment on I started watching tutorials, asking for help from mentors, and who knows what else. My brain was braining. Our team brain was braining, even as we were eating our free lunch. We gained momentum. I was completely focused on my laptop and Unity. In so little time we did so much, especially considering our limited gaming experience.</p><p>Everyone was working smoothly. Everything was going amazingly well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1004/1*OY1RanwgmIcx8SzGtXrkjQ.png" /><figcaption>The initial interface of our game</figcaption></figure><p>Then I looked at the clock. It was 6 pm. We had to upload the game at 8 am on Sunday, so less than 14 hours remained until the deadline. Our game was still a skeleton of our ambitious vision. The to-do list seemed to mock me: characters unfinished, gameplay mechanics incomplete, levels not even started. My confidence disappeared in two seconds. I told my guys that I needed a break. The few hours of sleep, my inexperience with Unity, and my brain on the verge of meltdown were not helping at the most important moment. Thankfully, my teammates understood and continued working on Unity while I took a short break.</p><p>As I decompressed in the rest zone, telling Mattia how many hours I had left before going totally insane, I heard a voice from my right directed to me and Mattia: “<em>Are you Italians?</em>”</p><p>It was Cosimo, a guy from Florence who had come to the hackathon on his own. He told us about the game he and his team were making. It had amazing graphics, sound design, and things that seemed impossible to conceptualize in less than 48 hours. I was shocked, like Matteo Renzi’s first reaction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/537/0*Y_j0xz_1wj7SqK1k.png" /><figcaption>Me listening to Cosimo’s game</figcaption></figure><blockquote><strong>Spoiler alert</strong>: Their game (Text To Survive) won the hackathon the next day!</blockquote><p>As I listened to Cosimo, I felt my confidence deflating like a sad balloon. We were, struggling to make basic mechanics, and they looked like they were ready to launch on Steam. I went back to our desk. Micheal and the others had done a huge sprint to implement a basic playable scene — a huge improvement. But that was just a baseline. We needed to do way more and the clock was ticking.</p><figure><img alt="level" src="https://cdn-images-1.medium.com/max/888/1*b7fWlI_y6wKYCHN6Z5SQOw.png" /><figcaption>The baseline: you can see how far we were from a complete level</figcaption></figure><h4>The Pivot</h4><p>With my batteries recharged and a free dinner fueling me, I was back to work, but something felt… off. The excitement from earlier was fading, and the Unity interface was slowly morphing into an incomprehensible mess of buttons and menus. I could sense that the others were feeling it too.</p><p>I tried to focus. My task now was to add a background image for the first level, making sure it matched our pixel art theme. I put my headphones on, switched to noise-canceling mode, and locked in. Then, through the silence, I heard something. Muffled voices. I looked up: Alessandro was looking at Felix’s laptop. “<em>Wait… is that the other game we were discussing to make yesterday?</em>” said Ale. Felix nodded. Ale, intrigued, leaned in: “<em>This actually looks super fun. Maybe… we should give it another thought…</em>”</p><p><strong>Silence.</strong></p><p>Did he really just suggest throwing away two days of work a few hours from the deadline? Was he seriously proposing a switch to plan B now?</p><p>We all crowded around Felix’s screen. The game wasn’t any different from what he showed me the day before. It was technically simple — probably too much, which is also why we didn’t pursue it initially — but it was polished; it looked good; it was functional, and most importantly, it was fun.</p><p>Then we looked back at our own game. Was it fully functional? No. Did it look good? Not really. Was it finished? Not even close. Was it fun? …not much.</p><p>Were we really abandoning plan A? At first, we hesitated. A heated discussion followed. Mentors got involved. Arguments were made. The clock was ticking — fast. The offices would close at 10 pm and we had to choose something — super fast. At 8 pm, we decided to pivot.</p><blockquote>From that moment on, <strong>plan B</strong> became our new <strong>plan A</strong>.</blockquote><p>It was not easy to drop plan A, because we spent time brainstorming, coding, asking for feedback, and — most importantly — losing sleep over it. It felt almost wrong to let it go. But deep down, we all knew the truth: plan B had a real shot. And in that moment we even switched mindset: we wanted to win.</p><p>This change brought a piece of good news for me: no more Unity! Unlike our original plan, this game required no game engine, so we split the tasks differently this time. Given the limited time, we focused on defining the rules, gameplay mechanics, and a basic UI.</p><p>We ran the new idea by the mentors, got feedback, made some quick improvements, and before we knew it, the clock struck 10 pm.</p><h4>The Night and The End</h4><p>We had to leave the office for the night. Felix headed to his hostel, Emiliano made the long journey back to beautiful Palaiseau, while Micheal, Ale, and I stayed at Mattia’s place — thankfully large enough to fit the four of us. But just because we left the office, it didn’t mean the work was over. Far from it. We kept refining the game to make everything intuitive from the moment players hit “Play.” We had until 8 AM to submit the final version, and every detail mattered. By 2 AM, we finally pushed the final version to the platform.</p><p><strong>That was it.</strong></p><p>After two days of hacking, pivoting, and sleepless grinding, our hackathon was officially over. I collapsed onto whatever available surface I could find and tried to get a few hours of rest to get ready for the last day.</p><h3>Day 3 — January 26, Sunday</h3><h4>Pre-Judging</h4><p>With all the games submitted, our fate was no longer in our hands.<strong> </strong>Everything now depended on two main things:</p><ul><li>The other teams’ games</li><li>The judges&#39; opinions</li></ul><p>We could still sell our idea with a strong pitch, but ultimately, the outcome wasn’t up to us anymore.</p><p>The day was structured like this: at 10 am, all 17 teams would pitch their games solely to a pre-jury of mentors and organizers. Each game would be scored on five criteria, and only the top six would advance to the final round — where we’d pitch in front of the official guest jury and the entire audience.</p><p>We arrived back at the office at 9 am, just in time for — you guessed it — our free breakfast. Then, we fine-tuned our pitch and the slide deck and by 10 am, we were ready.</p><h4>The Announcement</h4><p>The pitch went well and the judges seemed to genuinely enjoy our game. Afterward, we wandered around the office, making small talk, grabbing our final free meal, and hoping for the best.</p><p>At 1.30 pm, we all gathered in the main area as they introduced the jury. Then, the moment of truth: they started announcing the finalists.</p><p>First up: <em>Parental Control</em>.<br>Then: <em>NeuralJam</em>.<br>Followed by: <em>Text To Survive</em>.<br>And then: <em>Not Me</em>.</p><p><strong>Two spots left.</strong></p><p>We were starting to doubt that we had made it. Why? Each finalist’s game was displayed on the screen, alongside a picture of their team standing proudly beside a TV monitor showcasing their game. The photos had been taken moments earlier.</p><p>No one had come to take our picture.</p><p>Then we heard it:</p><blockquote>“M1X.”</blockquote><p>Wait, what?! <strong>We made it?! We’re in the top six?! LET’S GO! </strong>I guess it’s always useful to have plan B. And I guess Felix saw that we could have done great with that!</p><p>Strangely enough, they had announced our team name, not our game name, as they did for everyone else.</p><p>Oh, yeah, I forgot to tell you. Our game was called <strong>Think In Sync</strong>, and you can still play it at <a href="http://www.think-in-sync.com/"><strong>www.think-in-sync.com</strong></a> (not the original version from that day — this is a much-improved one).</p><h4>The Final Pitch and The Winner</h4><p>The order in which the finalists were announced also determined the presentation order, meaning we got to watch all the other teams before our turn. And wow — they were impressive. I had my personal favorites… but I’ll keep them to myself (<em>hihihi!</em>). Then, our moment arrived.</p><p>We had refined our presentation a bit since pre-judging, and it showed. The crowd was engaged, and we were having a great time presenting. If you’re curious you can see the pitches on <a href="https://www.youtube.com/live/nbLhijLu_f4">YouTube</a> (ours starts at <em>1:18:40</em>).</p><figure><img alt="people" src="https://cdn-images-1.medium.com/max/1024/0*xAiSZkiPzcbfgnQH" /><figcaption>The crowd</figcaption></figure><p>Of course, we knew we had no chance of winning — this time, for real. The other games were simply on another level — the level we had hoped to reach in those two days but didn’t quite get to. Then came the final announcements: second place to <em>Parental Control </em>and first place to <em>Text to Survive</em>.</p><p>And just like that, the hackathon was officially over.</p><p>Or was it?</p><h3>The Community Award</h3><p>While the in-person event had ended, we had one last prize— a Community Award, where the general public could vote for their favorite game. And honestly, that made perfect sense: a game isn’t just meant to be judged by a panel of experts — it’s meant to be played and loved by a community. We had one week to get as many people as possible to vote for our game on Hugging Face.</p><p>We added new features, fixed bugs, shared as much as possible, and got tons of likes on likes. In less than a week, our game was played thousands of times worldwide. We were beyond proud of the public reception!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iEreRrUFOO2BKnHTmHFtnQ.png" /><figcaption>Some stats as of February 4, 2025 (7 days after launch)</figcaption></figure><p>The last valid day to put likes was Monday 3rd and we ended up with 212 likes. We reached second place. But we had a very close and tough race with the winner: <em>NeuralJam. </em>Congrats!</p><p>Our prize for second place? A gigantic nothing.</p><p>Nonetheless, it was tons of fun. It felt like a small startup</p><blockquote>and I loved it.</blockquote><p>In case you are interested, you can try all the games of the hackathon on the <a href="https://huggingface.co/Mistral-AI-Game-Jam?sort_spaces=likes#spaces">Mistral AI Game Jam space</a>.</p><h3>I’m Done</h3><p>I have nothing else to say; actually, I think I said too much. Thanks to everyone involved in this, especially M1X.</p><p><strong>GG</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=508ebcab7529" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Read These Papers If You Want to Understand Diffusion Models]]></title>
            <link>https://medium.com/@sandromikautadze/read-these-papers-if-you-want-to-understand-diffusion-models-545a3639ca3c?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/545a3639ca3c</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[diffusion-models]]></category>
            <category><![CDATA[genarative-ai]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Tue, 14 Jan 2025 19:27:44 GMT</pubDate>
            <atom:updated>2025-01-14T19:27:44.717Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A glass of water with swirling red and blue dye mixing." src="https://cdn-images-1.medium.com/max/1024/0*rbRqlUeYQ5xJVJ3P" /><figcaption>Photo by <a href="https://unsplash.com/@chaozzy?utm_source=medium&amp;utm_medium=referral">Chaozzy Lin</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>You’ve read the title, so you expect suggestions on papers to study diffusion models. Before that, I think it’s worth starting with a brief, intuitive introduction to what diffusion models are. If you already have an intuition about them, feel free to skip this coming section and jump directly to the papers.</p><p><strong><em>· </em></strong><a href="#9353"><strong><em>A Simplified Explanation of Diffusion Models</em></strong></a><strong><em><br>· </em></strong><a href="#93d8"><strong><em>Papers To Read</em></strong></a><em><br> ∘ </em><a href="#3bc4"><em>1. Deep Unsupervised Learning using Nonequilibrium Thermodynamics (2015)</em></a><em><br> ∘ </em><a href="#750b"><em>2. Generative Modeling by Estimating Gradients of the Data Distribution (2019)</em></a><em><br> ∘ </em><a href="#7e29"><em>3. Denoising Diffusion Probabilistic Models (2020)</em></a><em><br> ∘ </em><a href="#39cd"><em>4. Denoising Diffusion Implicit Models (2020)</em></a><em><br> ∘ </em><a href="#b54d"><em>5. Diffusion Models Beat GANs on Image Synthesis (2021)</em></a><em><br> ∘ </em><a href="#e75b"><em>6. Classifier-Free Diffusion Guidance (2022)</em></a><em><br> ∘ </em><a href="#6db5"><em>7. Cascaded Diffusion Models for High Fidelity Image Generation (2021)</em></a><em><br> ∘ </em><a href="#d09f"><em>8. High-Resolution Image Synthesis with Latent Diffusion Models (2021)</em></a><em><br></em><strong><em>· </em></strong><a href="#83aa"><strong><em>To conclude…</em></strong></a></p><h3>A Simplified Explanation of Diffusion Models</h3><p>Diffusion models are at the heart of the recent buzz in generative AI. Initially developed for images, they have also demonstrated impressive performance with audio, video, and other modalities. For simplicity, let’s focus on how diffusion models work in the context of <em>images</em>.</p><p>In one sentence, <strong>diffusion models allow to generate images by iteratively denoising a completely noisy image</strong>. To expand on this, the core task of generative modeling is to create new samples from an (unknown) data distribution. Diffusion models achieve this through a two-step process:</p><ol><li><strong>Forward Process</strong>: <em>Adding Noise </em><br> Start with a real image <em>x_0 </em>and gradually add small amounts of Gaussian noise at each step <em>t=1,…T</em>. By the time the image reaches the final step <em>t=T</em>, the image <em>x_T</em> approximates random noise. Notice that this process is <strong>deterministic</strong> and completely under our control: we decide how much noise to add and how many steps to take.</li><li><strong>Reverse Process</strong>: <em>Removing Noise </em><br>Start from random noise <em>x_T</em> and iteratively remove small amounts of noise over <em>T</em> steps, until <em>x_0</em>, the true image, is reconstructed. The challenge lies in figuring out how to remove the noise correctly at each step — it is an <strong>unknown</strong> function. Here’s where a <strong>neural network</strong> comes into play: it is trained to predict and remove the noise added during the forward process at each step.</li></ol><p>Once the neural network has learned how to denoise, we can flip the process: <strong>start from random noise and let the model reverse the steps to generate a realistic image</strong>.</p><p>At a high level, we dissolve an image into pure noise (forward process) and then learn to reconstruct it step by step (reverse process). This mirrors the concept of physical diffusion —and that’s where the name “diffusion model” comes from.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/422/0*ZQ907NYQWzLGVmKL" /><figcaption><strong>Diffusion in the physical sense. </strong>The forward process from left to right; the reverse process from right to left.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*x1UeuBdVXfP792yg.png" /><figcaption><strong>Diffusion in the AI sense.</strong></figcaption></figure><h3>Papers To Read</h3><p>With the intuition in place, let’s dive into some of the main papers that have shaped the development of diffusion models. The recommendations below are ordered to provide a smoother learning curve but feel free to explore them in any order that suits you. I’ve also included bonus papers for additional context.</p><p><strong>PS: </strong>The descriptions are very simplified and just focus on the high-level concepts of the papers. Reading these summaries won’t make you understand diffusion models, only reading the papers will! <strong>Click the links, take notes, and spend time truly understanding the material.</strong></p><h4>1. <a href="https://arxiv.org/abs/1503.03585">Deep Unsupervised Learning using Nonequilibrium Thermodynamics</a> (2015)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a5vLi29UVHpHIqfOMvCv8w.png" /></figure><ul><li><strong>Overview</strong>: This paper laid the groundwork for diffusion probabilistic models by addressing a key question: <strong>how can we model a dataset in a way that is both analytically solvable and generalizable to arbitrary data distributions?</strong> Their solution is to use a <strong>Markov chain</strong> to transform a simple, well-defined distribution (e.g., Gaussian) into a target data distribution through a diffusion process. This choice ends up being:<br>- <strong>tractable </strong>as each step of the Markov chain is analytically solvable.<br>- <strong>flexible </strong>as diffusion processes can approximate any smooth target distribution.<br>Understanding this paper will help you appreciate the probabilistic machinery behind diffusion models. It provides essential mathematical foundations for diffusion models and explains why diffusion is a strong technique for generative modeling.</li><li><strong>Key Takeaway</strong>: Mathematical foundations of diffusion models.</li><li><strong>Bonus Paper</strong> <a href="https://arxiv.org/abs/1306.1091"><em>Deep Generative Stochastic Networks Trainable by Backprop</em></a><em> (2013)</em>: published earlier, it explores iterative training via Markov chains and offers useful context for why these methods are effective.</li></ul><h4>2. <a href="https://arxiv.org/abs/1907.05600">Generative Modeling by Estimating Gradients of the Data Distribution</a> (2019)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/1*Ra7xpAcfXHFCoPzdOMYmMw.png" /></figure><ul><li><strong>Overview</strong>: This paper introduces <strong>score matching</strong>, a generative modeling approach. The key idea is to approximate the <strong>gradient of the data distribution’s log-likelihood (the score) </strong>with a neural network. Once the score function is learned, new data can be sampled using <strong>Langevin dynamics</strong>, which iteratively refines samples by following noisy gradients. While not initially related to diffusion models, this paper lays the groundwork for later advancements by bridging the gap between data gradients and probabilistic modeling. It provides valuable intuition about what generating an image in pixel space means and how it can be conceptualized as a stochastic trajectory.</li><li><strong>Key Takeaway</strong>: Sampling is seen as a stochastic process in the image space.</li><li><strong>Bonus Paper </strong><a href="https://arxiv.org/abs/2011.13456"><em>Score-Based Generative Modeling through Stochastic Differential Equations</em></a><em> (2020)</em>: it generalizes score matching to the use of stochastic differential equations for sampling. It provides further intuition on how score-based models work.</li></ul><h4>3. <a href="https://arxiv.org/abs/2006.11239">Denoising Diffusion Probabilistic Models</a> (2020)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qohbTVwlLIQtU8Dl3pbW8Q.png" /></figure><ul><li><strong>Overview</strong>: This is the <strong>definitive paper</strong> on diffusion models. Building on earlier probabilistic and score-matching insights, it introduces a simple yet powerful framework for training diffusion models. Some of the main contributions include:<br>- <strong>Simplified training objective</strong>: Clever mathematical derivations enable training a neural network to predict noise at each diffusion step.<br>- <strong>High-quality image generation</strong>: Demonstrates that diffusion models can produce state-of-the-art images with a tailored architecture.<br>- <strong>Connection to score matching</strong>: Establishes that the sampling process is equivalent to Langevin dynamics in score-based models. This allows us to easily interpret what we are learning with diffusion approaches.<br>This paper shifted the field’s perspective by showing that diffusion models could achieve high-quality image generation. It serves as the framework for how diffusion models are trained today.</li><li><strong>Key Takeaway</strong>: If you had to choose only one paper to read, take this<strong>.</strong></li><li><strong>Bonus Paper </strong><a href="https://arxiv.org/abs/2102.09672"><em>Improved Denoising Diffusion Probabilistic Models</em></a><em> (2021)</em>: it introduces optimizations and enhancements to diffusion, including improved noise schedules and sampling techniques.</li></ul><h4>4. <a href="https://arxiv.org/abs/2010.02502"><strong>Denoising Diffusion Implicit Models</strong></a><strong> (2020)</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QGh42N7qoO5qZrhU1Ngk3Q.png" /></figure><ul><li><strong>Overview</strong>: One major drawback of standard diffusion models is the slow inference time: the reverse process requires <em>T</em> denoising steps, where <em>T</em> is often in the hundreds or thousands. This makes generating thousands of images a time-intensive task, taking hours or even days.<br>This paper addresses this inefficiency by introducing a <strong>non-Markovian reverse process</strong>, where dependencies across multiple steps are considered. This innovation allows for <strong>sampling in significantly fewer steps without compromising output quality</strong>. Specifically, the model can produce high-quality samples similar to those generated with <em>T</em>=1000 steps but in just 20 to 100 steps.</li><li><strong>Key Takeaway</strong>: Speed-up in sampling using non-Markovianity.</li></ul><h4>5. <a href="https://arxiv.org/abs/2105.05233">Diffusion Models Beat GANs on Image Synthesis</a> (2021)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1021/1*3m0aiwLbLdu_3faGI_Y6Zw.png" /></figure><ul><li><strong>Overview</strong>: This paper shows that, by developing an <strong>improved architecture</strong>, not only diffusion models can produce high-quality images, but also <strong>state-of-the-art (SOTA) quality</strong>. Another contribution of this paper is in the output guidance. Typically, diffusion models offer no control over the generated content, meaning we cannot direct the model to produce specific types of images. To address this limitation, the authors introduce <strong>classifier guidance</strong>: a classifier is trained on noisy images _x_t to predict their class. During sampling, the classifier’s gradients guide the reverse diffusion process, steering image generation toward a desired class.</li><li><strong>Key Takeaway</strong>: Classifier guidance for controllable image generation and SOTA results.</li></ul><h4>6. <a href="https://arxiv.org/abs/2207.12598">Classifier-Free Diffusion Guidance</a> (2022)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J2B8TzkRHlHwx4bu1aZdSg.png" /></figure><ul><li><strong>Overview</strong>: This paper addresses a key limitation of classifier-guided diffusion models: reliance on an external classifier. The authors propose <strong>classifier-free guidance</strong>, where a single model learns both conditional and unconditional score estimates. This is achieved by randomly discarding class information during training, allowing the model to generate diverse, high-quality images without needing an external classifier. The result is a simpler and more efficient approach that outperforms classifier guidance in both quality and diversity.</li><li><strong>Key Takeaway</strong>: Classifier-free guidance.</li><li><strong>Bonus Paper </strong><a href="https://arxiv.org/abs/2112.10741"><em>GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models</em></a><em> (2022)</em>: it compares classifier-free guidance and <a href="https://arxiv.org/abs/2103.00020">CLIP</a> guidance. Even though CLIP guidance does not achieve the best results, this paper marks <strong>the beginning of the diffusion-based text-to-image generation era</strong>.</li></ul><h4>7. <a href="https://arxiv.org/abs/2106.15282">Cascaded Diffusion Models for High Fidelity Image Generation</a> (2021)</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Gwodrq7MVDg2O4JVMDZ9UA.png" /></figure><ul><li><strong>Overview</strong>: The UNet architecture used in standard diffusion models is hard to scale to high resolutions. So the question is <strong>how can we scale diffusion models to higher resolutions in a simple way? </strong>The authors solve this by proposing a pipeline of <strong>cascaded diffusion models</strong>, where each model operates at a progressively higher resolution. The method leverages the ease of conditioning diffusion models on input images, with each super-resolution step simply concatenating its output to the next stage. Cascaded diffusion models offer a scalable and efficient solution for generating high-resolution images without the computational overhead of a single massive model.</li><li><strong>Key Takeaway</strong>: Simple technique to generate high-resolution images.</li></ul><h4>8. <a href="https://arxiv.org/abs/2112.10752"><strong>High-Resolution Image Synthesis with Latent Diffusion Models</strong></a><strong> (2021)</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/865/1*BFm5SWfB_09Zb_iPwpm9GA.png" /></figure><ul><li><strong>Overview</strong>: This is the <strong>paper that gave birth to Stable Diffusion</strong>, a significant milestone in generative AI. Standard diffusion models operate on the full pixel space, which is computationally expensive and inefficient. The key question this paper addresses is: <strong>how can we make diffusion models computationally efficient while maintaining high-quality output? </strong>The solution lies in encoding images into a <strong>lower-dimensional latent space</strong>, performing diffusion in this latent space, and then decoding the result back into the original pixel space. This approach works because the autoencoder captures the perceptual details of the image, while the diffusion process focuses on the semantic aspects. The result is a computationally efficient model that produces high-quality images. Another major contribution is the integration of <strong>cross-attention</strong> during the latent reverse process, enabling the model to handle tasks conditioned on text, images, classes, etc.</li><li><strong>Key Takeaway</strong>: Diffusion occurs in a learned latent space, drastically improving computational efficiency while maintaining high image quality.</li></ul><h3>To conclude…</h3><p>I claim that if you <strong>study </strong>these papers in the following order, you will have an easier life in understanding diffusion models. But let’s not kid ourselves: this won’t make you an expert at all. Why? Two big reasons:</p><ul><li><strong>We’ve only talked about images</strong>: Diffusion models are great, but they’re <strong>not </strong>just for images. They also work wonders on video or audio; so, if you think you’re done after mastering the image stuff, nope — there’s more homework.</li><li><strong>Diffusion models aren’t the whole story</strong>: Generative AI has a family, and diffusion models are just one cool sibling. To get the full picture, you’ll also need to hang out with <strong>GANs</strong>, <strong>VAEs</strong>, and <strong>flows</strong>. For example, understanding the architecture of certain diffusion models often requires knowledge of GANs or VAEs, while grasping the latest advancements in Stable Diffusion will involve concepts from flows.</li></ul><p>In general, diffusion probabilistic models for images are not the end of the story, but in my opinion, they are the best starting point. So, if you haven’t already clicked those links and read the papers — <strong>WHAT ARE YOU WAITING FOR?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/0*mdhdhZbJfZOklVNK.gif" /></figure><p><em>Did I miss any essential papers? Or did I write something that doesn’t quite make sense? Let me know in the comments! Your feedback makes this resource better for everyone!</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=545a3639ca3c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Some Random Thoughts on AI]]></title>
            <link>https://aiadvances.org/some-random-thoughts-on-ai-d14db44c055c?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/d14db44c055c</guid>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[intelligence]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-alignment-and-safety]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Wed, 29 Nov 2023 20:15:15 GMT</pubDate>
            <atom:updated>2023-11-29T20:15:15.810Z</atom:updated>
            <content:encoded><![CDATA[<h4>One Year Later ChatGPT’s Release</h4><figure><img alt="A surreal landscape of artificial neurons, pulsing and intertwining to create a powerful AI entity." src="https://cdn-images-1.medium.com/max/1024/1*yguphO23Whmmk4-lfUsS0w.jpeg" /><figcaption>Image by Author via Leonardo.Ai</figcaption></figure><p>On November 30th, 2022 OpenAI silently released a chatbot called ChatGPT. Fast forward one year and the field of artificial intelligence (AI) is advancing at an ever-increasing rate, with continuous excitement, problems, and innovations.</p><p>As an AI student myself, experiencing the front line of such a big phase transition of society is amazing and many (but like… <em>many many many</em>) opinions, events, and debates have caught my attention. So, after one year of being in the eye of the hurricane, I decided to collect <strong>some</strong> thoughts I heard or developed.</p><h4>Contents</h4><p><em>· </em><a href="#ccc8"><em>Data Is AI’s Fuel.</em></a><em><br>· </em><a href="#f9fe"><em>The Brain and Neural Networks Are the Same, but Not Really.</em></a><em><br>· </em><a href="#c950"><em>NNs Are One Step Ahead Because They Predict the Future. And We Do Not.</em></a><em><br>· </em><a href="#de86"><em>We’re Not Alone.</em></a><em><br>· </em><a href="#dfae"><em>Intelligence Is Like Astronomy.</em></a><em><br>· </em><a href="#ff61"><em>AI Is an Engineering Science.</em></a><em><br>· </em><a href="#c836"><em>Next, We Should Zoom In.</em></a><em><br>· </em><a href="#4f96"><em>Companies, Open Source, and Alignment</em></a><em><br>· </em><a href="#c85d"><em>There’s More to Language.</em></a></p><p>Each view is briefly summarized in a single point. The comments are short — even though they could be worth a separate article each — and not in any particular order. <strong>So, feel free to read whatever you like. </strong>Unfortunately, they are just a tiny fraction of what is truly discussed, but it’s still better than nothing to spark some reflections.</p><h3>Data Is AI’s Fuel.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4qInKGkT2LL5lphc" /><figcaption>Photo by <a href="https://unsplash.com/@peter_forster?utm_source=medium&amp;utm_medium=referral">Peter Forster</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>AI is like the Industrial Revolution, not because of the shift in society it will lead to, but because both explosions relied on <em>something else</em>. Here is what I mean: even though the Industrial Revolution spiked so much development in power, energy, electricity, etc., the <em>boom</em> was all possible because there were a bunch of dead dinosaurs lying underground. That is to say: “<em>Fossil fuels fueled the Industrial Revolution</em>”. Analogously, we are talking about an AI revolution because we have lots and lots of <strong>data</strong>: decades of Internet expansion first, and social media then, tremendously helped in feeding neural networks (NNs) with their favorite food.</p><p>Others could argue that AI’s fuel is the computational power, the algorithm, or the architecture, but I would just say:</p><blockquote>“<em>Nah… it’s the data</em>”.</blockquote><h3>The Brain and Neural Networks Are the Same, but Not Really.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_SGFwujYvDDyv5Cl" /><figcaption>Photo by <a href="https://unsplash.com/@artic_studios?utm_source=medium&amp;utm_medium=referral">Daniel Öberg</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Historically, NNs exist as a proposed copy-paste of our brain. In general, the influence that neuroscience had and has on AI is hardly quantifiable because it is huge. Regardless of the inspiration, artificial networks and brain networks turned out to <strong>diverge in their essence</strong>. As Hinton says, the brain does not seem to heavily implement anything like backpropagation in its internal learning system as NNs do. This is interesting because with way fewer parameters than those in our brains, modern large language models (LLMs) have much more knowledge than the average human. So could backpropagation be a new learning paradigm? At the same time, NNs require high amounts of training data with current architectures; instead, humans only need a few training examples to generalize well enough (look at kids). However, NNs are extremely <em>energy </em>inefficient whereas our brains are very efficient. The story could go on forever.</p><p>Probably NN learning systems would have never evolved naturally because of their inefficiency. Indeed, nature wants to minimize energy and the human brain has adhered to that, in a process that took hundreds of thousands of years. If I were to draw a line between the brain and the NNs, I would say the following: humans have <em>energy efficiency</em> encoded in their essence, but lose in terms of computational capabilities; on the contrary, NNs are way less energy efficient but have an unreachable computational power. Could this be a reasonable <strong>basic </strong>difference?</p><h3>NNs Are One Step Ahead Because They Predict the Future. And We Do Not.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xjQXjVME3SZO2JUs" /><figcaption>Photo by <a href="https://unsplash.com/@helloimnik?utm_source=medium&amp;utm_medium=referral">Nik</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The ease with which we can talk about the past is proof of how difficult it is to predict the future. We, humans, simply cannot predict it: there is so much to keep into account, right? Neural nets instead, only look at a loss function, and their computational power makes it <em>easy</em> to make predictions. An <a href="https://www.indiatoday.in/trending-news/story/this-robot-is-built-by-ai-and-can-walk-as-well-video-is-viral-2448844-2023-10-14">experiment at the University of Illinois</a> has shown that an AI program required only 26 seconds “to conceptualize and design [a] walking robot”. <strong>Evolution taught us how to walk in millions of years, not in 26 seconds</strong>. I believe that is something very interesting that deserves some attention.</p><h3>We’re Not Alone.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*_36jv-uNtlVxd1PC" /><figcaption>Photo by <a href="https://unsplash.com/@possessedphotography?utm_source=medium&amp;utm_medium=referral">Possessed Photography</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>It is the first time we can seriously argue about the existence of a more intelligent <em>thing</em>. We’re not alone, anymore! Hold on. I am not saying that AIs are intelligent or sentient, but people have come to the realization that there is something weird going on. We do not know what, why, or how, yet the <strong>emerging patterns in advanced LLMs clearly resemble intelligent human beings</strong>, whether you like it or not. The fact that the closest intelligent being is not an ugly rat anymore, but an algorithm that we can easily code on our computers, is super cool.</p><p>This is just the beginning of AGIs and our future relationship with them will depend a lot on us. Just like Noah’s Ark included all the animals that were to be saved, <em>we </em>will be in charge of deciding where AGI will have the right to jump on <em>our </em>boat.</p><h3>Intelligence Is Like Astronomy.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kNi-ck5wxGrfZSaR" /><figcaption>Photo by <a href="https://unsplash.com/@grakozy?utm_source=medium&amp;utm_medium=referral">Greg Rakozy</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>People have been studying astronomy for millennia, even from the Babylonians. They all started by looking up in the sky and wondering: <em>what is going on outside</em>? They started it to understand where we really were. Century by century, knowledge about the sky grew, until at some point we understood how to properly study stars with Galileo and the scientific method. After that moment, everything went downhill and we conquered the sky, literally.</p><p>In a similar fashion, we study intelligence to answer the question: <em>what is going on inside</em>? We want to understand our brain to understand who we really are. <strong>Right now, we are like the Babylonians</strong>. Surely, the scientific progress we currently have will make it unlikely to wait for centuries until we reach a paradigm shift, but the main idea still holds: we are just starting out and we will need time to truly have a theory of mind.</p><p>Maybe neuroscience will become useless to AI and, instead, AI will serve as a tool to study human intelligence. After all, there’s plenty of evidence that suggests the representations learned by the AI (both in vision and language) are very similar to those of the brain, so there must be something going on we don’t understand yet. Maybe the scientific paradigm we’re adopting is ill-posed for the study of human intelligence. Could AI be the new paradigm?</p><h3>AI Is an Engineering Science.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*QQI6hhKnyRomDKDm" /><figcaption>Photo by <a href="https://unsplash.com/@roblambertjr?utm_source=medium&amp;utm_medium=referral">Rob Lambert</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>AI moves fast. Its complexity and changing objectives make it hard to define a theory that, for example, can be as good as that of physics. So the <strong>best approach to AI development is simply experimenting</strong>, by first defining the problem and then solving it with the seemingly best approach in an iterative way (i.e. trial and error). Once that is established, the underpinning theory can then be rigorously defined.</p><p>Notice that there is nothing wrong with this approach. Talking about the Industrial Revolution again, Volta discovered the battery and nothing was known about electricity. Decades went on developing electric motors, generators, and telegraphs until Maxwell <em>et al. </em>developed a theory that fueled more knowledge and later more discoveries.</p><p>AI is currently at this point: lots of applications but not much theory. So if you have an idea about something, the best approach in AI is to <strong>just do it</strong>.</p><h3>Next, We Should Zoom In.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*wx06Yu0amvDlq-DL" /><figcaption>Photo by <a href="https://unsplash.com/@blueguy590?utm_source=medium&amp;utm_medium=referral">Jonathan Mabey</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Even though we may not be able to define a general theory off the bat, we can still search for some results by zooming in. Indeed, many big revolutions in science occurred when zooming in: cells, DNA, atoms, etc. What if the next transition in AI occurs by doing the same?</p><p>Now, a sure thing is that we know what the collective behavior of neurons produces (i.e. the output given the trained weights), but <strong>we do not have much theory into the relevance of each individual neuron in the learning process</strong>. A lot of creative works could emerge from here. In particular, analyzing the “behavior” of individual neurons, features, and circuits could be key in interpreting emerging phenomena in future advanced neural nets. Probably many tools from other disciplines, such as statistical physics or psychophysics, could end up beign useful, as the complexity of the problems will arise over time.</p><h3>Companies, Open Source, and Alignment</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yqBSz2nkK0yW0upM" /><figcaption>Photo by <a href="https://unsplash.com/@lukesouthern?utm_source=medium&amp;utm_medium=referral">Luke Southern</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The big competition between OpenAI, DeepMind, Anthropic, and other labs is fueling a lot of progress. Unfortunately, many state-of-the-art models are not released to the public. On the one hand, the community is penalized because more contributions could lead to huge leaps in a shorter time; on the other hand, companies are trying to prevent the models from being used by bad agents. Personally, I’d rather have the models to be private than open source. As a matter of fact, AI has the potential to bring many societal benefits but also be <strong>harmful and catastrophic if not handled by the right people with the right knowledge</strong>. So let’s keep them private for now.</p><p>That is why two further things are fundamental: government regulations and alignment. As for the former, it is crucial to adopt the right laws to prevent dangerous actors from intervening, now more than ever. As for the latter, there are many arising issues in the ethical uses of LLMs and other AI models in general. Thus, seeking a way to properly benchmark and tune the systems to align them with our values is crucial in order to construct the next generation of <em>safe</em> models.</p><h3>There’s More to Language.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2V9Je_XG_r35IZak" /><figcaption>Photo by <a href="https://unsplash.com/@flipboo?utm_source=medium&amp;utm_medium=referral">Philippe Bout</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I’ll be honest: there has been a boring (and at times annoying) <strong>over-indexing on LLMs</strong> during the last year. Yes, they are important and will be crucial, as most likely AGIs will need to produce and interpret language, but that should not be the entire focus. Currently, we have <em>intelligent</em> systems with language only, but human intelligence was developed by actively refining the senses and interactions with the environment. <em>So it was also about what we saw and did, not only what we said.</em> In today’s AI research, I believe that more focus should be put on vision systems, both in state-of-the-art models and alignment issues. There is a lot more to learn from pictures and videos than text because data is much more redundant and has more internal structure to learn from.</p><p>As a concluding note, AIs will become truly dangerous once self-aware in a robot. In a more proper way, the interaction between multi-modal (language, vision), and sensory-motor control models in robotics could be the defining characteristics of a serious AGI. If today we are focusing on existential risk on LLMs in chemistry, biology, etc. just for precautionary reasons, then, by analogy, we should do the same for what could be potentially more hazardous.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9p04Yl7Oze3L6Wc0kIAOCw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d14db44c055c" width="1" height="1" alt=""><hr><p><a href="https://aiadvances.org/some-random-thoughts-on-ai-d14db44c055c">Some Random Thoughts on AI</a> was originally published in <a href="https://aiadvances.org/">AI Advances</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Assisted, not Automated: Striking a Balance in Decision-Making with AI]]></title>
            <link>https://medium.com/predict/assisted-not-automated-striking-a-balance-in-decision-making-with-ai-5c7a87578f19?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/5c7a87578f19</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[automated-decision-making]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ethics]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Sun, 28 May 2023 18:51:24 GMT</pubDate>
            <atom:updated>2023-05-28T18:51:24.770Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A scale" src="https://cdn-images-1.medium.com/max/1024/1*0oxM3hn9DPHPXJ2RJr0Bnw.png" /><figcaption>Credits: Author via DALL-E</figcaption></figure><p>From Tesla’s self-driving cars to Boston Dynamics’ agile robots, from facial detection to image generators, from social media suggestions to map navigation systems, the impact that artificial intelligence (AI) has had in our lives has steadily increased in the last few years. Just like any other disruptive innovation, sooner or later AI will reach a connection with any imaginable aspect of our lives. One of these has been emerging lately and comes under the name of Automated Decision Making (ADS). ADS refers to the use of algorithms or other automated processes to make decisions without human intervention. In other words, ADS involves using technology to analyze data and arrive at a decision, without a person having to manually review the information themselves to take the decision.</p><h4>The Growing Influence of Algorithms in Decision-Making</h4><p>If this sounds futuristic and fictional, it is. One of the earliest examples of automated decision-making can be found in Isaac Asimov’s “I, Robot”, published back in 1950, in which many of the characters are robots and they are responsible for the decision-making of mankind. Today, there are countless many examples of why we are seeing this fiction become a reality. One basic and easy-to-understand example is credit scoring. Indeed, usually, lenders need to assess a borrower’s creditworthiness and decide whether to approve or deny their loan application; hence the decision maker is the lender and must choose whether to give the money or not. Automation comes into play here since an algorithm considers a variety of factors, including the borrower’s credit history, income, and debt-to-income ratio to reach a final decision. In a fast environment such as that of banks, automation allows lenders to quickly decide without being confused about the choice, leading to more efficient and consistent decisions. Another example is job recruitment where, in the first screening phases, algorithms consider education, work experience, skills, and other proxies to rank candidates based on their suitability for the role, which surely saves time and resources.</p><p>Other examples exist as well, yet what I believe is more worth mentioning is our common reaction to these facts: it feels strange to have decisions taken by an algorithm without human intervention. We are used to making decisions based on personal (and human) intuitions and judgments, which can consider several factors that are rarely mathematically quantifiable. In addition, the algorithms’ accuracy tends to be not perfect and is destined to make errors by construction. Thus, the following questions arise: is it fair and ethical to entrust automated decision-making to determine our fate? Are these decisions correct? What does “correct” even mean for a machine: is it risk minimization, or social acceptance maximization? What makes them more reliable than humans to fully base a decision on AI? These and others are critical questions that need to be addressed to fully comprehend the impact that ADS will have on society, in this era of unprecedented advancements.</p><p>In particular, the central vision that I believe in, at this level of technological progress, is clear: ADS should not stand for automated decision-making, but rather for assisted decision-making. While AI can be a powerful tool to support and enhance human judgment, we need to recognize that it cannot replace it entirely, and relying solely on algorithmic decisions can be extremely limiting for various reasons.</p><h4>Unveiling Automated Decision Making</h4><p>First of all, AI algorithms are only as good as the data they are fed. This means that if the data is biased, then the algorithm will be too, so using ADS, without having a clear picture of the information that the system is given, may lead to bad consequences. A clear example of what I think of comes from Redlining, a common practice adopted by American banks in the 20s. Bear in mind that even though it is not AI, the techniques used were very statistical and automated. The task was similar to that of the credit example quoted initially, namely: if you are a bank and you want to give a house loan or some insurance for houses, how likely is it that the person will pay and not get into default? What they did was analyze how the credit in a specific area was, so lower scores were assigned to areas that tended to not repay the debts. But as one might easily infer, inner-city neighborhoods, where lots of Afro-Americans lived, had a fewer score than rich neighborhoods; hence, in the long run, districts with a positive score became richer, more educated, and wealthier because they were given the money, while poor neighborhoods were impoverished even more. This shows how an initial bias over the data propagated significantly over the years because the decisions were only based on the output of data, without properly intervening with human oversight whenever needed. Similar cases can be applied to other examples, like crime prevention, in which data on crime patterns, demographics, and high-risk areas are gathered to allocate resources (like police patrols) more effectively. In fact, recently, an ADS algorithm used to predict recidivism rates has been found to be inaccurate and biased, leading to unjustified sentencing decisions. Hence, in a world where data is extremely messy and unclean, relying solely on ADS may lead to unwanted biases and unaccounted results.</p><p>Secondly, the lack of accountability increases the risk of trusting AI. Most of the decision-making tasks are so complex that black-box algorithms are used for them, meaning that even though they might make the choice, there is missing transparency in how the final decision was made, making it harder to trust the system. In addition, the variables that the algorithms might consider to output a decision may be unknown to humans. For example, a trading algorithm may make decisions based on market data that humans are not aware of, leading to unexpected market behavior. As this was not enough, when the mistake the algorithm makes is so big that the consequences become serious, current AI regulations do not provide guidance on the responsibility, making it harder to hold individuals or organizations accountable for their actions.</p><p>However, all of this does not imply that automated decision-making should be considered unethical a priori. Indeed, whenever a task is repeatable, can be made more efficient, and has an objectively correct choice, then ADS can be the right way to go. For example, consider passport scanning. By doing face recognition, airports can efficiently detect faces to match them with passport pictures and allow the passengers to avoid wasting time in boring procedures. Here the task is repeatable for sure and has a true answer (i.e., the face), so ADS is surely advised. Another example is in agriculture, where ADS can be used to optimize crop yields and reduce waste. Here the decision is clear, repeatable, and unique, and the farmer just must follow what the algorithm says, as to have the best results, because everything is well-defined for the algorithm to interpret. But when a decision becomes non-unique, that is when the A should stand for assisted, and not automated. To give an extreme example, self-driving cars belong to this category. Yes, they would reduce incidents on the street, improve overall driving quality, etc. as everyone says, yet, very simply put, the car’s decision-making should never be fully put in the hands of an algorithm but rather in the hands of the driver only. There are some ethical scenarios that we cannot fully agree on ourselves (MIT’s Moral Machine is an example of this), let alone a car that is supposed to be trained on human data and do the decision for us!</p><h4>Balancing Ethics and Efficiency</h4><p>In general, current ADS lacks the ability to understand complex human emotions and social contexts. This can be particularly problematic in cases where decisions are subjective, and intuition and empathy are essential. Again, the Moral Machine serves as a great example, but also the initial criminal justice instance can be useful to explain this. When the jury has to sentence a person, many factors like the offender’s personal history, mental health, struggles, etc. may fail to be understood by algorithms and should be judged upon human empathy and intuition. Still, it cannot be denied that when a low-risk and repeatable task can be automated, most people can benefit from it. This is why ADS should mean assisted decision-making and not automated decision-making. The stream of AI progress that we are undergoing right now is beneficial in the realm of decision-making, both mathematically and practically, as previous examples illustrate, but over-relying on those instruments will lead to a lack of independent judgment that we have seen to be fundamental in the decision process. Automation should function as a supporting aid to decision-making, rather than a replacement that fully automates it.</p><p>Finally, ADS has become an increasingly common reality in several fields, some of which were briefly aforementioned. While artificial intelligence can be an extremely powerful tool to support and enhance human judgment, it should not replace it entirely. For this reason, a more appropriate meaning for ADS should be assisted decision-making, not automated, where AI assists humans in their decision processes, still leaving the final choice to mankind. This is crucial given the potential biases that algorithms can inherit from the data they are given and problems with fairness, accuracy, and reliability that can limit trust in these systems. Thus, ensuring ethical and social values in the use of ADS is essential to avoid perpetuating social injustices. To conclude, as AI advances it becomes increasingly critical to address these ethical and moral challenges to strike a balance in the responsible usage of ADS, with humans at its core, not algorithms.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5c7a87578f19" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/assisted-not-automated-striking-a-balance-in-decision-making-with-ai-5c7a87578f19">Assisted, not Automated: Striking a Balance in Decision-Making with AI</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Killer Robots: The True Danger of AI]]></title>
            <link>https://medium.com/@sandromikautadze/killer-robots-the-true-danger-of-ai-52c924e6cd7?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/52c924e6cd7</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[future]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[war]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Tue, 11 Oct 2022 20:54:24 GMT</pubDate>
            <atom:updated>2022-10-11T20:54:24.343Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="A stack of missiles in a grey, foggy environment" src="https://cdn-images-1.medium.com/max/1024/1*4y3AtIMc6jSETyi4thJi4g.png" /><figcaption>“The danger of weapons”, AI-generated image by author.</figcaption></figure><p>Open Google Images and type «Artificial Intelligence». If you did, there’s a very high chance that you got pictures similar to these:</p><figure><img alt="Google Images on AI" src="https://cdn-images-1.medium.com/max/1024/1*QEdGSl0ryTR5cZUXr_o2aQ.png" /><figcaption>“Artificial Intelligence” on google.it</figcaption></figure><p>That is “artificial intelligence” for Google. Or, perhaps, for us? Think about it: our view is largely shaped by popular culture, mass media, movies, books, etc. in which the term AI is associated with an almighty-human-thinking robot.</p><p>No wonder one of the most undertaken fears is that self-aware AIs will gain control over mankind! Let me tell you an astonishing truth: that’s all baloney. Yet, in a sense, <strong>this phobia is not completely untrue</strong>. I’d say it is misguided, instead. Why? Because we should actually fear AI, not due to thinking robots, but due to the so-called <strong>killer robots</strong>.</p><h3>What’s a Killer Robot?</h3><p>Imagine an AI that recognizes people and targets them; imagine putting that piece of software into a weapon.. <em>et voilà.</em> <strong>This is how you get AI weapons</strong>, aka “autonomous weapons systems” (AWS) for the experts, aka “killer robots” for the mass media.</p><p>In even simpler words, they are dumb machines using the brain of a clever AI algorithm, created with the only purpose to kill without human intervention. They do not need some crazy person clicking a giant red button to kill the target. When the AI recognizes the target, it fires — automatically.</p><h3>How are they even a real thing?</h3><p>I got you. After all, the only encounter we might have had with AWS may have been watching Terminator or other sci-fi movies. We all know science fiction is just a lie, so why even be bothered by these “killer robots”?</p><p>However, I’m here to tell you another astonishing truth: autonomous weapons are not an issue of sci-fi; <strong>they are a problem of the present. </strong>If you are skeptical about their impact, I’ll try to give you a <em>glimpse </em>into it.</p><p>In 2016 <a href="https://www.engadget.com/2016-06-28-combat-ai-beats-air-force-experts.html">AI has beaten the Air Force’s top tactical experts in air combat</a>, and with a relatively simple algorithm. In 2020, <a href="https://www.alliedmarketresearch.com/autonomous-weapons-market-A13132">Azerbaijan used drones that could autonomously fly over an area and divebomb enemy radar signals to attack Armenian-occupied territory</a>. In 2021, <a href="https://www.newscientist.com/article/2282656-israel-used-worlds-first-ai-guided-combat-drone-swarm-in-gaza-attacks/">Israel used the world’s first AI-guided combat drone swarm in the Gaza attacks</a>. In Russia’s invasion of Ukraine, AI has been used to <a href="https://www.wired.com/story/russia-ukraine-war-ai-surveillance/?utm_campaign=The%20Batch&amp;utm_source=hs_email&amp;utm_medium=email&amp;_hsenc=p2ANqtz-9EOm1d3kC_SFPUH7mhgXpGX7WuSvgIJFdj43EHpj8Bcc89B70rO-fg77VV5gF_cIgtEWkk">intercept communications</a>, and <a href="https://www.wired.com/story/facial-recognition-identify-russian-soldiers/?utm_campaign=The%20Batch&amp;utm_medium=email&amp;_hsmi=221023141&amp;_hsenc=p2ANqtz-_VEjULQ5KVsXL6DyPsPTSMD3bhRJsl5tahljQLBFWg3Tf3O1Xbo3S6BUe5mm4tw5rztAtLc3OvelapSLJWjhod4PXCOE9AerMc-WlKvt-pk4u350Y&amp;utm_content=221015292&amp;utm_source=hs_email">recognize soldiers</a> —i.e. AI has been exploited for strategic and tactical advantages in the military!</p><p>In addition, the US alone budgeted $18 billion for autonomous weapons between 2016 and 2020. They also <a href="https://www.militaryaerospace.com/computers/article/14204595/artificial-intelligence-ai-dod-budget-machine-learning?utm_campaign=The+Batch&amp;utm_medium=email&amp;_hsmi=221023141&amp;_hsenc=p2ANqtz-8mNhVM0b2yDh7nI6D5pWNewtmVbMiPL_AF_RqoQYlyfEWfko_zcbXAEFnSbFviQR2S4nvoLd2OHL7WA2aBcQvrzDuPhr8cpeFaQbBj6d-ErZw7FhE&amp;utm_content=221015292&amp;utm_source=hs_email">requested almost $900 million</a> in AI R&amp;D in the military for 2022. Indo-Pacific allies recently <a href="https://ipdefenseforum.com/2021/11/indo-pacific-allies-investing-in-ai-technology-for-military-application/?utm_campaign=The%20Batch&amp;utm_medium=email&amp;_hsmi=221023141&amp;_hsenc=p2ANqtz-9ExJRWTeOvW58uK1Hez-Rtv9psiBcIVCuse0d3BmRD0RcVkZseYPxEokvRZB2NRwRl9-LT4rKlb5u4xkRPfY88pUEw1PZiS9amTrn9nE5SB0M9eWk&amp;utm_content=221015292&amp;utm_source=hs_email">started investing</a> in the same matter. In June, NATO announced a €1 billion fund for the development of technologies, among which (<em>guess what</em>) AI, for defense challenges. In June, the UK <a href="https://www.gov.uk/government/publications/defence-artificial-intelligence-strategy/defence-artificial-intelligence-strategy?utm_campaign=The%20Batch&amp;utm_medium=email&amp;_hsmi=221023141&amp;_hsenc=p2ANqtz-93GeIkFKIHRv92gVd4m6ub9MM-hnP9DVoZ17Dz1VGpWv3kW0wnEZMKU8FVWWEyXEiUUinQr1r_akhltftZE2VZcSRz8Dn2Pkc2hm2fi2VY-S_Foac&amp;utm_content=221015292&amp;utm_source=hs_email">devised new plans</a> for AI on defense strategies.</p><p><em>DO I HAVE TO GO ON?!</em></p><p>It seems like science fiction is a lie that tells the truth. What is the Truth about AI weapons? <strong>They are considered the third revolution in warfare</strong>, after gunpowder and nuclear weapons.</p><h3><strong>Houston, we have a problem!</strong></h3><p>Or perhaps more than one. The issues that derive from AWS are countless.</p><p>Consider <strong>unpredictability</strong>. AI algorithms are for the most part black boxes. Even experts in the field sometimes lack a fundamental understanding of the algorithms’ decision-making. This is just a fancy way of saying that predicting the behavior of such weapons could be extremely difficult.</p><p>Consider an <strong>AI weapon killing the wrong targets</strong>. Who is responsible for the mistake? The country? The programmer? The producer of the weapon? Nobody?</p><p>Consider the <strong>economy</strong>. These weapons might increase productivity while lowering the cost of production (<em>it’s just code</em>). What does this mean? Mass-production of autonomous weapons.</p><p>All of this, added to the rapid spread of AI, has made huge communities worry a lot. For example, in 2015 AI researchers wrote an <a href="https://futureoflife.org/2016/02/09/open-letter-autonomous-weapons-ai-robotics/?cn-reloaded=1">open letter</a> to warn against the use of autonomous weapons. To date, the letter has been signed by over 30k people, among which Elon Musk, Stephen Hawking, and some of the true pioneers of AI, like Stuart Russell, Nils Nilsson, and me.</p><p><em>(have a laugh)</em></p><p>Other activist movements are beginning to grow in these years. People are starting to get aware of this issue.</p><h3>Any pros?</h3><p>Whether you like it or not, <strong>AI will touch everything the military does</strong>. But this is not that bad.</p><p>Think about autonomous defense systems, driverless convoys delivering logistical supplies, and drones to leave medical aid to specific wounded targets. These are all pros!</p><p>Yet, not everyone is going to hold themselves to the same set of values. When China and Russia start deploying autonomous vehicles that can kill people, what is the move that others are going to make? Will there be a literal AI arms race?</p><p>And just to keep the momentum alive, <strong>the US, Russia, and China have recently said that a ban on AI weapons is unnecessary! </strong>That’s funny, isn’t it? I want to stress that it was not said by countries in the third world; it was said by countries that want to rule the world.</p><figure><img alt="Biden and Putin laughing" src="https://cdn-images-1.medium.com/max/1024/1*xZff6CI14r3qIew_CRzEfg.png" /><figcaption>“Biden and Putin laughing at the rest of the world”. AI-generated image by author¹</figcaption></figure><p>This tells us three things to deem possibly true:</p><ul><li>A ban is necessary.</li><li>An AI arms race is already going on.</li><li>Whichever country will have the best AI will probably achieve dominance on this planet.</li></ul><h3>To summarise…</h3><p>Mass media wants to convince us that the future of AI will be full of sentient human robots. That’s just sci-fi: machines do not think. But machines do what we ask and in recent years, we started asking those machines to kill people. <strong>This is not sci-fi anymore</strong>.</p><p>Autonomous weapons are a present danger that is getting increasingly big faster than ever before. Countries will need to come to agreements on the development of these technologies as soon as possible, otherwise, an AI arms race will be more inevitable than it already is.</p><p>Sure, there are advantages to the use of AI for the military, but if we stand still watching AI’s future unfold on its own, autonomous weapons could kill large amounts of people. Just like a factory machine can work on products day and night, autonomous weapons could kill 24/7. Not only that, AWS will spread across the globe very fast. Eventually, will we be able to shut off a system already distributed everywhere?</p><p>To conclude, as of now AI is surely an amazing tool. But being something completely new, there are some roads that are heading off the cliff. We, humans, must make a lot of effort so that we don’t take the human race off the cliff.</p><h4>Extras (optional)</h4><ul><li>[1] I wanted to put Jinping too, but MidJourney (the generator) told me that “the word jinping is banned”. Fair enough…</li><li>This is a topic worthy of a Ph.D. thesis. I claim not to be an expert on AWS in any way. If you spot mistakes, let me know.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52c924e6cd7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why DeepMind Is the (God)father of AI Innovation]]></title>
            <link>https://medium.com/@sandromikautadze/why-deepmind-is-the-god-father-of-ai-innovation-e54bb0148ed7?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/e54bb0148ed7</guid>
            <category><![CDATA[innovation]]></category>
            <category><![CDATA[ml-so-good]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Thu, 18 Aug 2022 12:16:20 GMT</pubDate>
            <atom:updated>2022-08-18T14:04:37.617Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yPvTSrA-YFBxyNzx" /><figcaption>Photo by <a href="https://unsplash.com/@nejc_soklic?utm_source=medium&amp;utm_medium=referral">Nejc Soklič</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>DeepMind has recently announced that they mapped most of the 200+ million proteins into their 3D structure using AlphaFold. This news got me thinking a bit.</p><p>Not because of its scientific relevance: I know very little about proteins and couldn’t care less about those 200 million molecules.</p><p>It got me thinking because DeepMind has been in the spotlight for groundbreaking innovations in AI nonstop.</p><p>This is remarkable since the AI market is like a whac-a-mole field full of new companies popping up every second. For one reason or another, most of these companies get whacked, but some of them are lucky and survive.</p><p>DeepMind not only seems to have survived fairly well, but over time it has also established a reputation for being one of the best AI research companies on the planet.</p><p>After reflecting upon it, I started wondering:</p><ul><li>What made DeepMind innovate so much in the AI field?</li><li>How come DeepMind has never failed to deliver so far?</li></ul><p>I tried to give some reasonable answers in the following paragraphs.</p><h3>In 2010 nobody focused on AI, but DeepMind did.</h3><p>Ok, the title might be a bit of an attention grabber.</p><p>Indeed, DeepMind was founded in 2010, and actually, in those years there were some people talking about AI. But the underlying idea of this title is true: DeepMind started focusing on AI when nobody else really focused a lot on it.</p><p>And exactly in the early 10s, AI had an exponential boom. So, when companies started looking at the present firms in the AI market, DeepMind was already there delivering good projects.</p><p>As a matter of fact in 2014, only 4 years after being founded, DeepMind was acquired by Google for $500 million. To date, it is one of Google’s largest acquisitions!</p><p>We could say that DeepMind trusted the AI’s long-term potential when nobody yet believed in it, and this let it gain a “fairly unfair” competitive advantage.</p><h3>DeepMind likes playing games.</h3><p>This title instead is not an attention grabber. Most of the initial projects of DeepMind were on games.</p><p>It first started with an AI algorithm called DeepQ Learning (DQL) that was able to play 49 different Atari games just from seeing the pixels and score on the screen. Here you can have a look at how it plays <a href="https://www.youtube.com/watch?v=W2CAghUiofY">Space Invaders</a> and <a href="https://www.youtube.com/watch?v=TmPfTpjtdgg">Breakout</a>.</p><p>Then DeepMind came back in the headlines with AlphaGo, a program that was able to beat the European and the world champions in the game of Go. It learned how to play Go by watching thousands of human Go games.</p><p>This project then translated into AlphaZero, which not only was able to play Go, but also Shogi and Chess. The difference is that it learned just with the rules of the game, without seeing any previous match.</p><p>It later changed into MuZero, an even more enhanced version of AlphaZero. Here the AI learned to play games with no knowledge whatsoever, not even the rules.</p><p>In July, DeepMind announced the development of a new AI that plays the strategy board game Stratego.</p><p>Games are not a casual choice. The CEO, Demis Hassabis, was into chess since he was 4 years old, developed games as a teenager, and worked in the game industry after a computer science degree at Cambridge.</p><p>So if there is a good moral we can draw from this, it is the following: <em>play games, build a company, and become a millionaire easily</em>.</p><p>Although games are related to the CEO’s passion, they are actually a really clever choice for AI. They are easily usable, accessible, and reproducible. This allows fast progress!</p><p>But, personally, the real fact about games is that computers seem to understand things that we don’t comprehend.</p><p>An interesting example is given by AlphaGo. In one of the games against Go’s world champion, the algorithm did a move — dubbed Move 37 — that had only a one-in-ten-thousand chance of being made by a human. That choice was initially considered a mistake (by humans), but it then turned out to be a game-changing move that caused AI to win the game.</p><p>By choosing games as the core of its projects, DeepMind has surely made important innovations in AI.</p><h3>DeepMind does not care about success. It cares about progress.</h3><p>Summarizing DeepMind’s mission is a very easy task:</p><ul><li>Step 1: <strong>solve intelligence</strong></li><li>Step 2: <strong>use it to solve everything else</strong></li></ul><p>They’re big statements, aren’t they? The concept of solving intelligence to solve everything else is as fascinating as it is utopian.</p><p>DeepMind is aware that its goals are highly complex, but it is also well aware that the mission won’t be achieved right off the bat. In the end, DeepMind’s mission is as valuable as that of going to the moon. People talked about going to the moon for centuries, simply as a creative tale. But we only went to the moon 53 years ago.</p><p>DeepMind knows that solving intelligence has been a dream of many people for centuries. Just like NASA was one of the first to make the moon landing come true, DeepMind wants to be the first to solve intelligence.</p><p>No wonder in 2015 the CEO of DeepMind, Demis Hassabis, said:</p><blockquote>The idea behind DeepMind was really to create a kind of Apollo programme mission for AI</blockquote><p>Its approach is simple: build small bits of intelligence, that grow in their level of generality. And with that higher level of generality, harder problems can be solved.</p><p>For example, this vision is conveyed by:</p><ul><li>The reusability of the projects: AlphaGo changed into AlphaZero, which later turned into MuZero. These projects have the same core — games — but increased in their level of generality.</li><li>The availability to the public: DeepMind surprisingly made AlphaFold open source shortly after announcing it. In 2022 we know the structure of almost all existing proteins thanks to AlphaFold. I bet we will hear about AlphaFold in a future Nobel prize…</li></ul><p>So, DeepMind does not value the success of a project. It values its meaning for human progress.</p><h3>DeepMind doesn’t just invent. DeepMind executes.</h3><p>You can invent an idea, that’s fine. But it is well known that with ideas you do nothing; you need to execute them, to see if they are really valuable. DeepMind knows it.</p><p>This is possible due to the diversity of their teams. I do not mean the diversity that every company “likes” to talk about nowadays; I mean working diversity. In fact, DeepMind hires researchers, engineers, neuroscientists, experts in ethics and society, artists, etc.</p><p>By doing so, they are able to invent, experiment, and execute. DeepMind has done it in the past, it is doing it now, and it will do it in the future. This ability marked its journey, making it one of the most renowned AI research companies around the globe.</p><h3>Conclusions</h3><p>DeepMind was born in 2010. AI started to gain success around that same period. In a sense, DeepMind is like a parent that has seen the growth of his child. We have seen in the previous points why that might be the case.</p><p>Yet, in these years other companies started innovating AI as well: Google Brain, OpenAI, etc.</p><p>So saying that DeepMind is the father of AI innovation may be too disrespectful for the other firms. Let’s go further from that, but not too far.</p><p>DeepMind is not the father of AI innovation. DeepMind is the godfather of AI innovation.</p><p>And Demis Hassabis — the CEO — is Marlon Brando.</p><p><a href="https://medium.com/mlearning-ai/mlearning-ai-submission-suggestions-b51e2b130bfb">Mlearning.ai Submission Suggestions</a></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fgiphy.com%2Fembed%2FADze48TCg6Hao%2Ftwitter%2Fiframe&amp;display_name=Giphy&amp;url=https%3A%2F%2Fmedia.giphy.com%2Fmedia%2FADze48TCg6Hao%2Fgiphy.gif&amp;image=https%3A%2F%2Fi.giphy.com%2Fmedia%2FADze48TCg6Hao%2F200.gif&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=giphy" width="435" height="245" frameborder="0" scrolling="no"><a href="https://medium.com/media/78504505700250c781e367e4bd3f0ebc/href">https://medium.com/media/78504505700250c781e367e4bd3f0ebc/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e54bb0148ed7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What Is the Art of Artificial Intelligence?]]></title>
            <link>https://medium.com/predict/the-art-of-artificial-intelligence-4b16b36df1bc?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/4b16b36df1bc</guid>
            <category><![CDATA[creativity]]></category>
            <category><![CDATA[revolution]]></category>
            <category><![CDATA[dalle-2]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[art]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Sat, 23 Jul 2022 13:25:11 GMT</pubDate>
            <atom:updated>2022-07-23T17:21:37.585Z</atom:updated>
            <content:encoded><![CDATA[<h4>How a New Impressive Tool May Revolutionize Art Even More Than Photography Did</h4><figure><img alt="A robotic human painting in an art gallery" src="https://cdn-images-1.medium.com/max/1024/1*9c9QDea_jOYZxi3DbVQ6AQ.png" /></figure><p>By definition, art is in constant evolution. Movements and ideologies come and go. But, at times, some changes go down in history. Now it can be one of those cases.</p><p>Joseph Nicephore Niepce, a French researcher, took the world’s first picture ever in 1826. Little did he know that photography would change art — and the world — forever. Yet, he felt that it was “<strong>the first uncertain step in a completely new direction</strong>”. He could somehow smell an incoming revolution.</p><p>Two centuries later, art is about to witness again a new far-reaching evolution, possibly never experienced before.</p><p>Not because of photography, neither due to art itself nor due to artists. <strong>Art will change because of artificial intelligence (AI).</strong></p><p>If you don’t yet smell the incoming revolution, don’t worry. Let’s try to uncover together why and how AI will transform the world (of art), perhaps even more than photography did!</p><figure><img alt="A view from a window, black and white" src="https://cdn-images-1.medium.com/max/1024/1*mCdi5oVrkaMzDV5x40zHMQ.jpeg" /><figcaption>Niepce, “View from the Window at Le Gras”, 1826. The world’s first picture.</figcaption></figure><h3>Why should AI and art have a link?</h3><p>I know you asked yourself that question. And rightfully so. After all:</p><ul><li>AI performs only specific tasks that can be mathematically identified; art, instead, deals with creativity, which is all but mathematical.</li><li>AI is literally 10 years old; art, instead, has existed since the dawn of the human age.</li><li>AI is full of boring nerds wearing black hoodies; art, instead, is full of happy creative, and inventive people.</li></ul><p><em>Just kidding. The last point is stereotype. Don’t ever think again that I’m about to become a boring nerd wearing black hoodies!</em></p><p>Anyway, the gist is the following:<strong> </strong>these two worlds are too far apart from each other.</p><p><strong>But what if I told you that it is not true?</strong></p><p>I know it is a very <strong>bold </strong>statement, literally. But whether you believe me or not, I kindly ask you to do one thing: go to the top of the article and look at the cover image — the very first one. Look at it and then come back here.</p><p>Have you seen it?</p><p>Are you sure?</p><p>Ok. I trust you.</p><p>Go on reading.</p><p><strong>That image was generated by artificial intelligence!</strong></p><p>Yes, AI did that. It is a sort of android painting in an art gallery. No human being has ever spent a second of his/her life creating that piece of art. But an AI algorithm did.</p><p>In other words:</p><blockquote><em>AI did Art.</em></blockquote><p>Do you still see these two worlds too far apart from each other? I hope that by now you will agree with me that AI and art seem to have a link. If this is still not enough…</p><h3>…let’s try to dig into it.</h3><p>That image was created using an <strong>image generator</strong>. The most famous of these is <a href="https://openai.com/dall-e-2/">DALL·E 2</a> by OpenAI and it has been released a few months ago. Others have become available just recently. They include <a href="https://www.midjourney.com/home/?callbackUrl=https%3A%2F%2Fwww.midjourney.com%2Faccount%2F%3FcallbackUrl%3D%252Fapp%252F">MidJourney</a> — used for the cover image — <a href="https://www.craiyon.com/">Craiyon</a> and others. There is even <a href="https://imagen.research.google/">Google</a>, but its project is not public yet (rumor has it that it is the best one).</p><p>How do they work?</p><p>If you’re eager to discover what is behind image generators, I’m sorry to tell you that I won’t discuss it in much detail. There’s plenty of information online that explains everything much better than I would.</p><p>Instead, I’ll try to be as simple as possible.</p><h3>What do you have to do?</h3><p>You, <em>a human being with almost non-existing artistic capabilities</em>, have to do the following things, in order:</p><ul><li>Open the image generator</li><li>Input a text prompt</li><li>Wait a few seconds</li><li>Look at the generated images</li></ul><p>Nothing more.</p><p>If you were worried about doing some strange math or writing code to get the images, I hope your fear is now gone.</p><p>It is all just like searching for an image on Google. The only significant difference is the following: <strong>you do not search for images, you create them.</strong></p><p>It is as simple as texting a friend and reading the response — <em>actually, this is not true. Some people take days to respond. So, AI creates images faster than people answer text messages.</em></p><h3>What does AI do?</h3><p>Those “waiting seconds” represent the true culmination of state-of-the-art techniques in AI, developed over the years. But they are not easy to explain for different reasons.</p><p>Firstly, it’s not easy stuff. You need to know many things about AI to deeply understand what is going on.</p><p>Secondly, each company might utilize different techniques. For example, this is how DALL·E 2 — the best image generator so far — works in simple words: the images are developed using a neural network model called a <em>transformer, </em>trained mapping images to text and vice versa. The neural network uses a technique called <em>diffusion. </em>Here random pixels are created and modified step by step, in order to match the text prompt as best as possible. This is how you get your image. But Google’s Imagen uses a slightly modified approach and other firms might use other variations.</p><p>This explains why different generators might understand the same prompt differently.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/668/1*7U8DPEwkBfKuF6LEs28aiw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ky_Enq4G0XSErT5S45cnDQ.jpeg" /><figcaption>“A quiet night at a Tuscan restaurant with an old man sitting, drinking wine, oil painting”. On the left, Huggingface’s output. On the right, Craiyon’s output.</figcaption></figure><p><em>All of these images look like real paintings! Isn’t it awesome?</em></p><p>I know I gave you a lot of technical information. You might feel overwhelmed. But, honestly speaking, whether you understood all of it or not, it still remains fundamental to appreciate that there is no human magic trick behind AI-generated art.</p><p>There is no creativity. There is no sentience. There is no feeling. It is only math. Yet, that math brings to life something truly outstanding that can touch humans.</p><h3>Sounds like a fun game. But where’s the revolution?</h3><p>That was exactly what I thought when I first discovered AI image generators: they were a fun game. I only saw the possibility of finally picturing a flying cow in the desert.</p><figure><img alt="Flying cow in the desert. 4 images" src="https://cdn-images-1.medium.com/max/512/1*UQ4X7HL7a-W8WOvTLC3wqA.png" /><figcaption>“Flying cow in the desert”. Generated using MidJourney.</figcaption></figure><p><em>That is the best flying cow I have ever seen and, probably, that I’ll ever see.</em></p><p>Admire this beauty. If it had been drawn by Magritte, Picasso, or any other notorious artist, everyone would have praised them for such inventiveness and imagination.</p><p>Now everyone must praise AI.</p><p>Jokes aside, <strong>the revolution is right in front of your eyes. </strong>There are endless possibilities: it’s all about imagination!</p><p>Imagine a child that wants to picture a flying cow in the desert. There you go! Children might benefit a lot from something like this.</p><p>Imagine a film director shooting the final scene of his apocalyptic movie that wants a sketch of the final shot: the “Earth exploding”. Try to draw it yourself. I know you don’t feel like doing it. Neither any director would. But with AI, in a few seconds, some ideas might come up:</p><figure><img alt="Earth exploding. 4 images" src="https://cdn-images-1.medium.com/max/512/1*j8fbcENREShF49qTWyVI_w.png" /><figcaption>“Earth exploding”. Generated using MidJourney.</figcaption></figure><p>Imagine a philosopher trying to describe what walking “into the unknown” is like:</p><figure><img alt="Into the unknown. 4 images" src="https://cdn-images-1.medium.com/max/512/1*Z9TXkaAPuxHwaGO4qCAVAw.png" /><figcaption>“Into the unknown”. Generated using MidJourney.</figcaption></figure><p>Imagine a game designer, that wants to create the main character of a videogame. Something like a “mysterious man with a trench coat and fedora hat walking through a dark gritty street in the cyberpunk world with neon street signs illuminating the background”:</p><figure><img alt="Mysterious man with a trench coat and fedora hat walking through a dark gritty street in the cyberpunk world with neon street signs illuminating the background" src="https://cdn-images-1.medium.com/max/1024/1*NEp6mZOvNsXzTdBEqG4Xow.png" /><figcaption>“Mysterious man with a trench coat and fedora hat walking through a dark gritty street in the cyberpunk world with neon street signs illuminating the background”. Generated using MidJourney.</figcaption></figure><p>I could go on forever. <strong>This is a spaceship for human creativity. </strong>I hope you see it too.</p><p>In the 19th century, with photography, you could capture anything you wanted and see the results after a few hours of exposure. It did not take long to change art and the world, forever.</p><p>In the 21st century, with AI, you can create anything you want — in seconds. I’ll say that again:</p><blockquote><strong>ANYTHING YOU WANT. IN SECONDS.</strong></blockquote><p>I guess it won’t take long to change things here as well. Don’t you smell the incoming revolution?</p><h3>What doors has AI opened now?</h3><p>AI image generators are a new thing. But some people have already taken true advantage of it.</p><p>The first person is me. Why? Because if I were to draw something for an art class as homework, I would copy everything from an image generator. That’s a guaranteed top mark!</p><p>As far as serious people go, “<strong><em>more than 3,000 artists from more than 118 countries have incorporated DALL·E into their creative workflows</em></strong>” states OpenAI. Check out <a href="https://openai.com/blog/dall-e-2-extending-creativity/">their blog article</a> for more info. You will be surprised how human beings have been able to reinvent these tools for their own benefit.</p><p>They are all spreading faster than Covid in 2020! And undoubtedly faster than photography in the 800s.</p><h3>What doors will AI open in the future?</h3><p>A revolution is not a one-time event. We are just at the beginning.</p><p>To give you an idea of where I think we are at the moment, recall that the first picture was captured in 1826. It was black and white. The first color photograph was taken by Maxwell (<em>yes, the physicist</em>) in 1861, 35 years later. That event changed things even more than the Niepce’s invention!</p><p>Now, the AI-ART couple is still at 1826 and is having lots of success. But we don’t know when we’ll start “seeing color” to achieve true success.</p><p>Who even knows a few years down the line what could potentially happen with animations or videos. They are just images stacked together, aren’t they? Imagine how social media might start implementing these tools in their apps for content creation!</p><p><strong>I don’t know what you feel, but I smell a strong scent of revolution!</strong></p><p>AI will open many doors, probably more than photography did.</p><h3>What other doors will AI close?</h3><p>Let’s be honest. AI is a newborn in our world. Art — taken in its most absolute sense — is not. There will be some inevitable clashes.</p><p>At the time of writing, AI will surely close my door which asks about prompts that include a face. Some of them were horrible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dw7wZ0-6eke-mynfuptG3w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ky8zZVDV1qbY7OKrKsTEyg.jpeg" /><figcaption>No caption needed.</figcaption></figure><p>Everyone can fail. Even AI. But these models improve day by day, so faces won’t be terrifying for long.</p><p>Talking about serious problems… first of all, there are huge ethical issues. You can easily imagine that a bad person could try to create fake images of other people doing bad things, like with deep fakes. Those images will be pretty realistic because we have seen that these AI algorithms perform very well.</p><p><strong>Something needs to be done about it.</strong></p><p>We still have no true global regulations on AI — I say global because Europe actually has one. The effects of “not doing anything until something bad happens”<em> </em>could be catastrophic. History knows something about it.</p><p>By the way, this is also the reason why most of these technologies are not fully public yet: companies want to be sure that their products are used responsibly.</p><p>These issues go back to those of fake news as well. For instance, I could write an entire article on a new sea animal fossil discovered under the ocean, attaching just one single image to prove it:</p><figure><img alt="A dead sea creature in a large fishtank" src="https://cdn-images-1.medium.com/max/1024/1*TgVAmGcj46XFwVctrD8_bg.png" /></figure><p>Is this image real? Is it AI-generated? I’m not going to tell you. This is what could really happen in the future.</p><p>Other issues could exist related to ownership. Who owns the picture? The writer of the text prompt? The downloader? The company that created the image generator? This is still unknown.</p><p>One problem, in particular, is dear to my heart: jobs will disappear.</p><p>This is the first time I truly understand the meaning of it. Why would a startup ever hire a designer to draw images for them, investing lots of money and time, when they could easily get an image of what they want in seconds for free?</p><p>I sincerely suggest you check out <a href="https://www.youtube.com/watch?v=MwAAH9tBoMg">Marques Brownlee’s video</a> to understand that the issue is not a joke. Maybe it is now. But in a few years, it will not.</p><p><em>Do I sound too pessimistic? I swear I’m not, for real.</em></p><h3>Finally, is it good or bad?</h3><p>I know it’s tempting to look at the cons! Most people do not cope much with such big changes. Especially if it’s about AI, for some unknown reason.</p><p>But someone once said: “We are sorry for the inconvenience, but this is a revolution”. I guess it was for a civil war or something, but the idea works here as well. That is, bad things could happen and there are aspects that will need to be readjusted. But when one door shuts, another opens.</p><p>Don’t be overly dystopian. The future of art is certainly unpredictable. That of AI too. But the odds seem to be going in a very clear and plausible direction, and we need to trust that it will be bright. You need to <em>trust the process</em>.</p><p>Try to be like Niepce. Try to see all of this as “<strong>the first uncertain step in a completely new direction</strong>”.</p><p>It is only by doing so that changes will become remarkable and go down in history. Just like it happened with photography, eventually, it will happen with AI too. That’s all it takes: trust.</p><p>I’m happy if you’ve come this far. I’ll award you with my favorite image so far: “eye of the universe”. Who cares if an AI did it. <strong>This is true art</strong>. And it is super cool. This is…</p><blockquote>The Art of Artificial Intelligence</blockquote><figure><img alt="Eye of the universe" src="https://cdn-images-1.medium.com/max/1024/1*Qck2MlIn9QSuMMhDo6zuTw.png" /><figcaption>“Eye of the universe”. Generated using MidJourney.</figcaption></figure><h3>Extras (optional)</h3><ul><li>To see more amazing pictures of the art of AI, check out <a href="https://www.instagram.com/openaidalle/">DALL·E 2’s Instagram account</a>.</li><li>Mark Zuckerberg has used AI image generators but in a unique way. <a href="https://www.instagram.com/p/CgCSWfTrkdv/">Check it out</a> on Instagram!</li><li>If possible, try to use the image generators yourself! It’s a lot of fun!</li><li>You still don’t know the text prompt of the cover image. Guess it!</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4b16b36df1bc" width="1" height="1" alt=""><hr><p><a href="https://medium.com/predict/the-art-of-artificial-intelligence-4b16b36df1bc">What Is the Art of Artificial Intelligence?</a> was originally published in <a href="https://medium.com/predict">Predict</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why TikTok Is NOT Optimized for Humans]]></title>
            <link>https://medium.com/design-bootcamp/why-tiktok-is-not-optimized-for-humans-c111f91b3dbf?source=rss-2630bbcfbc44------2</link>
            <guid isPermaLink="false">https://medium.com/p/c111f91b3dbf</guid>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[user-interface]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[tik-tok]]></category>
            <category><![CDATA[social-network]]></category>
            <dc:creator><![CDATA[Sandro Mikautadze]]></dc:creator>
            <pubDate>Fri, 08 Jul 2022 11:08:06 GMT</pubDate>
            <atom:updated>2022-07-09T10:03:28.268Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why TikTok is NOT optimized for humans</h3><h4>How a non-human-oriented choice let TikTok become the most used social media faster than ever</h4><figure><img alt="A hand holding a phone with the TikTok logo at the center" src="https://cdn-images-1.medium.com/max/1024/0*xgT2q-FWjVhujGCR" /></figure><p>Unless you’ve been living under a rock for the last few years, you are well aware that TikTok has become a new big player in the social media market and, most importantly, in an extremely short amount of time.</p><p>You don’t believe me? Well, suppose you were TikTok’s boss and that you had access to all necessary data to track the success of your business. In 2018, you would see revenue of 150 million dollars and 133 million users¹!</p><p><em>Don’t know about you, dear Mr. TikTok-boss, but I would be already pretty satisfied!</em></p><p>If we jumped to 2021 the situation would be the following: 902 million users and <strong>4.6 billion dollars</strong> of revenue, with <strong>1.8 billion users expected</strong> by the end of 2022¹.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CWi2Sx1sKo94BVPeeKSJfg.png" /></figure><figure><img alt="Graphs representing the revenue and the monthly active users on TikTok from 2017/18 to 2021" src="https://cdn-images-1.medium.com/max/1024/1*HukJ3pFyNph0mtH-c1ZyMg.png" /><figcaption>In blue is the revenue, and in red is the number of monthly active users¹</figcaption></figure><p>Just to give you an idea, Instagram was at roughly 150 million users in 2013 and reached 1 billion users in 2018². TikTok reached 1 billion in the third quarter of 2021¹, <strong>1 year faster than IG</strong>! It seems like TikTok is no longer a teen-dancing academy and is hitting <strong>user and revenue milestones</strong> <strong>at the speed of light</strong>!</p><p>Ok, I’m done with the boring economic data.</p><p><strong>Let’s jump straight into the point of the title</strong>: to understand why Byte Dance’s social network is <em>built different</em> we don’t<strong> </strong>need any fancy business chart with strange, incomprehensible and (at times) meaningless economic metrics. We don’t need to go interview the boss of TikTok. We don’t even need to have TikTok at all! All we need to do is just look at its user interface (UI)!</p><h4>Let’s try to understand…</h4><p>We live in an era in which we “<strong>consume immediate content</strong>”. That is, we want immediate satisfaction from things we immediately like. On the other side of the coin, there are social-media companies, whose main goal is to make money while we consume the content that they show us.</p><p>Nowadays, most companies use artificial intelligence (AI) to optimize the user experience and make us have a good time with their app. If you don’t grasp how this happens, let me ask you a question: <strong>have you ever wondered why, after putting a like to a cat picture </strong>(<em>why do cats always show up in picture examples?!</em>)<strong>, your homepage suddenly becomes filled with posts of cats?</strong></p><p><em>By the way, I don’t like cats, dogs are better.</em></p><p>In this case what Instagram does, in a nutshell, is <strong>collect data</strong> about your preferences and<strong> feed it to AI algorithms</strong> that then <strong>predict </strong>what other posts you are going to engage with the most. This makes you, the user, spend (<em>or waste</em>) more time on their app watching the content you enjoy and gives them, the company, tons of money.</p><p>Pretty much every company you can imagine does something like this: Instagram, Facebook, Snapchat, Twitter, Netflix, Amazon, you name it. TikTok too.</p><p>There is one thing, though, that sets this social media apart from the rest: <strong>TikTok is not optimized for humans; it is optimized for AI.</strong></p><h4>Let’s try to understand… even more.</h4><p>If you’re confused, bear with me. TikTok does exactly what I described before: it collects data, gives it to AI algorithms and deploys the results to increase user experience. But to maximize the success rate of these algorithms, companies look after <strong>which </strong>data is collected and <strong>how</strong>.</p><p>For example, if I wanted to predict what products you would buy in the future from my online store (<em>don’t reveal this secret to </em><a href="https://it.wikipedia.org/wiki/Jeff_Bezos"><em>Jeff </em></a>), I would better collect data on your search history and on the articles that you clicked on, rather than on the time at which you entered the website, wouldn’t I?</p><p>Similarly, for social media content, it is fundamental to collect the correct data, like interactions, viewers, watchtime, etc. in the best possible way. Now, remember what I said a few lines ago:</p><blockquote>“All we need to do is just look at its user interface (UI)”</blockquote><p>Hear me out: TikTok’s UI is built to extract data directly necessary <strong>for AI </strong>to be analyzed, not for humans to be evaluated. Turns out this is a <strong>game-changer</strong>.</p><p>In fact, consider Instagram’s feed: you have stories on top, you have a post at the center, which is either a video, a picture or a reel, you have the account name, you have the caption, the hashtags, the comments, the number of likes, you might even have some parts of another post showing up at the bottom of the screen… for AI algorithms <strong>IT IS A MESS</strong>!</p><figure><img alt="Instagram’s homepage" src="https://cdn-images-1.medium.com/max/417/1*NyT2jGoTzTFQwUqPw2zSRg.png" /><figcaption>Instagram’s feed</figcaption></figure><p>Instead, TikTok’s feed is much cleaner: you look at the video that is given to you and that is it. All the possible distractions, like the captions, comments, likes, etc. are “hidden”. In addition, you look at contents one at a time: you cannot scroll down and watch two videos simultaneously.</p><figure><img alt="TikTok’s homepage" src="https://cdn-images-1.medium.com/max/463/1*OdGyzsgztiTnNCC26UmNsA.png" /><figcaption>TikTok’s home page</figcaption></figure><p>Now you might be like “<em>Mh… so what?</em>”. <strong>SO WHAT?!</strong></p><p>Here’s the thing: if Instagram wants to suggest you a new post, it needs to know how much time you spend on a post, what kind of post that is, and a bunch of other stuff. You can deduce that Instagram’s feed makes this task not so easy, because of the many pieces of information that the homepage has. In a few words, this means that there’s some chance that you might end up having a feed full of cats when you never even put a like on a cat…</p><p><strong>With TikTok, this cannot happen</strong>. Or, at least, it is highly unlikely. Or, actually, way less likely than Instagram’s case.</p><p>TikTok’s UI is specifically built to collect <strong>accurate data</strong> on how much users enjoy a piece of content and does not have “background noise” that lowers the performance of the algorithms, like Instagram or many other companies have. TikTok’s interface is indeed the first one specifically designed for AI. This enables them to feed the AI algorithms with detailed-enough data such that the user experience can drastically improve in a short amount of time, without many corrections.</p><p><strong>This also explains in part why it is easier to gain followers on TikTok</strong>: people see what they like very easily, many times, and therefore decide to click the <em>Follow</em> button much more frequently than on Instagram or other social media.</p><p><strong>No wonder</strong> TikTok’s revenues have been sharply increasing in a short span of time. Indeed, data-driven businesses like those of social media work as follows: the more things you enjoy watching on the app, the more time you will spend on it; the more time you spend on the app, the more data the company can collect and the more money they get; the more money the company gets, the better algorithms they can develop and the better the user experience. Then the cycle begins again.</p><p>The following chart illustrates that <strong>TikTok seems to be the firm that has best captured the essence of the previous equation</strong>.</p><figure><img alt="A graph displaying time on the x-axis, and hours on the y-axis. The graph shows the average monthly hours spent per user of TikTok, Facebook, Instagram and Whatsapp. TikTok has 25.7 hours, Facebook 16 hours, Instagram 7.9 hours and Whatsapp 7.8 hours" src="https://cdn-images-1.medium.com/max/1024/1*1W0VcTb_xdx8x9V_etJCjQ.jpeg" /></figure><p>So, <strong>TikTok is winning by a landslide thanks to this UI</strong>. But the other competitors have never been standing still watching the rise of the Chinese company. Instagram added “Reels”, YouTube added “YouTube Shorts”, and other firms did other stuff that I don’t know.</p><p>The remarkable thing is that not only did they add these features to their apps, but they also literally copied TikTok’s UI! This explains why even these obviously copy-pasted products work well in ecosystems, like Instagram’s, to which they don’t originally belong. Interestingly enough, this UI turned out to have so many benefits that, guess what, “<a href="https://www.theverge.com/2022/5/3/23055496/instagram-full-screen-vertical-video-test-tiktok"><em>Instagram is testing a TikTok-like full-screen feed</em></a>”! It seems that nobody cares about photos anymore…!</p><figure><img alt="A new possible Instagram feed. Looks very similar to TikTok’s" src="https://cdn-images-1.medium.com/max/344/1*T0C5o-qukdz7sgf-oWIC9A.png" /><figcaption>This is what the new Instagram feed might look like… it seems even much cleaner than TikTok’s, doesn’t it?</figcaption></figure><h4>Let’s wrap it up</h4><p>Without a doubt, <strong>TikTok is the game-changer that nobody expected to have in the social media industry a few years ago</strong>. It has more usage than Instagram, Facebook, and WhatsApp, and it will steadily increase in the future. Its success is certainly a contribution of many different factors such as technology, marketing, innovation and engineering, not only a matter of user interface.</p><p>Yet, in all of this, it is fascinating to notice how building a system that properly collects data for AI can lead to a significant competitive advantage. No wonder other companies like Instagram and YouTube (who knows how many they actually are) decided to take inspiration from TikTok’s strategy, <em>just a fancy way of saying that they literally copied TikTok.</em></p><p>The future is uncertain, but the one thing we can be certain of is that if companies find the correct strategies to exploit AI, then they will create some huge benefits both for themselves and for the users. Humans seem to enjoy TikTok, even though, as we have seen, TikTok is NOT optimized for humans, but for AI. <strong>We’ll see how things will go!</strong></p><h4>References</h4><ul><li>[1] “Tiktok Revenue and Usage Statistics (2022).” <em>Business of Apps</em>, 30 June 2022, <a href="https://www.businessofapps.com/data/tik-tok-statistics/">https://www.businessofapps.com/data/tik-tok-statistics/</a>.</li><li>[2] “Instagram Revenue and Usage Statistics (2022).” <em>Business of Apps</em>, 30 June 2022, <a href="https://www.businessofapps.com/data/instagram-statistics/">https://www.businessofapps.com/data/instagram-statistics/.</a></li></ul><h4>Extras (optional)</h4><ul><li>Many pieces of information have been collected online, but I mostly drew inspiration (<em>just a fancy way of saying that I literally…</em>) from <a href="https://gianlucamauro.com/">Gianluca Mauro</a>’s videos and articles. Check him out, he’s really good!</li><li>There are many non-trivial technicalities that are oversimplified. I apologize to all experts of AI, TikTok and cats.</li><li>TikTok has tons of data about its users that we cannot even imagine. With the emergence of AI, there will constantly be a battle between data collection and privacy. Unluckily, TikTok is owned by China, a dictatorship in disguise, that is famous for all but keeping its own civilians’ privacy. The constant rise of this social media will make it much more difficult for us to maintain true privacy. But this is probably a topic worthy of another article!</li><li>I don’t have TikTok.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c111f91b3dbf" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/why-tiktok-is-not-optimized-for-humans-c111f91b3dbf">Why TikTok Is NOT Optimized for Humans</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>