YT YouTube2Text

YouTube2Text → Transcripts → Historias Innecesarias: Inteligencia Artificial - NADA es REAL...o sí

Historias Innecesarias: Inteligencia Artificial - NADA es REAL...o sí — Transcript

by Historias Innecesarias · 4,187 words · 639 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00How's it going? Welcome everyone to a
0:02new edition of Unnecessary Stories. How
0:04are you doing? In today's edition, we
0:15are going to talk about artificial
0:18intelligence. So, pay close attention.
0:26Right off the bat, I would like to
0:27clarify that this is going to be a
0:28video When I was a kid and spent
0:45excessive and concerning amounts of
0:47time on the computer doing nothing but
0:49that; it was common to hear my mom tell
0:52me: "You're on the computer all day,
0:54were you not using the computer or on
0:56the computer?" I was with the computer,
1:00it wasn't just another device or object
1:02in the house, it was almost an entity,
1:04a concept, a being. The computer. That
1:08stayed in my imagination for a long
1:10time and I never imagined that now,
1:12some 15 years later, that entity my
1:14mother attributed to the computer,
1:16perhaps a bit disparagingly, would end
1:19up becoming almost a reality, if it
1:21isn't one already. Artificial
1:23intelligence is the concept according
1:25to which, quote unquote, machines think
1:28like human beings. Normally, an
1:30artificial intelligence system is
1:32capable of analyzing data in large
1:34quantities, identifying patterns and
1:36trends, and therefore, automatically
1:38formulating predictions. In short, it's
1:41the computer trying to imitate the
1:43human being. It is a system that is
1:46present in applications we use daily.
1:49Siri, Alexa, the tag suggestions
1:51Facebook or Instagram make for you, is
1:53it your opponent when you play FIFA or
1:55Mario Kart alone? It's responsible for
1:58you being able to put your face in a
2:00movie, age yourself with an app, or
2:02it's the reason I can make a fake Dross
2:04voice say the following: "
2:07Hello, how are you? Welcome everyone to
2:10this new edition of Unnecessary Stories
2:13.
2:14Although the term artificial
2:16intelligence was only adopted in 1956,
2:18its beginnings date back a few years
2:20earlier with Alan Turing and the
2:22creation of the Turing test. An exam to
2:25test whether a machine can pass as a
2:27human or not. Spoiler, yes. In 1966,
2:31MIT developed ELIZA, one of the first
2:34programs to process natural language
2:36and converse through a series of
2:38programmed phrases. In 1979, a computer
2:42beat the world champion in Backgammon,
2:44Luigi Villa. Years later, in 1997, the
2:48computer Deep Blue, developed by IBM,
2:51beat the then-world chess champion,
2:54Garry Kasparov. I could spend hours and
2:58hours talking about concepts and
3:00significant facts, but given how fast
3:02this whole topic evolves and how much I
3:04have to show you, I prefer to focus on
3:06the present. Since 1990, the study and
3:10experimentation of this artificial
3:13intelligence concept has been focusing
3:15on two specific things: machine
3:17learning and deep learning. And get
3:20ready, because here comes the crazy
3:21part. If we say that artificial
3:24intelligence is, quote-unquote, the
3:26computer trying to imitate human
3:28behavior, it means the computer can
3:30learn. And the way a computer learns
3:34when we talk about machine learning and
3:36deep learning is through neural
3:38networks, just as we human beings learn
3:40. Let's dive briefly into this. In our
3:44case, as humans, what you are seeing on
3:47screen is the exact moment of the birth
3:50of a neuron, specifically the moment a
3:52human stem cell transforms and
3:54originates a new brain neuron. Every
3:58time we do something, it is the
3:59response of millions of these
4:01interconnected neurons. Each neuron
4:03receives electrical stimuli from other
4:05neurons. Upon receiving that stimulus,
4:08it processes it and, if necessary,
4:10stimulates many more neurons, and so on
4:13, until it forms an entire neural
4:14network responsible for, for example,
4:17making you click the mouse when we want
4:19to do something new. For example,
4:22learning to play the keyboard.
4:23Hi, how are you? We are going to learn
4:25a very well-known song. You can play on
4:28the white keys. We need different parts
4:32of the brain, or rather different
4:33neural networks, to work together.
4:37Playing the keyboard requires
4:38collaboration between parts like
4:40coordination, motor skills, and vision.
4:43Since I've never played the keyboard,
4:45now we're going to play the F and the
4:48E-F. Okay. Those neural networks aren't
5:00used to working together and I struggle
5:02to learn until, with a lot of practice,
5:04the neurons improve their communication
5:05, reinforce their connections, and
5:07consequently, I learn to play the
5:10keyboard. The work of many
5:43interconnected and functioning neural
5:45networks is what allows me to play the
5:47keyboard, pick up a glass of water, or
5:49understand this photo as a picture of a
5:51dog. Even the fact that you went on
5:54YouTube to watch this video is a
5:56consequence of millions of neurons
5:58working together and processing
6:00information in your brain. Now imagine
6:02that this whole complex process, which
6:05I’ve summarized in a very simple and
6:07perhaps misguided way, can also be done
6:09by a computer in its own way. That is
6:12exactly what artificial neural networks
6:15are: networks based on the functioning
6:17of biological neural networks, but
6:19which are made up of, in a nutshell,
6:21many mathematical formulas, circles,
6:23and lines. How these artificial neural
6:27networks work is very complex to
6:28explain, and if I’m being honest, I
6:30don't fully understand it myself.
6:33However, if you are interested in
6:35understanding how they work, I
6:37recommend these two videos, one by Dot
6:39CSV and the other by Ringatech. In the
6:42meantime, what we do need to understand
6:44is that these are systems that aren't
6:46programmed, but rather learn and form
6:48themselves just like our neurons do.
6:51Their goal is to try to solve problems
6:53in the same way the human brain solves
6:55them. They seek to translate, acquire,
6:58process, and understand the different
7:00images and sounds of the real world to
7:02be able to process them just like our
7:04brain. And while understanding how they
7:07function is complicated, it's very easy
7:09to see them in action and understand
7:11the concept. And for that, we’re
7:14going to use a simple example: cats.
7:17And I'll warn you right now that almost
7:19everything I do in this video, you can
7:20do too using the links I’ve left in
7:22the description. If we take a lot of
7:25photos of kittens and give them to the
7:27computer so it learns how a cat is
7:29composed, it will end up learning how
7:31to generate them. An example of that is
7:34this page, which updates every time we
7:37visit it with a computer-generated
7:39image of a cat that simply doesn't
7:41exist. They aren't Google photos, and
7:43they aren't even photos at all. They
7:45are the representation of what the
7:47computer understands a cat to be. We
7:50can even use other platforms to
7:51generate cats from simple sketches. No
7:54matter what we do, it will try to turn
7:57it into a cat. Even if we make cats
7:59that are physically impossible. And
8:02what happens if we train the computer
8:04with photos of people instead of cats?
8:06The same thing. It can create humans
8:08that don't exist. It imagines them. I
8:11know it might be hard to believe, but
8:13none of these people you are seeing
8:15exist in reality, or you can even train
8:18it with landscapes and have it process
8:20a really basic drawing and transform it
8:23into one. The computer understands
8:25which color is a tree, which is soil,
8:27which are rocks, which is a lake, or
8:29which are clouds, and it can generate a
8:31landscape even if the initial sketch
8:33looks like it was made by a child just
8:35learning to draw, even if it was mine
8:37and I'm 26 years old. Right. Other
8:40examples of deep learning are the
8:41deepfakes that exploded a few months
8:43ago, which allow you to put your face
8:45into any movie. Harry Potter. Come to
9:00die. didn't st it. I need to talk to
9:03the goblin. Instagram filters that
9:23change your look, or apps that can cut
9:25your hair, make you look old, like a
9:27child, or give you a simple and
9:28disturbing smile. are also examples of
9:31deep learning. All those things work in
9:34a somewhat similar way. It involves
9:37training the computer to see patterns
9:39based on image recognition, so that in
9:41this case, it understands the structure
9:43of a human face and can replace it with
9:45another, or add a different beard or a
9:47different effect. Now then, what
9:49happens if we train it with our
9:51language? What happens if an algorithm
9:54is developed that allows it to create
9:56the next best word to follow in a text?
9:59That is exactly what GPT-3 is, an
10:01artificial intelligence technology made
10:04up of various GPT-3 models, created by
10:06OpenAI, which aims to facilitate the
10:09development of more efficient and
10:11accurate machine learning models. It
10:14works by analyzing and learning from
10:16large amounts of data like texts,
10:18images, or videos, and then uses what
10:21it has learned to answer questions and
10:23perform tasks. And these last two
10:25paragraphs I just read, I didn't even
10:28write them; I asked the computer and it
10:30wrote them for me. GPT-3 can be used in
10:33an endless number of ways. You can ask
10:36it simple things like," What is the
10:37life expectancy in Argentina? "You can
10:42ask it to translate phrases into more
10:44than 24 languages. It can be used to
10:49write a restaurant review, or even to
10:58help you study or make a video about
11:00artificial intelligence and not forget
11:02any important points to mention. The
11:08uses of GPT-3 at the moment are not
11:10really a practical solution to major
11:13industrial problems, but the research
11:15community uses it constantly to improve
11:18it day by day. And if you are original
11:21enough and know how to communicate
11:23correctly with the computer, which just
11:25takes a bit of practice, it can
11:27simplify many things for you. I know
11:30people who use this to make, for
11:32example, radio columns or even
11:34summaries for college. Maybe if you are
11:37very clear, the computer can do your
11:39thesis for you, or at least help you.
11:42Small fact: a study conducted in 2020
11:45indicated that only 52%of readers
11:47detect which texts are created by GPT-3
11:50. Now then, let's go back to images. We
11:54already saw what happens if you train
11:56the computer with language, with cats,
11:57humans, or landscapes. But what happens
12:00if, once it already knows all those
12:02concepts and many more, it starts to
12:04combine them? That is exactly what
12:07DALL-E 2 is, an artificial intelligence
12:09that you can ask to make any image you
12:11want. Images that, in case there's any
12:15doubt, don't exist until the moment you
12:17generate them, to put it simply. The
12:20computer shows you what it imagines
12:22based on what you tell it. You’ve
12:24surely seen some strange combinations
12:26on social media like Voldemort at the
12:28hair salon, a Demogorgon playing
12:30basketball, or Mickey Mouse in prison.
12:33Those were made with DALL-E Mini, a
12:35free version that anyone has access to,
12:37and while it isn't as realistic, it’s
12:39quite good and clearly fun. On the
12:42other hand, we have DALL-E 2, a
12:44platform that doesn't have open access
12:46yet. When I set out to make this video,
12:49I took it for granted that I had to get
12:51it, so after months of insisting and
12:53insisting, I succeeded. DALL-E 2 works
12:56because, in a nutshell, they taught the
12:59computer millions and millions of
13:01images and concepts, and now it has
13:03enormous freedom to imagine anything in
13:06seconds. For example, we can ask for
13:09simple things like an ocean; something
13:12weird, like a 3D render of a race
13:14between horses and cars on the moon; a
13:16realistic portrait of an old woman; a
13:18close-up of a clown; the computer’s
13:21interpretation of hell, or a photo of a
13:23bald man with a pipe. We can ask for a
13:26real, detailed photo of a blue cactus
13:28with a red flower, or by changing just
13:31two words. We can ask for the same
13:33thing, but in the style of an old
13:35painting or, even simpler, a drawing.
13:38Or without knowing how to paint, I can
13:40make a picture of the moment I’m
13:41writing the script for this video or of
13:43a gathering with my friends while we
13:45play Mario Kart. DALL-E wasn't trained
13:47to know who Frida Kahlo is or what
13:49every type of painting is. It was
13:52trained with more than 650 million
13:54images that allow it to draw
13:55conclusions that it then shows us based
13:57on what we ask for. Something I find
14:01incredible is that if you ask for a
14:03realistic photo, you can even specify
14:05the camera model and the type of lens
14:07used. The more details you give it, the
14:09better it will be able to imagine what
14:11you’re asking for. It’s all a
14:12matter of, as I already said and even
14:14if it sounds crazy, learning to
14:16communicate with the computer. DALL-E
14:18has quite a few ethical boundaries.
14:20They are careful that the tool has a
14:22healthy use, and that’s why many
14:23words are blocked. Also, it doesn't
14:26generate images of well-known people.
14:28To be more specific, if you ask for Leo
14:30Messi eating pasta at a restaurant, it
14:33will show you people who don’t exist,
14:35but who have characteristics similar to
14:37what the computer understands as Leo
14:39Messi. The same applies if you ask for
14:41Mirta dancing with Hulk. And safety is
14:44precisely the reason why, at least
14:46until the time of this video's release,
14:49access to DALL-E is quite limited. For
14:52the moment, it's only being given to
14:54artists, and if you're interested, you
14:56can send them messages on Twitter or
14:57Instagram to tell them why you want
14:59access. They don't care if you have
15:02many or few followers, just that you
15:04are an artist. While it is forbidden to
15:08sell DALL-E creations as NFTs or sell
15:10the digital version to someone for
15:12money, they do allow you to print that
15:14image and do whatever you want with
15:16that physical version. Since mid-July
15:212022, you officially own all commercial
15:23rights to every image you generate,
15:26whether it's an image from scratch or a
15:28reimagining of the Mona Lisa. And yes,
15:32like everything in this video, that
15:34raises hundreds of questions that we
15:36don't really know how to answer yet. Am
15:38I really the owner? Will we all be
15:41artists? Recently, along with a
15:43photographer friend, we did an
15:45experiment. He posted several
15:47photographs made with DALL-E over a few
15:49weeks and no one noticed. You'll likely
15:52even have trouble telling which ones
15:54were created by artificial intelligence
15:56. No one realized, and in case you
15:58didn't either, they are this one, this
16:00one, and this roll of photos. Small
16:03side note: if you are interested in how
16:05DALL-E works in depth, I recommend this
16:07video by Vox. And to learn more about
16:10creating images with machine learning,
16:12this video by Tomás García. And we
16:15are talking about infinite
16:16possibilities, from something as simple
16:18as changing your phone wallpaper to
16:20something dreamed up by you—even if
16:22you don't know how to use Paint—to
16:23illustrating news articles or creating
16:25magazine covers as they did for
16:27Cosmopolitan. intelligence. We wanted
16:37to represent a powerful woman and we
16:39decided on an astronaut and I used
16:41DALL-E to generate options. Each time I
16:44adjusted my prompt over and over again,
16:46refining it to try to get the right
16:48image. And after many, many hours of
16:50trying hundreds of prompts, finally
16:52figured out the right one. It was a
16:55wide-angle shot from below of a female
16:58astronaut with an athletic feminine
17:00body walking with swagger towards
17:02camera on Mars in an infinite universe.
17:09You can generate ideas for
17:11illustrations for a client, or for an
17:14album or single cover, or even to fill
17:16space in a YouTube video as I did in
17:18the ant video, where several images you
17:21see were created by artificial
17:23intelligence, like this one, this one,
17:26this one, and this one. And even the
17:29video cover was made with DALL-E, or
17:31you can even print t-shirts with
17:33designs thought up by you but created
17:35by the computer, or print them and sell
17:37them as paintings. I insist, it is
17:40infinite and the only limit is your
17:42creativity and ingenuity. In short,
17:45DALL-E is like a being that knows how
17:47to draw, paint, design, and take photos
17:49according to what you tell it. Whatever
17:52you imagine, the computer interprets it
17:54. And I'll make a small parenthesis
17:56here to recommend this page, an online
17:58game where you have to try to
18:00differentiate which images were made by
18:01humans and which others by the computer
18:03. I'll leave the link in the
18:05description. Now then, let's move on
18:08from images and go to audio. What
18:10happens when you teach the computer the
18:13way someone talks? In that case, what
18:15happens is that the computer learns to
18:17talk like me. Of course, that opens up
18:20a whole lot of gray areas and unknowns.
18:22Is it my voice? Does it belong to me?
18:25Or is it simply the computer's
18:27interpretation of how I speak? Or I can
18:29make Luis Alberto Spinetta invite you
18:31to subscribe to this channel.
18:33Hello, being of light. How are you? My
18:36name is Luis Alberto Spinetta. I
18:38recommend that you subscribe right now.
18:40Thank you very much. I send you a big
18:43hug.
18:43Or we can try with Charly García.
18:45I totally agree with Luis.
18:47And what happens if instead of teaching
18:49the computer the tone of my voice, we
18:52teach it the way I modulate? In that
18:54case, the computer can learn it
18:56perfectly. What you are seeing is not
18:59Damián specifically, but rather the
19:01way I, the computer, imagine Damián
19:03speaks in different languages.
19:06If I want, I can give it a more
19:07Argentine tone,
19:08or if I want, I can speak languages
19:10that Damián never even heard in movies
19:12. What you just saw was done with a
19:26platform that I learned about from this
19:28DOT video, a video from which I also
19:30extracted a lot of clear information,
19:32and no, technically it is not me. To
19:36provide some context, Synthesia is a
19:38platform founded in 2017 by a group of
19:41artificial intelligence researchers
19:43with a clear goal: to allow anyone to
19:46make audiovisual content without
19:48cameras, microphones, or studios. That
19:51everything is done using artificial
19:53intelligence. And while it sounds crazy
19:56, it is quite real now. To achieve the
19:59avatar you saw, I just had to send them
20:01four one-minute clips of me speaking, a
20:03single recording. And I have an avatar
20:06to say and record whatever I want. You
20:08can also upload an audio file and use
20:11it. That means I can make you see this,
20:14while actually sitting very comfortably
20:16without having set up a single light or
20:18green screen needed to record something
20:20like this. None of my appearances in
20:23pajamas in this video were real, not
20:25even the introduction. Synthesia is a
20:28paid platform. Creating a custom avatar
20:30requires an extra fee, but with the
20:32base subscription you have access to
20:34over 30 actors and actresses to say and
20:36record whatever you want. It was the
20:39company in charge of doing some
20:40campaigns with Messi or David Beckham.
20:42This last one, in my opinion, is the
20:44most impressive. It was a campaign to
20:47raise awareness about malaria done in
20:49nine different languages. It is said to
20:57have killed more than half of the
20:58population that has ever existed. And I
21:11want to give a special mention to
21:12Synthesia, since they gifted me my
21:14avatar. If you are interested, I can
21:16make another video telling how I
21:18managed to convince them, since it was
21:20quite a fun job and it includes my fake
21:22mom trying to explain who I am.
21:31Can you imagine the future with all
21:33this? You can imagine it because we are
21:36already going through it and we can
21:38also imagine how all this will affect
21:40the development and consumption of
21:41audiovisual content. What would
21:44normally take hours and hours of
21:46editing, post-production, and rendering
21:49can now be done much faster. Everything
21:52we imagine can be understood by the
21:55computer and generate what is known as
21:57synthetic content: videos, images, text
22:00, and voices generated totally or
22:02partially by computers. With all this,
22:05the gap between the idea you have and
22:07the creation of the content is
22:08drastically reduced. It no longer
22:11matters if you don't know how to draw,
22:13animate, or if you are shy about
22:14speaking to a camera. It is estimated
22:16that in 10 years synthetic content will
22:18change the world. I, personally, am
22:21excited to imagine the huge changes in
22:23the creation and consumption of media
22:25that are approaching. Imagine the
22:27abysmal expansion of images that there
22:29can be in the stock image market.
22:31Landscapes that don't exist, settings
22:34that don't exist, and even humans that
22:36don't exist, which were imagined by the
22:38computer and can be put up for sale to
22:40be used by brands. All this at a much
22:43lower cost, since they don't have
22:45copyrights for technically not
22:47belonging to any person. We can already
22:49make different versions of the same
22:51video in different languages. We can
22:53already generate incredible images from
22:56just text. Imagine when we can write
22:58something and generate, for example,
23:00videos without actors, without a
23:02recording set, without moving to
23:04locations, et cetera, et cetera, et
23:07cetera. Again, hundreds of questions
23:09and it is perfect that it is so. What
23:12will be real? How much are we really
23:14going to teach computers? And to what
23:16extent will they learn? Will they
23:18eventually replace us in every aspect?
23:20If it learns from bad people, the data
23:22used to train or teach the algorithm
23:24usually carries a bias that the
23:26algorithm itself will reflect. For
23:29example, if you train a computer using
23:31text from the internet, gender or
23:33racist biases can easily slip in. A
23:36great example was Tay, the Twitter
23:39chatbot created by Microsoft in 2016,
23:41which became completely offensive,
23:43racist, and even a Holocaust denier in
23:46under 24 hours simply because it
23:48started learning from those it
23:50interacted with—the users. Another
23:54example is what happened with LaMDA,
23:55the most advanced conversational
23:57technology created by Google. You
24:00surely saw the news that an engineer
24:02was fired from Google for claiming that
24:04artificial intelligence had feelings.
24:07Those news reports claimed the computer
24:09said it felt used and even recognized
24:12being turned off as its own death. But
24:14we have to look at the story a little
24:16closer. The one who had the
24:18conversation with the artificial
24:20intelligence was engineer Blake Lemoine
24:22. He claimed the chatbot had feelings
24:24and even told him it wanted everyone to
24:26understand that it was a person. He
24:29published all the chat logs online,
24:30left the link in the description, and
24:32invited people to discuss whether or
24:34not LaMDA had feelings. Days later,
24:37Google suspended him for making the
24:39conversations public and violating the
24:41company's confidentiality policy, which
24:43caused even more of a stir. But the
24:46truth is that LaMDA repeats things
24:48found somewhere on the internet and
24:50makes decisions based on the datasets
24:52it was trained on. And considering that
24:55Blake Lemoine, the engineer who
24:57reported this, is a priest and even
24:59said in an interview that he is looking
25:01for God in artificial intelligence, the
25:03bias becomes clear. Saying an AI that
25:07repeats what it's taught is conscious
25:09is like saying a parrot is conscious
25:12and understands what it says just
25:14because it can repeat complex phrases.
25:17However, what is concerning is that an
25:20artificial intelligence can deceive a
25:22human brain and make it believe it has
25:24feelings. Once again, this sparks
25:27hundreds of questions. In an ideal
25:30future, artificial intelligence would
25:32be integrated in a way that
25:34significantly improves our quality of
25:36life. For example, it could help us
25:39manage our time better, as it could
25:41perform tasks for us like planning our
25:42meals or doing our shopping. It could
25:46also help us make better decisions or
25:48be more efficient and productive. In
25:51the worst-case scenario, it could
25:52threaten the very existence of humanity
25:55, decide that human beings are a threat
25:57to its existence, and take steps to
25:59eliminate us. In the meantime, we are
26:02left with the most fun part:
26:03experimenting, playing, informing, and
26:05asking ourselves questions. And we will
26:08only get the answers to those questions
26:10if we keep trying to understand all
26:12these new technologies. If we manage to
26:15use them well and link our neural
26:17networks to those of the computer, the
26:20possibilities are infinite. After all,
26:22this entire video is just that: a union
26:25between human and computer. Only time
26:28and our decisions will tell us what
26:30future awaits us. And since I couldn't
26:32think of how to end this video, I
26:34decided it would be better for the
26:36computer to do it. At the moment,
26:45artificial intelligence is in a
26:47learning phase, but it could soon
26:48develop to such a point that it exceeds
26:50our capabilities. This raises many
26:54ethical and moral questions, especially
26:57regarding privacy and security. Don't
27:07forget to subscribe to this channel for
27:10more useless stories.
27:17See you later.
27:19It's very short. Damián's channel is
27:21incredible. Everyone needs to subscribe
27:24urgently. Let's go Argentina. Yeah.

About this transcript

This page contains the full transcript of Historias Innecesarias: Inteligencia Artificial - NADA es REAL...o sí by Historias Innecesarias, generated from the public captions YouTube serves with the video. The transcript has 4,187 words across 639 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.