YT YouTube2Text

YouTube2Text → Transcripts → Experto en IA: La Gran Mentira de la Inteligencia Artificial | Dr. Ramón López de Mántaras

Experto en IA: La Gran Mentira de la Inteligencia Artificial | Dr. Ramón López de Mántaras — Transcript

by Vidas Ajenas · 12,564 words · 1,997 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00I have doubts, serious doubts that you
0:03can call something intelligent that
0:06simply gives the impression of being so
0:09. Behind it are very complicated
0:11algorithms developed by very brilliant
0:14and truly intelligent people. And yes,
0:17millions of people working in
0:19deplorable conditions to make
0:21artificial intelligence work. It is not
0:24harmless at all, as we are seeing, to
0:26believe, to believe that you are
0:29conversing or interacting with
0:30something intelligent and conscious. It
0:33is not harmless, quite the opposite.
0:35Headline: Open AI's ChatGPT, the latest
0:39version 5 point whatever, is now
0:41conscious. People believe it. Everyone
0:45believes it. My name is Ramón López
0:48de Mántaras Badía. I work in
0:51artificial intelligence. I do
0:55scientific research in artificial
0:57intelligence and I started working on
1:00this subject in 1975 with my doctoral
1:02thesis, so I’ve been in the field for
1:0550 years and I’ve seen its entire
1:07evolution, and today we’re going to
1:10talk about limits, where we are,
1:12dangers, etcetera. Thank you, Ramón,
1:28for being on Vidas Agenas. My pleasure.
1:32I'll get straight to the point.
1:33Artificial intelligence is intelligent,
1:37right? In the sense that we humans
1:41understand what it is to be intelligent
1:44, when we refer to our own intelligence
1:47or even the intelligence of many
1:49animals, which is increasingly
1:52surprising—the research by
1:54ethologists or people who work and
1:56investigate animal behavior—and
1:59surprising things are discovered about
2:02animals even very far from us on the
2:05evolutionary tree, like crows, for
2:08example. Well, no, in this sense they
2:12are not intelligent, meaning you can
2:16call them intelligent metaphorically if
2:20you want, or by cheating a little and
2:23defining intelligence in an ad hoc way
2:27so that the machine meets that
2:30definition. But well, I think terms and
2:34words are important, and they are what
2:36they are and have a big impact, right?
2:41And now, if they tell me: "Well, let's
2:43call artificial intelligence
2:44intelligent because when you interact
2:50with these systems, you realize that
2:53they give answers that are, well,
2:56acceptable or even good or persuasive
2:59or whatever you want, right?" Well, if
3:04you define it that way, if for some
3:06this is already a definition of
3:07intelligence, okay. But that’s up to
3:10them. But no, I have serious doubts
3:14that you can call something intelligent
3:18that simply gives the feeling of being
3:21so. This is very, very, very
3:23problematic, isn't it? The fact of the
3:26appearance, of the sensation. Uh,
3:29behind it are very complicated
3:32algorithms developed by very brilliant
3:35and truly intelligent people. That, yes
3:38, for sure. But they are based on
3:44statistical techniques, basically, and
3:45that's it. And well, if it's useful,
3:49well, if you can even make a business
3:51out of it, perfect, why not? No, but
3:55let's be clear and not try to pull a
3:58fast one or try to convince people that
4:01the machine is just as intelligent,
4:04equally as intelligent as us in every
4:07sense of the word, right? Because this
4:12can also create fears, right, in the
4:16general public.
4:19And on the other hand, to what extent
4:21is it artificial? Well, this is also an
4:25interesting question because we are
4:27seeing the latest trends in artificial
4:29intelligence, which are these so-called
4:32Chat GPTs, that everyone will know what
4:34I'm talking about. If I say Chat GPT,
4:37of course, it's like a magic word, it's
4:40like a wildcard, right? The famous
4:45know-it-alls at birthday meals and such
4:49. They talk about ChatGPT, for example,
4:54I mean, it has reached everyone,
4:55everyone knows what it is, right? But
4:57well, generative artificial
4:58intelligence, which is what it's called
5:00, is much more than this. There are
5:01others, there are others, there are
5:02others, right? But let's focus on this
5:04type because it's what people know most
5:06about these chatbots, which they also
5:08call conversational, right? You can ask
5:12questions and such, or "summarize this
5:14for me," things that people do, or even
5:17"write me an email," right? Or write a
5:21report about it, right? Let's see if
5:24they work. Well, they do work, and you
5:27have to acknowledge that they work well
5:30as long as you know a lot about that
5:32subject and are capable of making a
5:35critical analysis of what they tell you
5:37, because factual errors—what some
5:40call hallucinations, which is another
5:43term I don't like to use, but well,
5:45because it's very anthropomorphic—
5:48they're just errors, that's it. Errors,
5:51because they are based on probabilistic
5:53techniques and they make errors; they
5:56don't preserve the truth, right? So,
5:59for them to work even minimally well,
6:02what many people don't know is that
6:04there are millions—and I don't say
6:07millions just to say anything, it's
6:09proven. Millions of people who make,
6:13refine, and train these systems like
6:16ChatGPT, right? In the Philippines
6:19alone, it is estimated that there are
6:22at least two million young people, and
6:24millions more in Kenya, Pakistan, and
6:26India—generally in the global south
6:28—being exploited; pure and simple
6:31exploitation.
6:32Cheap labor.
6:33Yes, extremely cheap labor, 50 or 60
6:35cents an hour, and that’s when they
6:37actually get paid, because if things
6:39aren't done just right, sometimes they
6:41don't even get paid since there's no
6:43contract or anything. They are often
6:47paid through PayPal and things like
6:50that, right? These systems, right? They
6:53never see, they can never claim
6:55anything from anyone because they have
6:56no access to anyone, right? These are
7:01people, these are mostly very young
7:03people. They fill the internet cafes.
7:08The Post published a photo, a
7:10front-page headline of one of these
7:12internet cafes in the Philippines,
7:14packed with these young people at
7:16computers because they don't even have
7:18the financial means to have their own
7:20computer or laptop at home, or to pay
7:22for an internet connection, so they
7:24have to go to internet cafes. And what
7:28are they doing? Labeling images for
7:31autonomous cars, for example, right? So
7:34, this is a tree, this is a pedestrian;
7:37they are continuously labeling the
7:39images shown to them by the company or
7:42whoever it may be, right? And then they
7:45also show them answers from ChatGPT and
7:48other systems, and they have to tell us
7:50if the answer is correct or not. Often,
7:53they are presented with several
7:56alternative answers to the questions
8:00previously collected from people, right
8:03? When we use these systems, they
8:07collect all of this, they also collect
8:09the thumbs up and thumbs down when you
8:11say okay or not, and based on all this
8:13information, they make improvements so
8:15that when they release the next version
8:17, it works better. But it’s people
8:21who are doing it. This is not an
8:23artificial intelligence, let’s say,
8:26that is capable of analyzing anything
8:28and learning on its own. So, this is
8:32the job of these people, right?
8:35Moderating content, labeling images,
8:38and evaluating or giving preferences on
8:41which answers are more suitable than
8:44others. For the multiple questions that
8:47hundreds of millions of people are
8:50asking, that we are asking every day,
8:52to these language models. So, it
8:54doesn't improve in real time; it's not
8:57like you give a thumbs up and the
9:00algorithm immediately takes your
9:02opinion into account—that's not how
9:05it works at all. Offline, on the
9:08outside, meaning not in real time. It's
9:11these millions of young people, guys
9:13and girls, who take all of this, all
9:15these inputs, into account. That's why
9:19there have to be millions, because
9:22there is so much to process manually,
9:24and they improve for the next version,
9:27right? A startup. Now I'm going to tell
9:34another example, which is a very good
9:36example of how it's not as artificial
9:38as it seems, and in the example I'm
9:40going to tell now, it's not artificial
9:42at all. One thing is having human
9:46collaboration to improve it, but there
9:48really are algorithms working there
9:50when you consult with these bots, these
9:52chatbots, right? An artificial
9:57intelligence company that had secured
10:0050 million euros in investment from a
10:03large tech company and......it was to
10:09help programmers program. Well, this is
10:11one of the typical applications they
10:13are making now. Instead of chatting or
10:16playing a little, it's being used for
10:18more professional things, right? And
10:21one of them is helping you code, write
10:24code in a programming language, right?
10:28What a software engineer, a programmer,
10:30does, right? Well, the programmers who
10:33used this tool would go and say, "Well,
10:36I need a routine, a program, a software
10:38that does this and that and the other
10:41to solve this type of problem." Okay,
10:43let's go. After a while—it wasn't
10:46instantaneous—they would receive code
10:48, right? I mean, software, a program.
10:52Many times it didn't work. It must be
10:55said, but what happened is that it came
10:58to light and the company had to close;
11:00it had to shut down and is now accused
11:03of having defrauded its investors, and
11:06it turns out there was no artificial
11:09intelligence or anything, just 700
11:11programmers in India. Humans were the
11:16ones receiving these requests, making
11:19the programs, sending them to the
11:22company, and from there it went to the
11:25user who had made the request; 700
11:28programmers working non-stop to do
11:31these programming tasks, and there are
11:34plenty more examples like that, right?
11:40Autonomous cars in San Francisco that
11:41would get stuck every two or three
11:43blocks and wouldn't know what to do. A
11:46somewhat unexpected situation, right?
11:48That hadn't been foreseen in the
11:50training set. Well, you can't foresee
11:51everything, uh, you can't foresee
11:53everything. It's impossible. That's
11:55called the long tail problem. The long
11:58tail of a probability distribution is
12:00the likelihood. Imagine it has a shape
12:03that goes up, then down. Well, if it's
12:06a symmetric Gaussian, it's what they
12:08call the normal function, uh; but there
12:11are distributions that have, well, even
12:13multimodal shapes, with several peaks.
12:16But usually there is that, there is a
12:18long tail which are very unlikely
12:21things that don't, that don't happen,
12:23statistically speaking, they don't
12:25happen often.
12:27But they could happen,
12:28but they can occur, they aren't
12:29impossible. "Long tail problem" is what
12:32it's called in statistics, very well
12:34known, right? Well, when that which is
12:37unlikely to happen occurs, that's when
12:40artificial intelligence goes blank. So,
12:44okay, for example, a newspaper on a
12:47very windy day flying, flying past the
12:50windshield. A sudden hard brake that
12:54even depending on the speed the
12:55autonomous car is going, could even be
12:57a problem for those inside, right? No
13:01matter how much you wear a seatbelt,
13:02you might still get a mark here from
13:03the belt due to the sudden braking,
13:05right? Or an injury, right? Uh, that's
13:09an example of something that when these
13:12algorithms are trained, the software
13:15that pilots an autonomous car, well,
13:17they aren't taken into account. You
13:21can't take into account everything,
13:22absolutely everything that can be
13:23encountered in the real world. That is
13:25impossible. Uh, another very good
13:28anecdote is the kangaroos in Australia.
13:30Of course, these are usually trained in
13:35the United States, in Europe, and there
13:37are no kangaroos. So, the training set
13:39didn't have images of kangaroos jumping
13:41in front of the car. So, it didn't
13:44understand, it couldn't identify that,
13:46it couldn't identify that and, well,
13:47they got stuck. I mean, there are many
13:50reasons why an autonomous car can get
13:52blocked and just stay there completely
13:55stunned, frozen, doing nothing. Well,
13:59Waymo and Cruise, the two companies
14:01that have robotaxis in San Francisco,
14:03Los Angeles, Phoenix, and Austin, find
14:06that on average every 2 or 3 km, the
14:08car has to have intervention from a
14:10human teleoperator, and they didn't
14:13mention that. Yes, there are
14:16teleoperators who then get the car out
14:18of this blocked situation because
14:20otherwise it would stay there forever,
14:22right? It's like playing with a
14:25remote-controlled toy car, right? Of
14:28course, because the car sends images,
14:30they obviously see them, they see the
14:32surroundings because there are cameras
14:34pointing everywhere on the cars, many
14:36sensors. What the human teleoperator
14:39does is manually move the car and get
14:42it out. It's not that the car does it
14:44autonomously. Well, this also caused a
14:47scandal when this problem was uncovered
14:49in San Francisco, right? I mean, there
14:54are examples where there is a huge
14:57human intervention that they were
15:00hiding. There are so many, to such an
15:04extent that the Harvard Business Review
15:08also denounced it in an article that
15:11said, "The People Behind the AI Curtain
15:14." The Washington Post, that portal I
15:17told you about in the Philippines, it
15:19said, it didn't say artificial
15:21intelligence, but the truth about
15:22artificial intelligence: millions of
15:24people working in deplorable conditions
15:27so that artificial intelligence works.
15:30So, be very careful with what they try
15:33to sell us, the smoke and mirrors that
15:37many of these companies try to sell. We
15:41will come back to this later because I
15:43would like to talk a little more in
15:45depth
15:46and in detail, but I was thinking that
15:48perhaps there are people who think that
15:50this AI and artificial intelligence
15:52thing is something new, but research in
15:54artificial intelligence has been going
15:56on for decades. My question,
15:59I've been at it for 50 years, since '75
16:01when I started my thesis with a
16:03learning algorithm where an
16:04anthropomorphic robot with an
16:05anthropomorphic hand grabbed objects
16:07and learned; an algorithm learned to
16:09discriminate between objects of
16:11different shapes according to their
16:12form with an artificial skin that we
16:14called this, be careful. They are
16:17tactile sensors, the palm of the hand,
16:20the fingers covered in tactile sensors,
16:22some systems we could measure the angle
16:25of the phalanges according to this
16:27angle, this angle, and this angle,
16:29right? And with all this information,
16:32my doctoral thesis was an algorithm so
16:35that this robot with this
16:36anthropomorphic hand, that is, very
16:38similar to a human one,
16:40It was a marvel, mechanically speaking.
16:44I did it in France, in Toulouse, I read
16:46and presented the thesis in '77 and I
16:48started working on this, with this
16:50anthropomorphic hand, in '75, 50 years
16:53ago, so it's older; I didn't start with
16:55artificial intelligence, it started in
16:57'56, I started 18 years later. But yes,
17:01of course, it's been many years.
17:04I wanted to ask you what the ultimate
17:06goal of research in artificial
17:08intelligence is.
17:11I'll tell you what it was in '56, I'll
17:14tell you what it was in the 60s, what
17:16it was in the 70s, even the 80s, and
17:19even relatively recently, at least up
17:21until 2010 or 2012. Well, the objective
17:25, scientifically speaking, here we have
17:28to be precise, eh, one thing is a
17:30purely business objective, right?,of a
17:32company or what might the objective be,
17:34what could it be? Well, making money is
17:37just fine, eh? I'm not criticizing it.
17:39Companies have to make a profit,
17:41otherwise they have to close, right?
17:44The objective, scientifically speaking,
17:48is, eh, to see if it is possible for
17:52something that isn't biological, that
17:56isn't based on carbon chemistry like us
17:59, that isn't organic, that doesn't have
18:03cells, that isn't a living thing, can
18:07become intelligent in the human sense
18:10of being intelligent. Not with a trick
18:14definition, as I was saying before,
18:15right? This is a...in '56 they believed
18:19this was possible and that it wouldn't
18:21take too many years to achieve it,
18:23right? And look, we are still at this
18:28question, this great, this great
18:30scientific goal, achieving truly, truly
18:33intelligent intelligences, right? Eh,
18:37in the human sense of what is
18:38intelligent, of what an intelligent
18:40animal is, it hasn't been achieved.
18:43That is the scientific goal. Now, now
18:45that for a few years artificial
18:47intelligence has been in the hands of
18:50the "magnificent seven," as they call
18:52the seven big companies, right? Google,
18:56Amazon, we all know which ones they are
18:59, right? Meta, Facebook, Microsoft,
19:03Nvidia. Ah, I'm no longer very clear on
19:07what the goal of artificial
19:09intelligence is, apart from making
19:11money.
19:13And there was a distinction between
19:14general artificial intelligence and
19:16strong artificial intelligence.
19:18Of course. Yes. Let's see, because here
19:20there is...yes, this needs to be
19:21clarified because there is a lot of
19:23confusion too. Artificial intelligence
19:25up until now is what is called specific
19:28artificial intelligence. In English,
19:30it's called "narrow." Narrow, to
19:32distinguish it from general, right? I
19:36prefer "specific," because it's...I
19:37think it's a more appropriate antonym
19:39for general, but well. Which is the one
19:43you design to do one task very well,
19:46and the task is...you can define the
19:48task relatively well and get the
19:50digital intelligence to do it, not just
19:53well, but 1,000 times better than you.
19:57Playing chess, it's very clear. There
20:00hasn't been anything to do against
20:02software and systems that play chess
20:04for years. Everything is over here,
20:07right? Now with Go, with AlphaGo, it
20:10seems so too, although lately, there
20:11has been some human who has beaten the
20:14machine. That, that's curious, huh?
20:18When big news comes out, AlphaGo beats
20:21......Lee or Sedol or these great Go
20:27players. Wow, headlines all over the
20:30world. But it turns out there was a
20:33player, and not really a very good one,
20:35but because he did strange things, he
20:37completely confused the machine and the
20:40human won. Almost nobody heard about
20:41this. Sure, let's close this
20:43parenthesis, right? Eh, but let's see,
20:46these are all different specific
20:48intelligences, and like the saying goes
20:50, "he who grasps at too much loses
20:52everything," right? Well, on the
20:54contrary, if you grasp at little, you
20:56can grip hard, you can delve deep. And
20:59this is the key to the successes of
21:01artificial intelligence until recently:
21:03they were all super-specific, they did
21:05one single task well. It's clear: this
21:09specific or narrow versus general,
21:12which for me is multitasking. I mean,
21:15it's the opposite of this; that it's
21:17one artificial intelligence, not 1,000
21:19artificial intelligences that do 1,000
21:21tasks and you package them all together
21:23, no, that's not it. With a common
21:25architecture, a software architecture,
21:30a common cognitive architecture, that
21:32artificial intelligence would really
21:35have the capacity to use knowledge in
21:37one subject to transfer it to another
21:39—it's called transfer learning—that
21:41it truly learns to understand how the
21:44world works, learning cause-and-effect
21:46relationships. What happens if I push
21:49an object? What happens if I grab it?
21:51Eh, if the glass is full of water, I
21:55anticipate, because I have common sense
21:58, that it will weigh more than if the
22:00glass is empty, so I will apply a
22:01different amount of force, okay? Or if
22:05a glass is fragile because it's made of
22:07a material that breaks easily, I will
22:09possibly handle it with a little more
22:11care than if it were made of plastic,
22:14let's say, because the consequences of
22:16dropping it are not the same. So, all
22:19this type of knowledge about the world,
22:22about how the world around us works,
22:24which we learn by interacting—all of
22:27this, the day all of this is achieved
22:29and implemented in an AI, we could
22:31start talking about artificial general
22:33intelligence, that it knows the
22:35consequences of its actions. We're not
22:39talking about consciousness yet here,
22:40eh, careful, now he's going to bring up
22:42the strong one, right? But they perform
22:45many tasks, many tasks, not just one,
22:48and they do them quite well, right?
22:51Basically well, some better than others
22:53, but well, right? That is the general
22:55one that doesn't exist yet. That
22:58generative one, also called general
23:00purpose, right? No, notice that it's
23:04called general purpose, which some say
23:07is the stage toward the other one,
23:10right? I don't think so, but well, we
23:13can talk about that later if you want.
23:15Ah, that's the general one, and the
23:17strong one is the one that has mental
23:18states, consciousness, emotions,
23:20everything, everything. We, we are
23:22strong artificial intelligences. There
23:25is no doubt. There is no doubt. The
23:28strong AI would really be this one,
23:31proposed by a very renowned philosopher
23:33, John Searle, right? He did a very,
23:37very interesting experiment to
23:39demonstrate that if you are only
23:42capable of processing symbols in a
23:44machine, which would be processing ones
23:46and zeros, basically the binary
23:48language that is at the base of the
23:51machine, right? If it only processes
23:56symbols and nothing else. According to
24:00this experiment he proposed—a thought
24:02experiment, it wasn't actually
24:04performed—this philosopher, you
24:08cannot have mental states, nor can you
24:10have consciousness or anything of the
24:12sort. I mean, the strong one goes much
24:14further than the ability to perform
24:16symbolic calculations. Machines perform
24:19symbolic calculations and nothing more.
24:22So, based on this idea, strong
24:23artificial intelligence would not be
24:25possible. I am one of those who
24:28believes that John Searle was right,
24:29that it is not possible,
24:31that the strong one is not possible
24:34with current technology; maybe with
24:38technologies that don't currently exist
24:42or are very incipient, we could think
24:46that something not born of a mother
24:49could be conscious. We have to think
24:53about the replicants in Blade Runner.
24:55The replicants in Blade Runner are not
24:57software running on hardware. They are
25:02beings based on carbon chemistry. They
25:04have real cells, they don't have
25:05artificial neurons, they have real
25:07neurons. They are exactly the same as a
25:10human, but they weren't born of a
25:12mother. This is the only difference. So
25:15, thinking that something that is alive
25:18and is just like you, just like me,
25:20could eventually develop consciousness,
25:23there I can admit it. But with
25:26technology based on the service, a
25:28transistor that conducts or doesn't
25:30conduct, which is essentially a one or
25:32a zero—if you like, it's a bit more
25:33complicated than that, but just to
25:35illustrate it, right? And on top of
25:38that, there's software programmed by
25:40people in whatever programming language
25:42you prefer, it's the same, it doesn't
25:44matter. To think that this can be
25:47conscious, to be aware that it's alive,
25:51to have emotions, feelings, to feel
25:54empathy—well, having consciousness
25:57and intentionality implies all of this.
26:02To me, it's utter nonsense to say that
26:04with the technology current artificial
26:07intelligence is based on, you can
26:10create or have strong artificial
26:12intelligence. I rule it out completely.
26:15General multitasking. Maybe. Yes, yes.
26:18I admit this. I admit that we might
26:21reach it in the medium-to-long term; in
26:24the very short term, a year or two out,
26:27I don't think so, but it's possible to
26:30achieve artificial intelligences that
26:32are much more general, practically what
26:35we could already call general. It can't
26:38be ruled out. It can't be ruled out.
26:42Before continuing with the subjectivity
26:43of artificial intelligence, I would
26:45like to do a quick review of the things
26:47it is actually contributing to. In your
26:49book, "100 Things You Need to Know
26:51About Artificial Intelligence." There
26:52are dozens of examples in which
26:54there are like 60 examples or more. Yes
26:55,
26:55there are many. In health, science, and
26:57the environment.
26:58Exactly. Yes.
26:59We can pause on a couple of those
27:01examples. Yes, of course, as one is
27:04obviously—I always tell it when I
27:07talk about the great successes of AI
27:09because, fortunately, even though I am
27:12very critical, if I didn't see what it
27:15can contribute and that it provides
27:18many benefits, I wouldn't be doing what
27:20I do. I would be doing something else,
27:24right? Which, by the way, won the Nobel
27:30Prize in Chemistry for Demis Hassabis
27:33and John Jumper, who is a colleague of
27:37Hassabis. It's the folding problem, of
27:41how a protein folds in
27:43three-dimensional space, okay? Proteins
27:46are chains of amino acids and they fold
27:50in very capricious ways. It is very
27:53complex, and predicting how a protein
27:57folds was, until recently, very
28:00difficult, very difficult.
28:02Crystallographic techniques were used,
28:05which they love, and what they did was
28:08to visualize it, right? Once folded,
28:10and this was very expensive, right? I
28:16believe that in the body, in humans, we
28:19have I don't know how many hundreds of
28:21thousands of proteins, I don't remember
28:24, but in terms of all chemistry, eh,
28:26we're talking about many millions and,
28:28of course, there was a small percentage
28:31of proteins, relatively small, whose
28:33structure was known. This is important
28:38because when the function, what the
28:40protein does, its function depends on
28:43how it folds. It is very important how
28:47it folds for what it does, how it
28:48functions later, right? That is to say,
28:51its functionality. And thanks to the
28:53fact that now, thanks to this AlphaFold
28:56—the software is called AlphaFold—
28:58it solves this problem, although it
29:01still makes many mistakes. But well,
29:03basically a very high percentage is
29:06correct; the software's predictions are
29:09verified and are correct. Eh, well, it
29:12facilitates, eh, the design and
29:15synthesis of new drugs, even vaccines,
29:18right? And they have made it freely
29:22available, I mean, open, right? I mean,
29:25open to the entire community of
29:27molecular biologists and biologists who
29:30want to use this software, right? This
29:34is something I tell in one of the
29:37chapters and, for me, it is undoubtedly
29:40the greatest result of artificial
29:42intelligence in recent years. It has
29:46nothing to do with models like ChatGPT
29:48and all that. Eh, this is another
29:50interesting thing to point out: it is
29:52not generative AI. They have in common
29:55that they are artificial neural
29:56networks, yes, but they are
29:57architectures that have nothing to do
29:58with each other. Language models are
30:02based on a concept called a neural
30:04network known as a transformer and a
30:06thing called vector embedding, which is
30:08transforming words into vectors of
30:10thousands of dimensions, very long
30:12strings of numbers; it has practically
30:14nothing to do with this, they are
30:15different things. It has 32 integrated
30:19algorithms working in cooperation in
30:21AlphaFold. 32 algorithms, eh, of
30:25various types with various artificial
30:28intelligence techniques. A magnificent
30:31example of how the great success of
30:33artificial intelligence in recent years
30:35is precisely not chat GPTs and
30:37companies, but rather AlphaFold. That
30:40is an example I really like to talk
30:41about. Then there are many others,
30:43right? To change fields, instead of
30:47medicine, or moving to a completely
30:49different field, well, it has to do
30:52with topics, for example...I have many
30:54chapters on climate change, like
30:57monitoring the evolution of glaciers,
31:02avoiding having to solve, eh, super
31:06complex equations of fluid mechanics.
31:12that required enormous computing power
31:16to solve these systems of equations
31:19that allowed us to see what was
31:22happening, to finely monitor the
31:26evolution of a glacier, right? Now,
31:30there is an artificial technique that,
31:34by sensorizing the glaciers in a way,
31:37artificial intelligence can manage to
31:40monitor, that is to say, and predict
31:43the speed at which the glacier will
31:46evolve and how it will evolve. It's a
31:49very good example, too, right? Then
31:51there is another one that is very
31:52important because it has to do with
31:54energy, which is one of the problems we
31:56have in the world, right? Uh, everyone
32:02has heard of nuclear energy, obviously
32:04not everyone knows what a nuclear power
32:07plant is. The nuclear plants we have
32:10now are fission-based. Atoms are
32:14bombarded with neutrons, the atom is
32:17split, okay? And as a consequence of
32:20this splitting of the atom, there is a
32:22variation in mass and, applying
32:24Einstein's equation, energy equals mass
32:26times the speed of light squared, this
32:28variation in mass, by splitting...eh,
32:30and this causes a chain reaction, eh,
32:33by the way, eh, controlled, fortunately
32:35, otherwise they would be atomic bombs.
32:37If they are not controlled, they are
32:38atomic bombs.
32:39I mean, a nuclear plant and an atomic
32:41bomb are essentially the same thing.
32:44The difference is that in a nuclear
32:46plant you control the chain reaction
32:48and, on the other hand, in the bomb,
32:50you don't. Well, we already know the
32:54problem of waste, of nuclear waste,
32:57right? Because radioactive materials
33:00are used. The raw material is
33:02radioactive. But there is something
33:04else called nuclear fusion, right?
33:07Fusion. Instead of splitting an atom in
33:11two, two light atoms, usually isotopes
33:14of hydrogen like deuterium and tritium,
33:17are forced to fuse, and the same thing
33:20happens there, but in reverse; there is
33:23also a variation in mass between before
33:26and after the fusion, and with this
33:29variation in mass, the famous Einstein
33:32equation, you generate a brutal amount
33:35of energy. It has a great advantage.
33:38You are not fusing atoms of any
33:40radioactive material there. It's
33:43hydrogen and its, well, the isotopes of
33:46hydrogen. It's fine, it doesn't have
33:49nuclear waste, it's very clean. It is
33:52not. So, the problem with these fusion
33:56reactors is that they work at, I don't
34:02remember now, tens of thousands of
34:04degrees. Well, the sun, the sun, our
34:08star, what the sun does is nuclear
34:10fusion continuously. Inside the sun,
34:13there is nuclear fusion. It is that,
34:16nucleosynthesis, right? What we want
34:18is that. I mean, what we want with
34:20nuclear energy would be like having a
34:23sun, in quotes, confined and controlled
34:26, right? So it generates energy for us
34:30as we need it. But since these are so
34:33intense, they cannot be put inside a
34:36container because any material would
34:38melt. So, how is it achieved, this beam
34:42of—this is, it’s called plasma?
34:46Because it’s hydrogen and its parts
34:49are ionized, meaning they respond to
34:52magnetic signals; being ionized, they
34:55have electrical properties despite
34:58being a gas, right? So, what is
35:03achieved is using very powerful
35:05electromagnets to keep this plasma
35:08spinning inside a device, without it
35:11ever touching the walls, because
35:14everything would melt, right? Due to
35:17the incredibly high temperatures there.
35:21So it is very important to control the
35:23plasma flow so it always goes through
35:25the center and never approaches the
35:26walls. These fusion reactors are
35:29toroidal, usually donut-shaped, uh,
35:31that toroidal shape, but a big donut,
35:33of course. And they are spinning around
35:38there, and they are encouraging atoms
35:41to fuse. Uh, before it’s stable, you
35:44have to maintain it for a certain
35:45amount of time, right? Okay, and you
35:49also have to finely control the shape
35:52of the beam, uh, the shape, I mean,
35:55it’s not very clean, it can zig-zag,
35:58but well, that’s what the shape of
36:02the beam is controlled with—
36:04electromagnets. Until now, these
36:08electromagnets were very difficult to
36:10control. Each electromagnet had a
36:12controller, and coordinating everything
36:14was enormously difficult. The
36:16experiments always failed. Now an
36:18artificial intelligence is capable of
36:21coordinating all the magnets at once,
36:23all the electromagnets at once, and
36:26with a very fine hand, right? With
36:28classic controllers from something
36:30called automation, from control theory,
36:33which is what was used before. Now with
36:36artificial intelligence, you can
36:38control all the electromagnets
36:40simultaneously and with great precision
36:43, which allows the plasma flow to be
36:45maintained for much longer. We are
36:48still talking about a few seconds or
36:50even fractions of a second, meaning we
36:53are a long way off from having an
36:54operational nuclear power plant and
36:56electricity reaching our homes from
36:58nuclear fusion. That's still, wow,
37:00light-years away. But one of the
37:03important recent advances is this
37:06chapter of the book where I talk about
37:09the control of a nuclear fusion reactor
37:12.
37:14And now, thank you. It's just that I
37:17think a discourse that's too critical
37:19is leaving aside the contributions it
37:21is actually making, right?
37:23Sure, sure. That's why, yes, we're
37:25talking about this now. Great
37:26contributions. Yes, yes. So, now I
37:29would like to move on to that part I
37:30call subjectivity, but it's a bit about
37:32the brain and consciousness.
37:34Sure, it’s basically that. Yes.
37:35Do you think the brain is a replicable
37:38machine?
37:39Well, let's see, replicable, in what
37:41sense replicable? Within a, I mean, to
37:43make an upload of the brain or the mind
37:46into a computer. What do you mean by
37:48that? The singularity, no,
37:50that the functioning of the brain from
37:52carbon chemistry can be translated to
37:55the functioning of a neural network.
37:57Ah, artificial.
37:58Artificial. Ah, okay, okay. Yes, I
38:00understand you now. I mean, to make a
38:03copy, in great detail, down to the
38:06finest level of detail of how the
38:08biological neuron works and use this
38:11knowledge to develop a...Well, in my
38:14opinion, there is a problem to begin
38:17with: for this to be done, is that in
38:20the brain (we now know, or the human
38:23nervous system, the peripheral, and
38:25even the brain) it is much more than
38:28just the brain, for starters, and there
38:33are some cells, which there are many
38:36more of than neurons, called glial
38:38cells. Until recently, it wasn't very
38:43clear what the functionality of glial
38:46cells was. Well, it's a kind of, they
38:51control neurons, they are at a
38:53meta-level and they control, but it
38:55wasn't very clear, but in recent years
38:58it is known that they play a
39:00fundamental role in information
39:02processing by the brain. Artificial
39:06networks don't even come close to
39:09modeling glial cells, for example.
39:13There are other things they don't model
39:15. All chemical activity is left out of
39:17artificial neural networks. They
39:20approximate the electrical activity
39:22more or less well, right, of when the
39:25neuron fires and the connection with
39:27other neurons. They adjust the weights
39:31between neurons, which is a replica,
39:33well, an approximation of how
39:36biological neurons fire; what is called
39:38firing together, which means they are
39:41collaborating, cooperating on a task,
39:44right? And so now it is known through
39:47techniques that look, they can see
39:49images of the brain, right? Magnetic or
39:53others, one can see with quite a lot of
39:56precision depending on the task a
39:58person is doing, the task, task would
40:00be the word in Spanish, from playing
40:04music improvised or playing with a
40:06score, which is a very nice example.
40:09The difference between improvising or
40:11not improvising music, you can see
40:13which parts of the brain are the most
40:15active or not and which neurons fire,
40:17fire in a group, right? groups of
40:19neurons that are synchronized, that is
40:21the core, right? And this divided into
40:23this is simulated or emulated
40:25mathematically with the weights between
40:27the neurons, right? Of artificial ones.
40:29Agreed? Well, the electrical part we
40:32see is what is being done, but the
40:35brain is of such immense complexity
40:37that we still don't even know. The
40:41brain is still a great unknown. We know
40:44very little. We actually know more than
40:4610 years ago, we know more every time.
40:49But there is much left to discover. So,
40:52to think that this complexity, which is
40:54the result of millions of years of
40:56evolution, can be, as they say,
40:58snapshotted or looked at at a
41:00microscopic or nanoscopic level to see
41:03what is happening in there, and to
41:05think that you can then reproduce this
41:07digitally. I find it very hard to
41:10believe. Besides, there is a lot of
41:13analog activity, not just digital in
41:15the brain. Yes, above a certain
41:18threshold, boom, the biological neuron
41:20fires. That does resemble the
41:24artificial one, but nothing more than
41:26that. It is a very crude approximation
41:29because there is much more; the entire
41:31chemical side is not considered. In
41:34here, we have processing based on
41:37electrical and chemical themes, not
41:40just electrical, it is electrochemical,
41:43right? So, well, the neurotransmitters,
41:47in short, there is a huge amount of
41:50complexity that is unknown, and even if
41:53we were to know it one day in the
41:56smallest detail, could we replicate
41:59this brain digitally? I also have my
42:03doubts. Well, I also have my serious
42:05doubts because, what do you do? Okay,
42:07you replicate the whole brain, the mind
42:09, but what about all the experiences,
42:11where do you put those? How do you
42:14somehow decode the language the brain
42:17uses to then apply it to, I don't know,
42:21to a robot, a robot that is based on
42:24this artificial brain? Uh, biologists
42:28and neuroscientists often laugh quite a
42:31bit when computer scientists or
42:34artificial intelligence experts or
42:37proponents of the singularity like Ray
42:40Kurzweil and others talk about how
42:43we're going to replicate the brain.
42:46They say, "They don't know what they're
42:48talking about, they don't realize the
42:50enormous complexity of the brain and,
42:52besides,
42:52one does very different things and
42:54it also interacts with a body. Right,
42:56exactly. That's where the" brain in a
42:59vat "thought experiment comes in, right
43:01? The brain in a test tube. It's a
43:04philosopher, what was his name? Hilary
43:07Putnam. Hilary Putnam, I think, said
43:10that and proved that it's absurd
43:13because you need the body. Exactly.
43:17Well, this suggests that Putnam's idea
43:19was what the Wachowskis, who made
43:22Matrix, based themselves on. I mean,
43:25Matrix is" brain in a vat, "it's the
43:28brain in a test tube. So, whoever has
43:32seen the movie will remember that there
43:34are the humans, the real ones, the real
43:36ones, not the ones in the simulated
43:38environment, they are there in a kind
43:40of, like a placenta, inside, it looks
43:42like they are inside a, a, a placenta
43:44or something, and they are hooked up
43:46with cables, right? Right. These are
43:50the real humans who are there, they are
43:52there, right? Motionless, right? And
43:54there are all these cables coming out
43:56of their brains and these cables
43:59receive all the information from the
44:02Matrix, which is this software, this
44:05great simulation of the world, right?
44:09And they are there but they don't do
44:12anything, well, these brains are
44:15supposed to not, not distinguish
44:18from reality.
44:18They don't distinguish from reality. I
44:20mean, well, this is the movie. Very
44:22good, it's a great movie, right? But
44:24the" brain in a vat "thing doesn't make
44:27any sense, because of that, because the
44:30brain itself, it's fine, it's very
44:32important, but it's not enough to
44:34explain everything that is human
44:37intelligence and how we function,
44:39mental states, and all that. We need a
44:42body. Without a body, we're in trouble.
44:47Uh, maybe we've gone too far and I
44:49don't know if it's because of the
44:50language. Look, an example. Yesterday I
44:52was using ChatGPT,
44:54yes. And I don't know, I don't remember
44:56what I asked it, but the response said"
44:58thinking. "
45:01Thinking.
45:02And it's
45:02yes.
45:03That's humanizing it too much, right?
45:04Completely.
45:05It's not thinking, it's calculating.
45:07Right.
45:08It's calculating. Yes, it is definitely
45:09calculating. It's performing a
45:12multitude of operations with
45:14multidimensional vectors, matrix
45:16operations. It's operating with real
45:18numbers. It's dealing with numbers. I
45:22don't know if we've gone too far by
45:24claiming that generative AI reasons.
45:28Of course, it doesn't reason, obviously
45:29.
45:30And that, at some point, we'll just
45:32wake up one day and have generated a
45:34consciousness.
45:36Right, that's what some are saying,
45:38isn't it? Dario Amodei from Anthropic
45:40or Ilya Sutskever, Hinton himself,
45:45Geoffrey Hinton. It surprises me that
45:48Hinton says it, but well, he's the
45:50Nobel Prize winner in physics. That
45:53consciousness is, well, just around the
45:55corner, right? That these machines,
45:58it's clear that if they aren't
45:59conscious, they're almost conscious
46:01already. I don't fully understand why
46:07they say this. I understand those who
46:10say it for business interests. I
46:12understand Dario Amodei saying it, I
46:14understand Sam Altman saying it, I
46:16understand Ilya Sutskever saying it,
46:18because, of course, they need
46:20multi-billion dollar investments since
46:22they don't make any money. None of
46:24these companies are profitable. They
46:26burn through billions. I read that
46:32OpenAI was burning around 5 billion a
46:35year, and their revenue, the income
46:39they receive, is tiny. It's tiny. So,
46:43what do they receive? Subsidies on one
46:46hand, but on the other, venture
46:48capitalists who have put in a lot of
46:51money, right? Why have they put in so
46:55much money? Because, obviously, what
46:58they sell them, what they tell them, is
47:00that artificial general intelligence—
47:03maybe even conscious and
47:05superintelligent—is just a matter of
47:07time; that they should be patient
47:10because we're almost there, and when we
47:12have it, they'll make money they won't
47:15even believe. What they call ROI in
47:19English, return on investment, they
47:21promise an astronomical return on
47:23investment, they convince them. They
47:26are very skillful at convincing. These
47:28people are excellent salespeople,
47:31excellent salespeople, they're young
47:34people. I don't know, I assume they
47:38have some knowledge of computer science
47:40. I hope so, although many of them are
47:43dropouts, quite a few who haven't
47:45finished their studies. I don't say
47:48that because they aren't intelligent,
47:50but by choice, they surely saw there
47:52was a lot of money to be made quickly
47:54by deceiving people, right? So they did
47:56it, didn't they? They are certainly
47:58smart, and they are certainly skilled,
48:01but many of them lack much training in
48:03areas like neuroscience, philosophy,
48:06and so on, which I think one needs to
48:08know a bit about. Uh, you can't be
48:12one-dimensional and wear blinders, only
48:14knowing how to write code. To me, this
48:19is a mistake. The only one I save, who
48:22doesn't say these foolish things, is
48:23Demis Hassabis, whom we mentioned
48:25earlier regarding AlphaFold. Demis has
48:27an enormous background in neuroscience,
48:30for example, right? And well, perhaps
48:34another who doesn't say it and who is
48:36very reasonable is Yann LeCun, the head
48:38of artificial intelligence at Meta. He
48:40says that these systems, these language
48:43models, are not leading us toward
48:45general intelligence. He said in
48:50English that it is an" off-ramp, "an
48:54off-ramp, a highway exit ramp. Uh, he
48:59says, he was saying on the road, on the
49:02path, on the freeway toward general
49:04intelligence, generative AI is an
49:07off-ramp. In other words, it doesn't
49:09bring us closer, it diverts us. Yann
49:12LeCun. And Hassabis are the only ones
49:17among all of them who say that the body
49:20is important for developing an
49:23intelligence that can be called smarter
49:26than what we have now, right? Both have
49:32powerful projects in their respective
49:34companies or labs on what they now call
49:37physical AI, which I and many others
49:40have called embodied artificial
49:42intelligence all our lives—with a
49:45body, with anthropomorphic robots—
49:47which is better than if they weren't,
49:50since they resemble the human body more
49:53. So you can replicate experiments as
49:57if they were humans, right? Interacting
50:00with their environment, with arms, with
50:02stereoscopic vision, and so on, right?
50:06So, well, Hassabis has a project called
50:10PLATO, in honor of Plato—Plato in
50:13English—where he places an embodied
50:17intelligence, a robot, in a situation
50:20similar to that of a young child. It is
50:26based on Jean Piaget's theories on
50:28development, cognitive development in
50:32humans, and in children in particular,
50:33right? So it begins to interact with
50:36objects and begins to learn
50:37cause-effect relationships. For example
50:40, learning the cause-effect
50:41relationship by interacting randomly,
50:44discovering that if it pulls on an
50:46object tied to a string, if it pulls
50:48the string, the object moves, and if it
50:50pushes the string, it doesn't, right?
50:53For example, it's not a typical
50:55cause-effect relationship that this
50:57system learns, right? Interacting with
51:00physical objects, with its physical
51:02environment, and some also have
51:05embodied AI projects as well. I
51:08published a paper back in 2015 about
51:10this, about cause-effect relationships,
51:13a robot that learned to relate where it
51:15placed its finger on a virtual keyboard
51:18and the note that played. After
51:21experimenting more or less randomly,
51:23the algorithm allowed it to associate
51:26the cause—finger—with the effect—
51:28musical note, which was do, re, mi, fa.
51:32And to verify that it had effectively
51:34learned this cause-effect relationship,
51:36we would say," Now play this melody,
51:38your turn. "The robot plays it
51:43physically, you know, ta, ta, ta, do,
51:45re, mi, fa. You would see an impeccable
51:51robot, 100 times in a row playing these
51:53four notes, well, well, hitting them
51:55every time, right? Sometimes, as is
51:58everything, there is a certain lack of
52:01precision in many of these humanoid
52:04robots. They aren't industrial robots
52:07that require nanometers of precision,
52:10so there are always small variations in
52:13the dynamics of the arm, and maybe the
52:16finger lands, hits the note next to it,
52:19but the robot realizes and corrects
52:21itself. That is another interesting
52:24thing we did, right? I mean, even
52:27before these gentlemen, Kun and Hazabi,
52:30got into making embodied artificial
52:32intelligence to learn cause-effect
52:34relationships, we were doing it in 2015
52:37.
52:40So, Ramón, are we exaggerating
52:45the capabilities of artificial
52:46intelligence when we talk about
52:47consciousness and talk about
52:48Of course. Yes, yes, yes. I'm telling
52:51you, those who exaggerate are those big
52:53companies I mentioned earlier, for
52:54economic interests, because of what
52:56investors say, right? Others, well, I'm
52:59not so sure, because there are other
53:02people, other folks who also believe
53:05that consciousness will be possible in
53:08machines with current technology, right
53:12? With current digital technology. I
53:15don't know. Look, I...some cases
53:18surprise me: colleagues who defend that
53:21this is feasible. And these are people
53:25I know, because I know them, who have a
53:28good background as computer scientists,
53:31you know, as software engineers and
53:34such, and who have and know many things
53:36even outside of software, and for them
53:39to say this surprises me, it surprises
53:42me. I don't know, I don't know why they
53:44say it. In any case, let one thing be
53:46clear: the burden of proof lies with
53:48the one making the claim. I mean, the
53:52burden of proof for when someone tells
53:56us," Hey, my machine, my system is
54:01conscious, "lies with them, not with us
54:05to prove that it isn't. Let’s not get
54:09this wrong. The person making the claim
54:11is the one who has to prove it.
54:13That’s how science works. I mean,
54:15it’s not enough for OpenAI or some
54:18company or Sam Altman to tell me,"
54:21Artificial General Intelligence is
54:23finally here. "Bravo, bravo, bravo,
54:26let's toast. No, that doesn't mean
54:28anything to me. It's useless to me. To
54:32start with, my suspicion is that
54:34they’re probably just pulling our leg
54:36, right? But until it’s
54:40scientifically demonstrated,
54:42reproducible, and they use the
54:44scientific method to convince the
54:48community, it’s meaningless.
54:52Unfortunately, it might be enough for
54:55careless journalists, those who follow
54:58the rule of never letting the truth
55:01ruin a good headline, right? to run
55:04with it and show up in newspapers all
55:06over the world. OpenAI has achieved
55:09General Intelligence. This will happen.
55:12We’re going to see it. I’ll bet
55:14whatever you want that we’ll see
55:16headlines like that in a relatively
55:18short time. And that worries me a lot,
55:23you know? For my sake, because I
55:25won’t believe it at all, but there
55:28will be many people who do. Many people
55:32will believe it. And later, when they
55:35interact with that AI, they’ll be
55:38convinced that behind the interface
55:40there is something, or someone, who
55:42cares about them, understands them, has
55:45emotions—and we’re already seeing
55:47it now. There are people using ChatGPT
55:51as a psychotherapist. For God’s sake,
55:54stop using ChatGPT as a psychotherapist
55:57. Do you realize the insanity of what
55:59you’re doing? But of course, the
56:02problem, why do I say this? Because to
56:05me, the serious problem is that
56:08you’re exposing yourself; you said,"
56:11No, the privacy issue goes much further
56:14. "You are revealing very, very
56:17personal things. You don’t talk to a
56:22psychotherapist about yesterday’s
56:24game or whether it’s sunny or going
56:26to rain. Very intimate things; they
56:30don't realize where they think that
56:32information goes, they think it stays
56:34on their computer screen and doesn't go
56:36anywhere. Please, uh, please, let's see
56:40if one day someone with a lot of bad
56:43faith sells this person's profile to
56:47some company and you want to get
56:50insurance and they deny it to you. And,
56:54oh, why did they deny my insurance?
56:56Well, maybe because they've concluded
56:58that you're mentally unbalanced, that
57:01you're schizophrenic,
57:02you've given them all
57:03because you've given them all this
57:06information and it's proprietary, it's
57:08proprietary software, please. Absolute
57:12nonsense is being done with these
57:15language models. People don't get it,
57:17they don't realize it, they have no
57:19critical thinking and this worries me a
57:21lot. The loss of critical thinking and
57:23common sense in society is accelerating
57:25. I'm going to tell another example.
57:28This, all of this is documented. Uh,
57:29I'm not making anything up. Uh, be
57:31careful. Uh, it has come out in major
57:33media outlets like the Washington Post,
57:35New York Times, which I subscribe to
57:37and and I see I see things that alarm
57:39me almost every day. A mother, a 5-
57:41year-old girl, her daughter, with spots
57:44on her skin, everything, spots all over
57:46her arms, her chest, everything. J.
57:49What's going on? What's going on?
57:50What's going on? Instead of taking her,
57:54taking her to a human doctor, she goes,
57:56takes photos, right? And she uploads it
57:59to a system, a generative AI, these
58:01multimodal ones, you can upload photos,
58:03I don't remember which one it was. It
58:06doesn't have to be ChatGPT, eh, don't
58:07hold me to that now. No, I don't want
58:09to say something false. A language
58:11model, I don't remember which one it
58:13was right now at this moment and it
58:15tells her, it diagnoses her very
58:17convincingly, very persuasively and it
58:19doesn't just diagnose, it says," Give
58:21her this, go to the pharmacy and do
58:23this treatment. "Without any critical
58:25thinking at all and without any common
58:28sense at all. This this mother buys the
58:30medicine and gives it to her and rushes
58:32to the ICU. An exaggerated allergy, I
58:37mean, the girl was enormously allergic
58:40to what she gave her. Well, she barely
58:45survived, eh, by the way, the news I
58:48read is that she hadn't hadn't passed
58:51away, but what a scare. Cases of
58:56suicides of people who have been
58:58interacting with chatbots and have been
59:01persuaded and have even felt something
59:03like like falling in love like Her, the
59:05movie Her is no longer science fiction
59:08and there are people now who are
59:09already uh experiencing what happens to
59:12the guy in the movie, right? With the
59:15protagonist, right? they end up
59:17practically falling in love with the
59:20artificial intelligence system. And
59:24there was also a case in Belgium where
59:27the wife of the man who committed
59:29suicide is now in legal proceedings in
59:32court against this company, right? In
59:36the United States, there was also
59:38another suicide case involving a 14-
59:40year-old girl. The Belgian guy was,
59:43like, so enormously in love with this
59:49character, because it was, ah, it was
59:52Character.ai, which has...it’s a
59:54conversational robot where you can even
59:57choose which character you want it to
1:00:00play. So, what’s her name? The
1:00:04protagonist from Game of Thrones.
1:00:06Daenerys, something like that.
1:00:07Targaryen. Dragons.
1:00:09Yes. The one with the dragons.
1:00:12And he was convinced he was talking to
1:00:15Daenerys. Convinced, uh, brutally
1:00:19convinced, and saying," I’d like to
1:00:21spend the rest of my life with you, "
1:00:24but, of course," where are you? "No,
1:00:27and it says," No, well, it's... ", and
1:00:30he asks," And if I die, will I be able
1:00:33to be with you? "the." Spend my life
1:00:36with you, "and it answered yes,
1:00:39basically encouraging him, and the guy
1:00:41goes and shoots himself. And the wife,
1:00:46who is a lawyer, by the way, she is now
1:00:49...well, I don’t know, I don’t know
1:00:51how it has evolved because I haven't
1:00:54seen any more news. The only thing I
1:00:57saw was that she had filed a complaint
1:00:59and that, in the first instance, it had
1:01:01been admitted for processing against
1:01:04this software company. Yes. I mean,
1:01:07look at where we’re getting to with
1:01:09the consequences, because it seems
1:01:11it’s not harmless at all; as we are
1:01:14seeing, believing you are conversing or
1:01:17interacting with something intelligent
1:01:19and conscious is quite the opposite.
1:01:23Perhaps, from a journalistic
1:01:24perspective, regarding that future you
1:01:26mentioned, one day we’ll wake up and
1:01:28the headline will be:" OpenAI now has
1:01:30consciousness. "
1:01:31Will we see this headline?
1:01:32Probably not. Even if it’s false,
1:01:34there won’t be a retraction either.
1:01:38Well, there will never be a retraction
1:01:40because, for example, even if I go on a
1:01:43media outlet or a podcast and say no,
1:01:45that as long as they don’t prove it
1:01:48scientifically, it won't go anywhere.
1:01:51Surely, in my opinion, I don’t think
1:01:53it would interest the headlines of
1:01:553,000 newspapers around the world, for
1:01:58instance, right? And the perception
1:02:01that it’s real reminds me a bit of
1:02:02the Wizard of Oz, right?
1:02:04Yes, yes, yes.
1:02:06That is behind a curtain and everyone
1:02:07believes it.
1:02:08It is manipulated.
1:02:11Or the chess player by von Kempelen, a
1:02:14machine from the 18th century. Von
1:02:19Kempelen was an inventor, an engineer
1:02:22who was very skilled and built a drawer
1:02:25like this, a box about this size, and
1:02:28the torso, they called it The Turk,
1:02:31they called it the mechanical Turk,
1:02:33because it wore a turban and was
1:02:36dressed like a Turk, right? And it
1:02:40moved pieces. Besides, it had a
1:02:44mechanical arm where it moved the
1:02:46pieces on a chessboard; I mean, on top
1:02:49of this table was the chessboard, okay?
1:02:53Well, he toured it all over Europe. Ah,
1:02:58royal houses, kings, queens from all
1:03:01over Europe, impressed that it played
1:03:04chess so well, because it played chess
1:03:07very well, very well. It beat almost
1:03:11everyone. Uh, even Napoleon is
1:03:16mentioned because one version stopped
1:03:18and then someone, after Kempelen died,
1:03:21someone else restored it or I don't
1:03:24know what happened, and it kept touring
1:03:27Europe, right? And well, it didn't take
1:03:35long to find out that inside there was
1:03:38a small-statured man. Now, I know the
1:03:43word dwarf cannot be used, but well,
1:03:45I've already said it. It seems it's
1:03:47politically incorrect. Who fit in there
1:03:50comfortably, and was a great chess
1:03:53player. So, he would see, because there
1:03:55was a transparent part there, he saw
1:03:57the positions of the pieces, right? He
1:04:00saw the positions of the pieces and it
1:04:03was the person inside who controlled
1:04:06the Turk's arm, this figure of this
1:04:08kind of dummy, to respond to the moves.
1:04:13Well, what is impressive, and this
1:04:15connects with what I said before, about
1:04:17the sensation of consciousness, and now
1:04:19I'm going to say something else, I
1:04:21think it's interesting, is that, even
1:04:24though when the truth was discovered
1:04:26and even though people already knew it
1:04:28was a trick, it continued to impress. I
1:04:35mean, since they couldn't see the one
1:04:37inside, it was like nothing happened. I
1:04:40mean, it kept touring and impressing
1:04:43people. I'm not saying everyone who
1:04:46went to see it knew there was a trick,
1:04:48but it was already quite vox populi
1:04:50that there was a trick. Okay? So there
1:04:53we have a problem. This is very well
1:04:55studied psychologically. There is the
1:04:59Müller-Lyer illusion, I think it is.
1:05:04That many of you will have seen. They
1:05:06are two segments of exactly the same
1:05:09length. But exactly, I mean, on the
1:05:12computer screen you make a line and
1:05:14then you could duplicate it and move it
1:05:16down. They are equal; it’s the same
1:05:18line that you’ve duplicated. Then, on
1:05:21one of the lines you put two closed
1:05:23arrows. That is to say, this is the
1:05:25line, imagine that this is the line. I
1:05:27put one arrow here, like this, and here
1:05:29an arrow like this, and on the other
1:05:31one the opposite, with the arrows open.
1:05:34Doesn't this illusion ring a bell? Well
1:05:36, where the arrows are open, with the
1:05:39two little arrows at the ends pointing
1:05:42outward, it looks longer. You see it,
1:05:45and you see it as longer. No matter how
1:05:48much you know—that's the interesting
1:05:51part, psychologically speaking—no
1:05:53matter how much you know that the two
1:05:55segments are exactly equal, equal down
1:05:57to the sub-millimeter, despite knowing
1:05:59it, and it happens to me and it happens
1:06:01to us all, we continue to see one as
1:06:03longer than the other. How is it
1:06:07possible that after knowing it, your
1:06:09perception still deceives you? I mean,
1:06:13this is called—it has a technical
1:06:15name—it's what's called something
1:06:18cognitively impenetrable. It means that
1:06:23neither reason nor knowledge of
1:06:24something necessarily modifies your
1:06:26perception of that something. And that
1:06:30will happen to us, it is happening, and
1:06:33it will happen with artificial
1:06:35intelligence, because no matter how
1:06:37much you know there is no consciousness
1:06:39, and no matter that through the
1:06:41interface you interact with a language
1:06:44model, no matter how much you know for
1:06:46a fact that it is not conscious, you
1:06:51will perceive consciousness. That's
1:06:54just how it is. That's how it is. So,
1:06:58well, it’s of little use in that case
1:07:01, because no matter how much you tell
1:07:02people" it is not conscious, "they will
1:07:04continue to see consciousness. I mean,
1:07:06in a way, it’s a lost battle. It’s
1:07:09a lost battle, that is, telling people,
1:07:11" No, no, no, this has no consciousness
1:07:13at all, "because even if they say," Ah,
1:07:16okay, okay, okay. "But like the arrows,
1:07:18right? Like the segments, they will
1:07:20continue to see, they will continue to
1:07:22think, perceiving, perceiving
1:07:23consciousness. You said in your book,
1:07:26and I think it’s related to what
1:07:28you're talking about, what you’re
1:07:29commenting on now, that the problem is
1:07:31not Frankenstein, the problem is Dr.
1:07:33Frankenstein.
1:07:34Of course, of course. Yes. This is
1:07:36relevant to the part of the book
1:07:38chapter where I talk about ethics and
1:07:40social and ethical explanations of
1:07:42artificial intelligence. Right, and
1:07:44this connects with the consciousness
1:07:46issue. Just imagine the consequences of
1:07:49the fact that something could be
1:07:52considered conscious at a societal
1:07:55level. Society includes judges, let's
1:07:59not forget that. Imagine that something
1:08:02, an artificial one, as I was saying
1:08:04before, the headline, OpenAI, ChatGPT,
1:08:07the latest version 5 point something,
1:08:09is already conscious. And that people
1:08:14believe it, everyone believes it.
1:08:20Imagine the examples I was giving
1:08:22earlier about suicide, or about the
1:08:24mother with the spots on her daughter's
1:08:27skin. Imagine that something similar
1:08:31happens, which will surely happen, and
1:08:33you file a lawsuit. Like the widow in
1:08:38Belgium who said," Yes, the lawsuit has
1:08:39been admitted for processing. "Why?
1:08:43Because there is a judge with common
1:08:46sense who said," Yes, sir. Let's see,
1:08:50agreed that this person possibly
1:08:52already had a certain tendency to
1:08:55commit suicide, they were already
1:08:57vulnerable, but the chatbot played a
1:08:59role here. It's the straw that broke
1:09:06the camel's back, right? So, well, that
1:09:11case is admitted for processing, but
1:09:14imagine that in a while there is a
1:09:17social consensus or the claim that it
1:09:20is conscious, and someone sues because
1:09:23someone committed suicide, the son or
1:09:27someone, or it gave me this treatment
1:09:30and I died because I was allergic.
1:09:35Terrible things, with really serious
1:09:37consequences, and you go and sue,
1:09:42you're wasting your time because
1:09:44something that is conscious has duties
1:09:46and responsibilities. If you are
1:09:48conscious, you have intentionality, you
1:09:50have everything. I mean, saying that
1:09:52something is conscious carries a series
1:09:55of things behind it that we don't
1:09:56realize. It’s just terrible. So, well
1:10:00, a judge could calmly say, "Ah, look,
1:10:02if artificial intelligence is conscious
1:10:05, it is a moral agent, and
1:10:06automatically, if it is a moral agent,
1:10:09it is responsible for its actions.
1:10:13Therefore, I cannot go after the
1:10:15company, or the programmer, or Mr. Sam
1:10:18Altman, or whoever it was that released
1:10:21, that deployed this, this, this
1:10:24chatbot. It’s brilliant for these
1:10:28people because they shake off
1:10:30responsibilities. I have sometimes come
1:10:33to think that these claims, that people
1:10:36say general AI will be here in a year
1:10:38or two and that it will be conscious
1:10:40and such, do not only obey what I said
1:10:43before about return on investment,
1:10:45about getting investors who put in a
1:10:47lot of money. Maybe, I'm perhaps
1:10:50speculating a bit here, eh? But well,
1:10:52it’s my opinion. It could be that
1:10:55they already have this in mind too.
1:10:58Meaning: well, we're very interested in
1:11:00having people believe, having everyone
1:11:02believe that what we're going to
1:11:04release is truly conscious, this moral
1:11:06agent, it bears all its own
1:11:08responsibility; because if something
1:11:10serious happens, they won't be able to
1:11:12go after us, I mean, they will have
1:11:14shaken off the responsibility. This, be
1:11:17careful with this consequence too,
1:11:19right?
1:11:21I mentioned before, we are in the realm
1:11:23of ethics, and before
1:11:25no, sorry, of course, that was the
1:11:26Frankenstein part.
1:11:27Exactly,
1:11:28in this case, if it's conscious, they'd
1:11:31say," No, no, the problem is
1:11:33Frankenstein, not Dr. Victor
1:11:34Frankenstein who created it. "It's the
1:11:37same as saying the problem is the
1:11:39artificial people, not Mr. Sam Altman,
1:11:41not Mr. So-and-so who developed it.
1:11:43That's where it comes from, okay? Sorry
1:11:45. Let's wrap up the question.
1:11:50I was saying that, well, Hinton, who I
1:11:51mentioned earlier, the Nobel Prize in
1:11:53Physics, right?,from last year.
1:11:55Last year, yes.
1:11:57He warned about the relationship
1:11:59between artificial intelligence and
1:12:01fascism.
1:12:03Why can it generate the right context
1:12:05for that to emerge by widening the gap
1:12:07between the rich and the poor? Do you
1:12:10think there are reasons to
1:12:11Yes, yes, are yes this? I think that's
1:12:14where I actually agree with him. I
1:12:18don't agree with him on many things,
1:12:20but on that, yes. And others have also
1:12:23said it, Timnit Gebru, Emily Bender,
1:12:27people who, well, really think and
1:12:30reflect a lot on these aspects. Ah, yes
1:12:35, yes. Have they mentioned this
1:12:39possibility, right? Yes, yes. Well, we
1:12:43saw it with Cambridge Analytica, right?
1:12:45And in how social media guides you
1:12:47toward
1:12:48Exactly. How they can be manipulated,
1:12:49right? Look, there's a very interesting
1:12:51thing. This was said by Daniel Dennett,
1:12:54a philosopher of science who passed
1:12:55away less than a year ago,
1:12:57unfortunately, who was a man who knew a
1:12:59great deal, he had an incredible
1:13:00scientific and theological background,
1:13:02aside from originally being a
1:13:04philosopher. Daniel Dennett said in an
1:13:08interview, let's see, isn't it true
1:13:11that counterfeiting money is punishable
1:13:14under the penal code and you can go to
1:13:17jail for it? Well, counterfeiting
1:13:20people should be even more serious. And
1:13:24that's what we clearly see now with
1:13:26generative AI, the ease with which you
1:13:29can falsify not just still images, but
1:13:31videos. Videos where you could see me,
1:13:37for example, committing a crime
1:13:42red-handed, right? Doing something,
1:13:46Look, this is brutal, all of this, the
1:13:49consequences again of all of this,
1:13:51right? And I attended, no, I wasn't in
1:13:55the debate, I attended a debate a
1:13:57couple of years ago, maybe a year or
1:14:00less. A debate that took place about
1:14:03artificial intelligence and issues of
1:14:05justice and legal matters. And in the
1:14:09debate, one of the panel members was a
1:14:12judge, an ex-judge, and he said:" The
1:14:14day is coming when they will show us a
1:14:17video in a trial of the defendant
1:14:19committing the crime red-handed, and if
1:14:22we don't have 100%certainty that it's
1:14:25not fake, that it hasn't been generated
1:14:27by AI, if we don't have that certainty,
1:14:30we won't be able to admit the video as
1:14:33evidence. "Wow. Wow. Hey, watch what
1:14:38we're saying, that's heavy, eh.
1:14:42So, it would have to be regulated in
1:14:44some way.
1:14:45Of course, of course. Regulate it, make
1:14:47it mandatory to disclose. Watermarks,
1:14:51some way, something, something that
1:14:53appears there that says this video...
1:14:54But of course, those who make these
1:14:56videos aren't interested. What's the
1:14:59point of making a video trying to go
1:15:02against a political opponent or whoever
1:15:04to crush them? What nonsense. If this
1:15:08video isn't fake, nobody will do it.
1:15:11This, yes, you can regulate. I think we
1:15:15have to regulate this, by the way,
1:15:17because if you get caught, you pay, at
1:15:20least, right? But chemical and
1:15:22biological weapons are also regulated,
1:15:24and they are still being used. Now,
1:15:27since it's regulated and forbidden, if
1:15:30you get caught, you pay. This is
1:15:33important.
1:15:34What things do you think, Ramón, that
1:15:35we should never delegate to artificial
1:15:37intelligence?
1:15:39Well, let's see, nothing that, for
1:15:42starters, nothing that poses a danger
1:15:45to our cognitive abilities. For example
1:15:48, imagine that—and now I'm going to
1:15:50go to something else that isn't
1:15:52artificial intelligence, but it serves
1:15:54as a way to transition to talking about
1:15:56artificial intelligence. Imagine that a
1:16:004 or 5-year-old child—I don't know
1:16:02when kids start school now to start
1:16:05adding, subtracting, these most basic
1:16:08arithmetic operations. Imagine that
1:16:11instead of teaching them how to add,
1:16:13how to subtract, the concept of
1:16:14addition, right? With examples, and
1:16:17having them assimilate the concept of
1:16:19what it means to add, to subtract, and
1:16:22how it's done with pencil and paper.
1:16:25Imagine that instead of that, you give
1:16:27them a calculator. That child will know
1:16:30, will learn to use the calculator by
1:16:33dialing the number, pressing the plus
1:16:35or minus sign, whatever, and adding
1:16:38another number. Hit equal and there you
1:16:40go. That's great, isn't it? But we are
1:16:44doing terrible damage to this child
1:16:46because we are depriving him of
1:16:48acquiring a, what a, a skill, a
1:16:50competence. It's super essential. It is
1:16:53essential. This, this same child once
1:16:57he has learned it, right? Obviously, if
1:17:01it's the calculator, and he has already
1:17:04well learned, well internalized, well
1:17:06assimilated, eh, adding, subtracting,
1:17:08dividing, more things, whatever. He
1:17:12goes to high school and has to do a
1:17:14math problem, I don't know, eh, where
1:17:16there are some derivatives, a
1:17:18calculation of maximums and minimums,
1:17:20something a bit more, of a certain
1:17:22slightly more advanced math, right? And
1:17:26one of the steps of the problem
1:17:27solution he has to solve involves an
1:17:30arithmetic calculation. Well, there you
1:17:33can allow the calculator, there it's no
1:17:36problem because he already has this
1:17:38essential skill learned from when he
1:17:41was, from when he was 5 or 6 years old.
1:17:44Now it's fine that he uses the
1:17:45calculator, it no longer matters. Okay,
1:17:48now ChatGPT, do we no longer have to
1:17:56teach students to write an essay or to
1:17:59make a, I don't know, to write, in
1:18:02short, right? Or maybe we shouldn't
1:18:06even teach spelling, right? Because,
1:18:09well, since the machine corrects it for
1:18:11you anyway and now it even writes it
1:18:13for you, there we have a problem
1:18:16because at a certain age, again, this
1:18:18is also essential that they learn it. I
1:18:21mean, in school or university you don't
1:18:24, you don't make students write an
1:18:27essay just so there's another essay in
1:18:30the world, in addition to the 50
1:18:33billion that already exist. That is not
1:18:36the goal of writing an essay. The goal
1:18:38of a student writing an essay is for
1:18:40them to learn to express themselves
1:18:43well, to know how to think, to reflect.
1:18:46That is the goal, not the essay itself.
1:18:49Amen. So, if we are now allowing, if
1:18:52you allow students to use these tools
1:18:54instead of them learning to write, we
1:18:56are heading in the wrong direction. We
1:18:59are heading the wrong way. Furthermore,
1:19:01there is an article recently published
1:19:04last year, 2024, March '24 in Nature
1:19:07Human Behavior, a very prestigious
1:19:10journal from Nature, right? It is very
1:19:13well known in the world of science, it
1:19:15is perhaps the most famous of all, the
1:19:17most prestigious. And human means human
1:19:19behavior, as we know, right? Where they
1:19:22have a very interesting article that
1:19:25says writing is thinking. Writing is
1:19:29thinking. Writing is not vomiting one
1:19:32word after another like ChatGPT does.
1:19:36It's not chaining words together, no,
1:19:39it's thinking, revising, seeing, facing
1:19:41the blank page. Let's see, I want to
1:19:45express this, I would like to write
1:19:47about this topic, right? Let's see how
1:19:50I do it. Reflection, you pace around, I
1:19:53mean, you are setting in motion a
1:19:54series of brutal cognitive brain
1:19:56mechanisms, much more than you think.
1:20:00And then basically, you more or less
1:20:02manage to, right, do your writing,
1:20:04right? Your essay, whatever it is,
1:20:07right? And besides having the
1:20:10satisfaction of having successfully
1:20:12faced the blank page, which for me is
1:20:14important to begin with, you've also
1:20:17done a brutal mental exercise of
1:20:19thinking. That's why writing is
1:20:21thinking. If we delegate writing to
1:20:24artificial intelligence, we don't just
1:20:26delegate writing, we delegate thinking.
1:20:29We can lose skills, we can lose the
1:20:31cognitive ability to think. What is
1:20:34called deskilling, there's even a word
1:20:36for this now, deskilling, right? The
1:20:38loss of skills. And this could be
1:20:41terrible. So, be careful when casually
1:20:44using these language models or these
1:20:46AIs, delegating your question was about
1:20:49what I wouldn't delegate. Well, I
1:20:52wouldn't delegate these things that
1:20:55pose a risk to competencies, skills,
1:20:58essential things that any human should
1:21:02know. Obviously, I also wouldn't
1:21:04delegate many other things if we talk
1:21:06about privacy, I wouldn't use it as a
1:21:08therapist.
1:21:09Or autonomous weapons, all of that, as
1:21:11I mentioned before, right? Obviously, I
1:21:13could make a long list of things I
1:21:16wouldn't delegate, but to synthesize a
1:21:18bit, anything that could jeopardize my
1:21:20cognitive abilities, I have a
1:21:22preventive priority, which I've almost
1:21:24lost already, not entirely. Still, when
1:21:28I don't have access to Google Maps, I
1:21:30can still get places. Google Maps. How
1:21:34many people, many of us, surely those
1:21:36of us here and those listening, if they
1:21:38reflect, will realize that years ago,
1:21:40when there was no Google Maps, they
1:21:42reached their destinations or looked at
1:21:44a map, a paper map, a typical map.
1:21:47Sometimes you had to turn it because
1:21:49otherwise you couldn't interpret it
1:21:51well. Those fun things people used to
1:21:53do with maps, right?
1:21:54Not knowing how to fold it,
1:21:55not knowing how to fold it correctly.
1:21:57What a mess. Well, it doesn't matter
1:21:59what breaks. Well, it's all the same.
1:22:01Okay, but we could orient ourselves, we
1:22:04were able to map what was on the paper
1:22:08to what we saw, meaning important parts
1:22:11of the brain were also at work. Now
1:22:16with ChatGPT—sorry, we're talking
1:22:18about Google Maps—aren't we getting a
1:22:21bit dull? Many people no longer know
1:22:23how to orient themselves. I've realized
1:22:27myself that I've lost some of my sense
1:22:31of direction from using Google Maps a
1:22:34bit more than I probably should. Sure,
1:22:39look, it's practical, you get there
1:22:41sooner, if you're in a hurry, it works
1:22:43out great, doesn't it? Using Google
1:22:46Maps or the car's navigation, right? Eh
1:22:48, but well, it's another example that
1:22:52we can lose skills. If you had the
1:22:55capacity to reach millions of people,
1:22:57Ramón, what would you tell them?
1:23:00Well, for them to be very cautious, to
1:23:02use artificial intelligence, yes, but
1:23:05with a lot of common sense and critical
1:23:07thinking. I mean, I'm no Luddite, mind
1:23:10you. It's not about throwing artificial
1:23:15intelligence into the trash bin. No way
1:23:19, not at all, right? Far from it. But
1:23:22now that it has reached our homes,
1:23:24unlike two or three years ago, right?
1:23:29Now that we have it there at a click,
1:23:32be careful, think it through, reflect
1:23:35well on what you're going to use it for
1:23:38, and if it's really, truly essential
1:23:41or not. I'll tell you something very
1:23:45funny. Someone who had to send an email
1:23:47to their boss—this is real, too, eh?
1:23:50And well, they wanted it written. But,
1:23:54of course, they put a list of items, or
1:23:58bullets, right? Point one, bam, point
1:24:02two, bam, a sentence, a sentence, five
1:24:04or six sentences saying," Based on
1:24:07these, now generate the email. "Wow, it
1:24:10was something super well done, super
1:24:12well written, flowery language,
1:24:14everything perfect, two or three pages
1:24:16—it went a bit overboard, maybe it
1:24:18rambles on a lot, eh? These systems
1:24:21ramble a lot, okay? They send it, super
1:24:24happy, they send this to their boss.
1:24:27The boss receives it. What is the first
1:24:29thing the boss does? Copy and paste. He
1:24:33copies it, pastes it into chat, and
1:24:34says," Rewrite this for me. "I mean,
1:24:40it's just, it's brutal, isn't it? There
1:24:45was no need to go through the chat, he
1:24:48could have just sent the email with the
1:24:50five or six points, short sentences,
1:24:51boom, and sent it directly to the boss.
1:24:55Of course, he wanted to show off, look
1:24:57how nice the email is. And what the
1:24:59boss did is," Wow, so long. Come on,
1:25:03summarize it, summarize it in a series
1:25:05of bullets, right? Of of items. I mean,
1:25:08damn, that's a great anecdote, I think.
1:25:10"
1:25:11And I'll change the question, although
1:25:14it's similar, if you had the ability to
1:25:16reach the CEOs of big tech companies
1:25:20and give them a message. Ah,
1:25:23but it's a message they would actually
1:25:24listen to.
1:25:25No, no, none of them. No, no, no.
1:25:28Whatever I say, they won't listen to me
1:25:29, because the message I'd tell them
1:25:31would be very critical. Obviously I
1:25:33would tell them:" Be more honest, tell
1:25:37the truth, don't go around fooling
1:25:40people saying that in a year AI will be
1:25:44conscious, uh, I mean, warn them, be
1:25:47careful, for certain things you don't
1:25:50do it. "It's not a doctor, it's not a
1:25:53whatever, it's not a lawyer, uh, it
1:25:56generates many errors. Ah, if they
1:25:58start if they start saying all this,
1:26:00they are going to run out of venture
1:26:02capital investors. Maybe it's just,
1:26:06it's obvious, right? They can't, they
1:26:09can't. They have to sell a fantasy and
1:26:12it won't matter much what I tell them,
1:26:15uh, not at all. I perhaps finish with
1:26:20this, of all the research we've been
1:26:22doing on artificial intelligence, what
1:26:24has it taught us about humans?
1:26:29Oof! Well, I well, I I can speak for
1:26:33myself, uh, I in the years when I
1:26:36started in 74 75, the 80s, also in the
1:26:3990s when I developed new machine
1:26:42learning algorithms. Ah, I was very
1:26:47enthusiastic and I saw that I don't
1:26:50know that that that that well, I was
1:26:53super interested in the subject, uh, I
1:26:57only saw advantages in some way, right?
1:27:01And but now I realize that it has
1:27:08gotten out of hand in some way, right?
1:27:11I mean, look, the academic world, the
1:27:14academic world in a way, until
1:27:16relatively recently, was setting the
1:27:18pace for where we’re going with
1:27:21artificial intelligence, right? But now
1:27:25, these big companies are setting the
1:27:28pace because they have the necessary
1:27:31ingredients for generative AI. Lots of
1:27:35data, lots of computing power, and
1:27:38enough financial resources to hire the
1:27:41most brilliant talent and, well, um...I
1:27:46see where we’re heading with AI with
1:27:50concern, but I think your question had
1:27:53another component, right?
1:27:56Perhaps what it has taught us, right?
1:27:59If, if that, if the AI...
1:28:00It has taught me, I’ve been somewhat
1:28:02disappointed by what we’re seeing now
1:28:04, right? Now, there is also something
1:28:06positive. It has also taught me that
1:28:09back in the days when I was so
1:28:11enthusiastic, I wasn't aware of the
1:28:14enormous complexity, the enormous
1:28:17complexity of the human brain, of human
1:28:20intelligence, right? Of what it means
1:28:24to be human and all that, right? I
1:28:27think I have a very different view now,
1:28:30and I’m valuing what it means to be
1:28:33human much more than before, while at
1:28:36the same time, I'm downgrading
1:28:39artificial intelligence. Meaning,
1:28:42before I perhaps overestimated
1:28:45artificial intelligence and perhaps
1:28:48underestimated human complexity. Now
1:28:51it's the other way around. Now, I
1:28:53overestimate, because I believe it’s
1:28:56true, the complexity of the human brain
1:29:00and everything. Ah,
1:29:04and you underestimate artificial
1:29:06intelligence. At this moment, I would
1:29:10say I do that. That’s why I’m
1:29:12critical.
1:29:13Okay, Ramón, thank you very much for
1:29:15your time and for your work.
1:29:17Thank you.

About this transcript

This page contains the full transcript of Experto en IA: La Gran Mentira de la Inteligencia Artificial | Dr. Ramón López de Mántaras by Vidas Ajenas, generated from the public captions YouTube serves with the video. The transcript has 12,564 words across 1,997 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.