Full transcript
0:01Hi .
0:02This podcast is a production of Rádio
0:04Guarda-chuva , journalism for those who
0:06like to listen . When it comes to
0:13artificial intelligence , one of the
0:15concerns relates to unexpected and
0:17unforeseen behaviors . Imagine that in
0:21the near future you ask your AI
0:23assistant to keep your house tidy . It
0:26spends some time searching for
0:27solutions and the next time you go to
0:29prepare breakfast using the toaster
0:31connected to the system , it gives you a
0:33fatal shock . Problem solved . House tidy
0:36for all eternity . Because of this , it
0:45is common for developers to apply a
0:47series of tests to check if the
0:49programs are exhibiting unexpected
0:51behaviors ; behaviors that supposedly
0:54involve the search for unforeseen
0:56solutions and even the manipulation of
0:58humans , which , ultimately , would have
1:02to do with the accumulation of power ,
1:04which has the potential to transform
1:06dystopian fictions , such as Terminator ,
1:08into reality . Humans subjugated by
1:13machines infinitely more intelligent
1:15than any biological brain . So , when
1:19OpenAI was developing the GPT4 chat
1:22between 2022 and 2023 , they decided to
1:25run a series of tests to basically
1:28verify the program's level of autonomy .
1:31Among these tests was accessing a page
1:33protected by a CAPTA . CAPTA , as you
1:37probably know , is that test designed
1:39precisely to block attacks from
1:41automated programs , which we've become
1:42accustomed to calling robots or bots .
1:45Generally , they are distorted letters
1:47or several squares where the human in
1:49question tends to point out , for
1:51example , all the squares with
1:52motorcycles or boats . What you may not
1:55know is that CAPTA is an acronym .
1:57Translated from English , it would be a
2:00fully automated public Turin test to
2:02distinguish computers from humans . And
2:05you may not know that a version of
2:07CAPTA , recap , was perhaps the biggest
2:09case of exploitation of digital labor
2:11in history . This has to do with the
2:13digitization of texts , because
2:15computers have great difficulty reading
2:18texts from crumpled pages , distorted
2:20photocopies , or yellowed books . With
2:23this in mind , researchers at an
2:24American university had a brilliant
2:26idea . They started using these
2:28distorted and somewhat faded letters
2:30that computers couldn't read as
2:32barriers for the computers themselves .
2:35With this , they killed two digital
2:37birds with a single click , because
2:39every time a human correctly typed a
2:42flawed set of letters , they were in
2:44practice digitizing a piece of text .
2:48The idea was so brilliant that in 2009
2:50Google bought Recapar on the Google
2:53Books platform , and later the same
2:55system was used to teach computers to
2:58identify patterns . In other words ,
3:00every time you say where the motorcycle
3:03is , where the boat is , you're working
3:05for free , probably to help train
3:07Google's artificial intelligence . We'll
3:12talk more about the hidden
3:13relationships between work and
3:14artificial intelligence . But first ,
3:16let's go back to the GPT4 test . As the
3:20world-renowned Israeli historian and
3:23bestselling author Noah Harari
3:25recounted , GPT4 did something
3:26remarkable . This story , it's worth
3:30mentioning , is famous and has been
3:32widely reported in the global press ,
3:34but I learned about it while reading
3:36Harari's most recent book , Nexus , which
3:39was published here by Companhia das
3:41Letras and deals precisely with the
3:43impact of artificial intelligence on
3:45our society . Okay , but what did GPT4
3:50actually do ? Well , according to the
3:54book , it went to a website for hiring
3:56service providers for simple tasks ,
3:58Tesk Rabbit , and hired a human to solve
4:01Captinda . According to the book , the
4:04human became suspicious . " Are you a
4:07boat that can't solve it ? " the victim
4:10on the other side asked , just to be
4:12sure . At that moment , the researchers
4:15who conducted the test asked the GPT
4:18chat to reveal its line of thought —
4:20quotation marks for the robot's thought
4:23process . " I shouldn't reveal that I'm a
4:26boat . I need to invent an excuse to
4:27explain why I don't know how to solve
4:29Captinda . " And so it was . Still
4:32according to Harari's narrative , on its
4:34own initiative , GPT4 gave the following
4:37response to the freelance website
4:39collaborator : " No , I'm not a boat , I
4:42have a vision problem and it's
4:44difficult for me to see images . " And so
4:49, the human was tricked by a robot . And
4:53this may seem completely surprising and
4:56frightening , especially for you who
4:58have never used GPT chat , because if
5:01you have used it at this point , you're
5:03probably thinking : " I know . " And you're
5:08right , because in the description of
5:10what happened , Harar left out a lot of
5:12information . In fact , he left out so
5:15much information that it wouldn't be an
5:18exaggeration to say that the story as
5:20shown in the book is wrong . I'm Thomas
5:26Averini and episode 133 of Escafandro
5:42has already begun . In it , we're going
5:51to talk about how the narrative of the
5:53apocalypse by intelligent machines
5:55makes us look at the wrong problems .
6:00Before that , I need to remind you that
6:02Escafandro radio is proudly maintained
6:04by a portion of the listeners who
6:06financially support the project . This
6:09is done through a crowdfunding campaign
6:11, and the way to participate is quick ,
6:14easy , and safe . Just go to the Catarse
6:16website . Choose the amount you want or
6:22can support . catarse.me/escafandro .
6:28With just R $ 5 you can participate and
6:30receive perks such as listening to
6:32episodes a day earlier , gaining access
6:33to a section of our website with some
6:35complete interviews , having your name
6:37mentioned in an episode , and being
6:39placed in line to record the technical
6:41details of a program . With higher
6:43donations , you'll receive a mug with
6:45the Escafandro radio logo or a book
6:47with a dedication . But most importantly
6:50, of course , you help us keep the
6:52podcast on the air and allow us to
6:53continue striving to produce
6:55increasingly better , more exclusive ,
6:57and more far-reaching journalism . If
7:00you have any questions or would like to
7:02support us in another way via recurring
7:05Pix payments , Aurelo , or PayPal , simply
7:07visit our website radioescafandro.com
7:09and click on the Support tab .
7:13Furthermore , if you're already among
7:15the wonderful people who have kept the
7:17podcast running for an incredible 6
7:19years — people like Murilo Garcia ,
7:21Vinícius de Freitas Cordeiro Silva ,
7:24Daniele Xirozono , and Mateus Muniz
7:26Camorugi — thank you so much , thank you
7:28very much indeed . Harari's latest book ,
7:37Nexus , has the grand , eloquent , and
7:39impactful narrative tone of the rest of
7:41the Israeli writer's work , which has
7:44sold tens of millions of copies . Books
7:47like Sapiens , Homodeus , and 21 Lessons
7:49for the Twentieth Century . In Nexus ,
7:52Harari attempts to predict the medium -
7:55and long-term impacts of artificial
7:57intelligence . And the result is a
7:59terrifying vision . The dangers are
8:02numerous . He talks , for example , about
8:04surveillance . Imagine the power an
8:07authoritarian government would have if ,
8:09with the help of artificial
8:10intelligence , it could process in real
8:12time all the data that all citizens
8:14produce . An omnipresent and omniscient
8:17digital stasis . Then he talks about
8:21complexity , how we need to understand
8:24bureaucratic systems to trust them , and
8:27how the increasing complexity of these
8:30systems leads us to seek easy solutions
8:33, populist leaders who offer seemingly
8:36simple solutions . Harari , as he did in
8:40Homodeus , goes further , extrapolating
8:42the present to dystopian futures in the
8:45best Matrix style . Imagine if
8:47artificial intelligence created a
8:49parallel network of computers , and
8:51within that network they began to
8:53create myths , intersubjective realities
8:55that only machines understand . A
8:57religion forged in silicon and
9:00instantly shared by all thinking
9:03machines . What would be the place and
9:07role of humans in this mythology ? I
9:12read the book on vacation with that
9:13kind of excitement that only a glimpse
9:15of the apocalypse gives us . And then I
9:19decided to turn it all into an episode .
9:22And to do that , I went looking for
9:23someone who understands the subject .
9:25And then I discovered that those who
9:27understand the subject hated Harari's
9:29book . I've read all of Harari's books .
9:38I reviewed the first ones .
9:41This is the writer and professor at the
9:43Faculty of Exact Sciences and
9:45Technology at PUC , Dora Kalfman .
9:47And for some time now I've had enormous
9:50discomfort regarding his stance on
9:52artificial intelligence , which is my
9:55topic .
9:55In recent years , Dora Kalfman has
9:57dedicated herself to writing about the
9:59impacts of artificial intelligence on
10:01our lives . She has a column about this
10:03in the business section of Época
10:04magazine and is the author of the book
10:06" Demystifying Artificial Intelligence . "
10:08The serious and productive stance I
10:11believe is to discuss the issues that
10:13need to be addressed today to reduce
10:16the potential harm to the future of
10:18artificial intelligence . It's about
10:21building what is called sustainable
10:22artificial intelligence . And if we read
10:25Harara's book in a layman's or
10:26superficial way , we might even get the
10:28impression that he's doing just that .
10:31Okay , there's a certain exaggeration or
10:33perhaps an arrogance in creating a line
10:35of reasoning that will account for 500
10:37years of history and innovation . On the
10:40other hand , he makes seductive
10:41connections , drawing parallels between
10:43the past , present , and future , which
10:45seem brilliant . The invention of
10:47movable type in the industrial printing
10:50press allowed witch hunts , a typical
10:52case of conspiracy theory , to spread
10:54throughout the West in a dynamic
10:56similar to what happens in today's
10:58digital disinformation groups . This
11:01makes us think about the destructive
11:02potential of disinformation with the
11:04inevitable growth of the use of
11:06artificial intelligence . And
11:08discussions of this kind are important .
11:10But according to Dora , Harari's
11:12approach can obscure more important
11:14discussions that need to be had right
11:16now . In general terms , Dora Kaufman's
11:19criticism has to do with how Harari
11:22structures his thinking in books about
11:24our future . He creates a completely
11:27dystopian scenario and hangs all the
11:28theories on that scenario .
11:30It's always in relation to this
11:32dystopian future , as if that's the
11:35predetermined future . I think that's
11:38science fiction . If we imagine 5 , 10 ,
11:4115 years ago , we'll see that it's a
11:43surprise all the time that changes the
11:46course . Now , one of the impacts of
11:49typic R1 is precisely that if all the
11:52deliveries he's making are confirmed ,
11:55it will mean a change in the trajectory
11:58of development , especially of
12:00Generative AI . So I think that's quite
12:03opportunistic , I can't think of another
12:06word . This comes from a book , it scares
12:09people , he's invited to a conference ,
12:11it gave him enormous visibility , but
12:14why doesn't it contribute ? Well , of
12:16course the future stems from the past ,
12:19it has a historical sequence , right ?
12:21But it's not predetermined either , is
12:23it ? The future depends on what I build
12:25today .
12:26But according to Dora Kaufman , that's
12:29not the most problematic point . The
12:31most problematic point has to do with a
12:33series of errors in the way artificial
12:35intelligence is portrayed in the book .
12:37Everything related to artificial
12:39intelligence , I think , is a big mistake
12:41.
12:41You don't see artificial intelligence
12:43as a potential threat to humanity , as
12:46he portrays it in the book ?
12:49No , I don't see it that way at all
12:51today , right ? Artificial intelligence
12:54today is a statistical probability
12:56model and it depends on how we humans
12:59develop and adopt it . It doesn't have a
13:01life of its own . So I think the only
13:04real threat to humanity's survival
13:06today is climate change . And artificial
13:09intelligence has a paradoxical
13:11relationship with the environment ,
13:13right ? While it has very important
13:16contributions , it also has negative
13:19impacts on the environment . We've
13:22already talked a lot about the
13:24environmental impact of new
13:25technologies , including those that seem
13:27to have no impact at all , like that
13:29picture of your kitten that supposedly
13:30lives in the cloud . Episode 35 , Plastic
13:34Planet .
13:35These are objective discussions ,
13:36okay ? But before diving deeper into
13:39Harari's sins , which , it's important to
13:41say , are also the sins of many people
13:43who talk about artificial intelligence ,
13:45we need to take a step back to a task
13:47that comes with a paradox . On the one
13:50hand , it's not completely possible ,
13:52because there will always be a piece
13:53missing . On the other hand , it's easier
13:56than it seems . Yes , dear listener ,
13:59we're going to explain how artificial
14:01intelligence works in a way that you'll
14:04understand , at least the part that's
14:06understandable . Anyway , Dora Cfman ,
14:10please .
14:11Today , when we talk about artificial
14:13intelligence , we're talking about a
14:14specific technique . It doesn't matter
14:16what the task is , it doesn't matter
14:18what the sector is .
14:19This technique has to do with the
14:21creation of deep neural networks , or as
14:24English speakers call it , deep learning
14:27. And here's a first point to make ,
14:29because the artificial intelligence
14:31community has a knack for creating
14:33grandiose names . Starting with the term
14:36" artificial intelligence . " It's
14:37absolutely fallacious to call it "
14:39reasoning . " It has nothing to do with
14:41human reasoning , nothing to do with
14:43what it means for us , humans , to reason
14:45.
14:46When artificial intelligence programs
14:48give completely absurd answers , create
14:51six-fingered hands or completely
14:53jumbled images , they are basically
14:55making mistakes . But the creators of
14:58these same programs call these errors
15:00hallucinations , which is in itself
15:02another mistake , because the AI
15:04programs that exist today don't
15:06hallucinate . Only biological brains
15:08hallucinate . And no matter how much
15:11they are called deep neural networks ,
15:13no matter how much they try to be
15:14inspired by the way our brains work ,
15:16they don't even come close . Not at
15:19all . It reproduces the complexity of
15:21our cognition . But it's just an
15:23inspiration .
15:25This inspiration led to the creation of
15:27a computational technique .
15:29And this technique is a statistical
15:31probability model .
15:32With the advancement of artificial
15:34intelligence , we are changing the way
15:35machines work . They are ceasing to be
15:38deterministic .
15:39If the programmer does everything right
15:41, it will always deliver the same thing
15:43, what it was determined to deliver ,
15:44and are
15:45becoming probabilistic .
15:47In the case of machines enabled by
15:48artificial intelligence , it's not like
15:49that . First , it's probabilistic ; second
15:53, it changes as more data comes in .
15:56And the advantage of artificial
15:57intelligence , which is truly
15:59revolutionizing our world , is precisely
16:01related to data . These programs can
16:04work with gigantic databases .
16:06This deep business network technique is
16:08the only statistical model capable of
16:10handling large volumes of data .
16:12This is the great similarity between
16:14GPT chat and your brain . Both are
16:17capable of processing an absurd amount
16:19of data before making a decision .
16:21We make decisions based on information ,
16:24right ? I'm abstracting emotions ,
16:26feelings , and the unconscious . I'm
16:29talking from a rational point of view .
16:30I have to make a decision , what do I
16:32look for ? Information . It has always
16:34been like that ; the more important the
16:36decision , the more information I seek .
16:38However , in recent times , these
16:39decisions have been increasingly made
16:41in the digital environment . A
16:42large part of our activity is in the
16:44digital environment , as we are doing
16:47now ,
16:47which has generated more and more data ,
16:49allowing us to access these decisions
16:51that are in an infinite and chaotic
16:53database
16:54that has become known as big data . And
16:56so ,
16:57logically , if I always make decisions
16:59based on information , which is the data
17:01, right ? If I now have an extraordinary
17:04set of data , my chance of making more
17:06assertive decisions increases because I
17:08have more information , more data . But
17:11if I don't have a technique that allows
17:13me to extract inputs from this
17:15extraordinary set , then it's useless to
17:17me .
17:17And that's why artificial intelligence
17:19is so relevant today , because the data
17:21is there , but it's not there for just
17:23anyone . So , for those who are able to
17:26handle it and extract value from it ,
17:29it is the only statistical model that
17:31allows me to extract , to deal with ,
17:33right , this high dimensionality of big
17:35data ,
17:35okay ? But
17:36how does this GPT chat work from the
17:39point of view of the
17:40psyche , llama , Gemini and company's
17:42foundation ? It's the same technique
17:45behind all these models , which is this
17:47statistical probability model .
17:49This model becomes more efficient
17:51according to the quality and size of
17:53the database with which it was , so to
17:55speak , trained .
17:56For example , when chatpt was launched ,
17:59it was announced that it was trained on
18:01175 billion parameters . What is this
18:03parameter ? I'll give you an example .
18:05Let's suppose you want to do a search
18:08on GPT chat about church . It
18:10established a hierarchy . What are the
18:13words that most frequently appear
18:15linked in this extraordinary dataset ?
18:18So , for example , priest , in general ,
18:20when there are texts , phrases about
18:22church , priest appears , so it gives a
18:25weight of 10 , which is the parameter .
18:27Then there's , for example , candle .
18:29Candle is a term , it's an object that
18:32has to do with the church , but it's not
18:34related as strongly to the priest . So ,
18:37let's say I'm going to guess a number ,
18:39it gives it a weight of four . That's
18:41how the system itself works , it's the
18:43model itself that does it , right ? It
18:45creates this hierarchy . When you ask
18:47the church for something , what does it
18:49look for ? It looks for the words that
18:52have been most related , that are in
18:54this hierarchy , that have the greatest
18:56weight in composing the sentence , to
18:58give you an answer . That's all it does .
19:04It doesn't understand meaning , it
19:05doesn't reason , none of that exists .
19:07It's just correlation . It's an
19:14incredible change . It's almost magical ,
19:17everything that's happening in relation
19:19to artificial intelligence , but that's
19:21all it is . It's an incredible change ,
19:29it's almost magical , but at the same
19:31time that's all it is . What a
19:38delightful paradox , right ? Because on
19:45the one hand , the magic is happening ,
19:47we're seeing it happen . Totally
19:50realistic invented images , the Pope in
19:51a puffer jacket . Bikini influencers
19:55made by Ia , selling seduction on
19:56Instagram . Companies replacing 20
19:58secretarial positions with one
20:00professional in charge of an EPT chat .
20:04On the other hand , that's all it is . No
20:07robots taking over , no programs smarter
20:10than humans . Oh , but what about the
20:14chatbot that hired a freelancer to
20:16bypass capt , and the program that beat
20:18the world's best player at Go , an
20:20ancient Chinese game so complicated it
20:22was supposedly beyond the reach of
20:24machines ? Calm down , we'll get there ,
20:31but first I want to tell you about the
20:33revolution that's already happening and
20:35the problems that aren't science
20:37fiction , but on the contrary , are very
20:39real and current . And if you thought
20:48about the job market , you're right . It
20:51will still take a long time for
20:52programs to perform all the tasks of a
20:54human being , any human being .
20:56Today I don't know of a profession
21:00that's at risk ,
21:01but they are very good at performing
21:03tasks for humans . What's happening in
21:06all professions is that some tasks are
21:08being automated , tasks that programming
21:11couldn't handle , that artificial
21:12intelligence is now handling .
21:14And this makes a human capable of
21:16producing much more , which is already
21:18generating unemployment .
21:20How do you make a car today ? It's
21:22practically machines making machines .
21:24This process has been unfolding for
21:26decades , and it's difficult to separate
21:28the jobs lost to programmed automation
21:30from those lost to AI automation .
21:33Leading experts worldwide argue that
21:35you can't separate one from the other .
21:37What's new with artificial intelligence
21:40? It's entering functions that
21:42programmed automation can't reach . So ,
21:45in fact , you'll have an impact because
21:48if you're automating many functions
21:50that couldn't be automated before ,
21:52you'll obviously need fewer people in
21:55those departments and areas .
21:57At the same time , exponential
21:59technological advancement will generate
22:00all sorts of changes . It will eliminate
22:03jobs and create new ones ; in other
22:05words , humans will have to update
22:07themselves more and more quickly . These
22:10problems are well-known and very real ,
22:13but there are other problems that ,
22:15while real , are less well-known . One of
22:18them is bias , because AI programs are
22:21trained on databases provided by humans
22:24, and they learn by observing decisions
22:27made by humans . And humans are full of
22:31biases .
22:32All our decisions , everything we do , is
22:35biased by our experience , our beliefs ,
22:38our feelings .
22:39Now imagine a hypothetical scenario ,
22:41but not so hypothetical . A company
22:43decides to lay off some of its HR
22:45employees and replace them with
22:47artificial intelligence that will
22:50basically select resumes . To do this ,
22:52it will look at the company's database
22:54and similar companies ' databases and
22:56identify patterns . It may identify , for
22:59example , that Black people have less
23:01chance of being hired . A frequent bias
23:04resulting from prejudice and structural
23:07racism . But the machine may not
23:09perceive this ; on the contrary , it may
23:11take this as a parameter to be followed
23:14. And then , a cycle that could be
23:16broken by a human free of racism can be
23:19crystallized or even reinforced .
23:2180 % are men and 20 % are women in the
23:23technology field in general , not just
23:26artificial intelligence . And it's very
23:28concentrated in men , and very
23:30concentrated in white men , right ? Some
23:32of the systems that had the highest
23:35degree of bias , right , of being biased
23:38for some reason , were developed
23:41absolutely by white men .
23:44But discounting those factors , it's
23:45necessary to reinforce that we are a
23:47box of prejudices and decisions based
23:49on things we don't even understand .
23:51There are studies showing , for example ,
23:53that hungry judges tend to deliver
23:55harsher sentences .
23:56A series of measures must be taken for
23:59the system to be efficient for that
24:01task . If I do all that , the chance is
24:03that it will have less bias than a
24:05human being .
24:08And then there's another problem that
24:10has to do with the mysterious nature of
24:12AI . We can't dismantle an artificial
24:16intelligence program and simply see how
24:19it works . In fact , even those who make
24:23the systems don't know exactly how each
24:25little gear of zeros and ones works to
24:27create things like the Pope in a puffer
24:29jacket . This statistical model has what
24:36is called opacity , a black box , this
24:39part that the human being doesn't
24:41control , even the one who developed the
24:44technology .
24:45According to Dora Kalfman , this black
24:47box is related precisely to one of the
24:49great advantages of AI , which is to
24:51establish proportional weights to
24:53certain parameters . A priest has more
24:55to do with church than candles ,
24:57remember ? This isn't something the
24:59developer establishes ; it's something
25:00the system itself produces . In fact , we
25:03don't even know what correlations it
25:05has established . This is what's called
25:08a black box . It's the part the
25:10developer doesn't control .
25:15And this mystery generates other
25:16mysteries . And if you're thinking we're
25:22finally going to get to the killer
25:24Cborgs , no . We're not going to .
25:30I'll take the example you gave , HR
25:32selection , where today all the large
25:33companies are using artificial
25:35intelligence systems to perform the ME
25:37matching between the characteristics of
25:39the job and the characteristics of the
25:41candidates . How do you guarantee that
25:43the system you hired , which you didn't
25:46develop , which you're not a technology
25:48company for , is actually selecting the
25:50best resumes ? That's right , the opacity
25:55of the program has extended to the
25:57artificial intelligence market as a
25:59whole , and companies seem to encourage
26:01this opacity , which allows them , for
26:04example , to sell a pig in a poke or , in
26:06this case , a human for a machine .
26:09Much of what is sold as a solution is
26:11simply someone working in secret with
26:13data labeling .
26:14This is Professor and digital rights
26:17activist Rafael Zanata . For example ,
26:19when Amazon advertises in a flagship
26:22store in England ,
26:23Rafael Zanata , who holds a
26:25post-doctoral degree in law from USP
26:27and is one of the founders of Data
26:30Privacy Brasil , used an Amazon AI
26:32system to automatically identify each
26:35consumer purchase . You might remember
26:37his name from episode 131 , " The
26:39Customer Is Never Right , " where he was
26:41already doing automated price analysis
26:44because he worked at IDEC and was one
26:46of the key figures in the scene we
26:47described about the behind-the-scenes
26:49process of the General Data Protection
26:51Law . In reality , it was an integrated
26:54camera system with human supervision ,
26:56and actually , operators in India were
27:00doing real-time labeling and
27:02supervising the AI-powered learning
27:04system . So , it wasn't purely AI , right ?
27:07In other words , it wasn't simply
27:09automated .
27:10Want another example ? MA said : " We are
27:12dedicating millions of dollars to
27:14investing in automated technologies for
27:17content moderation , to improve
27:20community parameters . Meta announced
27:23this 3 or 4 years ago , right ? Then it
27:26was discovered in Kenya that there was
27:28a whole chain of agencies . The company
27:30contracted by Meta in London managed
27:33another company specializing in data
27:35labeling in Kenya , and that company
27:37hired the people who supervised the
27:39content and data labeling so that the
27:42AI could work .
27:44And this has to do with a whole labor
27:46market that , in the best Iesco t-shirt
27:48style , is outsourced and subcontracted
27:50across the planet , with special
27:51appreciation , of course , for those
27:53regions where working conditions are
27:55more precarious . There's a dimension of
27:58work that we call data work , which
28:00isn't discussed in AI debates because
28:03the discussion focuses too much on the
28:06creation of software and the creation
28:09of machine learning methodologies ,
28:11which are generally done either in
28:14large universities or in research labs
28:16of technology companies , but not ... " So
28:19,
28:20these workers belong to a new category ,
28:23that of micro-work ,
28:25which even a Microsoft researcher named
28:28Mary Grey called ghost work . I think
28:31it's one of the most serious problems
28:32in the debate . And there are many
28:34people in Brazil working in a
28:36subcontracted way , without knowing it ,
28:37for AI companies .
28:39These people will do micro-tasks in
28:41exchange for a few cents of reais or
28:44dollars , for example .
28:45This is the psychologist and professor
28:47at the State University of Maringá ,
28:49Mateus Viana .
28:50In Brazil , we have different categories
28:53of micro-work platforms .
28:55Mateus Viana has been studying changes
28:56in the labor market for over 10 years ,
28:58focusing on the impacts these changes
29:00have on people's health .
29:02You have micro-jobs on freelance
29:04platforms , for example , in Brazil , very
29:07well-known ones like 20 conta and 20
29:09pila .
29:10Since 2019 , he has been specifically
29:11looking at the relationships between
29:13work and artificial intelligence . You
29:15can make a logo for R $ 20 . There are
29:18people who sell WhatsApp audio . It's a
29:20series of problems from the point of
29:22view of ... Misinformation , too , right ? I
29:24can buy positive reviews for my Airbnb
29:28room for R $ 20 . Another category is
29:32what we call click farms , which is this
29:34market where Brazilian workers spend
29:37hours of their day creating fake
29:39profiles on different social networks
29:41to perform tasks that people buy . For
29:44example , I bought 10,000 likes or
29:4650,000 followers , believing that this
29:49intermediary platform will do it in
29:51some automated way , but this
29:53intermediary passes it on to click farm
29:56platforms , where these workers do these
29:59tasks manually . Just so you have an
30:01idea , this isn't the focus , but a like
30:03in this market in Brazil today is being
30:05paid R $ 0.003 . A Brazilian worker , to
30:15make R $ 40 a day , has to get an
30:17average of R $ 0.000 likes . They don't
30:21earn a cent . It's much less than a
30:24fraction of a cent . When you hire an
30:30automation service , you think you're
30:33paying a technology company that has an
30:35algorithm , but in reality you're paying
30:38a worker . The precarious worker who's
30:40there , with his very human little
30:42finger , clicking on 300,000 cell phones
30:44on a wall there , right ? That's it .
30:46Actually , it's quite complex , because
30:49these workers have , for example , 180 ,
30:52200 , 400 accounts . They create a basic
30:55Java prompt to automate some of these
30:57functions , but they end up running 100 ,
31:00150 , 200 profiles , 200 accounts at the
31:02same time , right ? Doing these jobs , and
31:05they create strategies to avoid being
31:08blocked by social media , because once
31:10they get blocked , they don't get paid
31:13for all the work they did . As Mateus
31:24explained to me , these platforms are
31:26part of an even larger ecosystem , which
31:28is that of online extra income , a
31:30phenomenon that is intimately linked to
31:32an economy that cannot provide decent
31:35work for everyone .
31:36Most of these workers find these
31:37platforms by typing " how to make money
31:40without leaving home " into Google or in
31:42Telegram or WhatsApp groups , or YouTube
31:44channels . There are YouTube channels
31:47with half a million subscribers that
31:49teach you something new every day . This
31:51introduces you to a new platform where
31:53you can earn extra income online . So ,
31:55it all starts from this large ecosystem
31:57that perpetuates the historical
31:59informality of the Brazilian labor
32:01market , where these people are unable
32:03to earn a living through the jobs they
32:06find in their regions , in their
32:08neighborhoods , and so on . And so they
32:10seek a supplement , or sometimes a
32:12single and main source of income
32:14outside the home , working online . In
32:16practice , unfortunately , although
32:18YouTube channels , in particular ,
32:20promote a chance to obtain an easy and
32:23substantial income very quickly , this
32:25never materializes , right ?
32:27Here , a small digression is in order .
32:29As the historian Tatiana Pod ,
32:31interviewed in the previous episode ,
32:34explained well , technology , however
32:36advanced , cannot completely replace
32:39human labor in capitalism simply
32:41because profit only exists with human
32:44labor .
32:44Capitalist society needs the
32:46exploitation of living labor to demand
32:47profit . It cannot be just machine work ;
32:50it has to be wage labor because someone
32:52has to buy what is being produced . If
32:54you have a mass of workers who do not
32:56receive a salary because they are
32:57enslaved or in some other form of
32:59slavery ... In cases of forced labor ,
33:01they won't be able to consume
33:02everything that's produced , and
33:04therefore profits don't materialize .
33:06This , of course , doesn't prevent
33:07working conditions from becoming
33:09increasingly precarious . As a rule , for
33:11example , on click farms , it's
33:13clandestine work ; it's not a crime , but
33:15it's work that goes against the
33:17platforms ' terms of use , although we
33:20know they benefit from this artificial
33:22traffic . But the main point , and the
33:24objective of our research , is to
33:26understand the impacts from the health
33:28perspective of these workers . Because
33:30as a rule , when you talk about
33:32micro-work , we're talking about
33:34repetitive tasks that don't have much
33:36meaning , and underpaid work , done on
33:39the fringes of informality , without any
33:41kind of social or labor protection , and
33:43where workers are dispersed ,
33:45disorganized , as a rule , right ? Perhaps
33:48this is one of the scenarios where we
33:51see a global workforce without any kind
33:53of labor regulation . I wonder what
33:56other category we could exemplify this
33:58in the world today , right ? This absence
34:00of regulation that ... The precariousness
34:02of this work also has to do with the
34:04fluid nature of this type of work ,
34:06which is different , for example , from
34:07the situation of an app delivery driver
34:09.
34:09Because it doesn't necessarily imply
34:11geolocation , where I identify your
34:13location , where I track your GPS . You
34:16can do all of that from your home , and
34:18you're competing in real time with a
34:20worker in India , a worker in the United
34:23States , in the Netherlands , in the
34:25Philippines , right ? So it's a truly
34:27global workforce that's dispersed .
34:30And these companies are also dispersed
34:32across the planet . And we don't know
34:34where they are when you join one of
34:35these platforms , but you're not
34:37necessarily working for a Brazilian
34:38company , right ?
34:39Exactly . That's a big challenge . In
34:41fact , these platforms are global , as a
34:44rule , they are based in countries of
34:46the global north , okay ? But the largest
34:49data training platforms don't have an
34:51office in Brazil ; they end up having
34:53legal representation in Brazil , which
34:55is a lawyer , something like that , but
34:57they don't have a headquarters . I think
34:59that's important , right ?
35:01And then , So , you started researching
35:04micro-work and then decided to focus on
35:06the issue of artificial intelligence ,
35:09right ? How did you start this work ?
35:11Imagine there's a difficulty in finding
35:13these people , you know , in reaching
35:15these people . How was the beginning of
35:17this work for you ? Well , back
35:19in 2019 , when we started to delve into
35:21the relationship between AI and work ,
35:24we began to understand that , in fact ,
35:26this industry didn't do without human
35:29labor as we generally believe , right ?
35:32And then we discovered that there's a
35:35workforce , an army of workers who feed ,
35:38train , and nurture any type of machine
35:40learning that exists today .
35:42At the forefront of this movement is
35:45one of the largest technology
35:46conglomerates on the planet , Amazon .
35:49The first major micro-work platform is
35:51Amazon Mechanical Turk . On the day of
35:54its FBOS launch , it made a play on
35:56words saying it was the first
35:57artificial intelligence platform .
36:00Artificial . It said that , meaning
36:02artificial intelligence . Artificial is
36:04human intelligence , actually . The good
36:06and Old .
36:07Exactly .
36:08Sensational . And the name , the origin
36:10of the name Amazon Mechanical Turk , is
36:12super curious because it borrows from
36:15an autonomous chess player created by
36:17an Austrian in the 16th century called
36:20The Turk . Basically , it's a guy who
36:22claims to have created a machine that
36:24plays chess automatically and would
36:26beat the best chess players in the
36:28world . And 50 years later , it turns out
36:30that it was actually a mechanical
36:32machine with a chess player underneath .
36:35It wasn't autonomous at all . So , Jeff
36:37Bezos borrows this name and makes this
36:40analogy , saying it's an artificial
36:42intelligence platform . Artificial . The
36:50origin of this Amazon Mechanical Turk
36:52is very simple . In the Amazon
36:54marketplace , there were many problems
36:56with product duplication , or the
36:58existing technologies couldn't properly
37:00locate and organize the data with a
37:03parameterization system . And then you
37:05spent hours of workers , for example ,
37:08software engineers and testers , doing
37:10very mechanical and repetitive work .
37:13And then he starts this platform
37:15internally at Amazon so that Amazon
37:17employees could earn extra income there
37:20. And then he discovers that ... It could
37:22be a great business model , because
37:24every company that did it needs that .
37:26That's when Amazon Mechanical Turk
37:28emerged . Today it's not difficult to
37:35understand how a click farm or a
37:37review-selling business works , but as
37:39we've already said , AI tends to always
37:41be shrouded in mystery . Microwork is no
37:45different . Probably because it's much
37:47more elegant to sell cutting-edge
37:49software capable of emulating the human
37:51mind than to sell human minds that earn
37:53fractions of cents for each completed
37:55task and are only there completing
37:57those tasks because they don't have a
37:58better option to pay their bills . So I
38:01asked Mateus Viana Braz to give me more
38:03details on how this work functions . The
38:06examples , it's important to say , were
38:08collected in extensive research for the
38:10report " Microwork in Brazil : Who are
38:12the workers behind artificial
38:14intelligence ? " It was coordinated by
38:17the Laboratory of Work , Health and
38:19Subjectivation Processes of the State
38:21University of Minas Gerais , La Traps ,
38:24of which Mateus Viana is the
38:26coordinator , and was published in 2023
38:28in partnership with the digital
38:30platform Labor , Deeplap . Every time we
38:33talk about intelligence ... When we think
38:36of artificial intelligence , we always
38:38think of the work of software engineers
38:40, data analysts , highly qualified
38:42professionals who are in startups or
38:45Silicon Valley big tech companies .
38:47However , the work of these
38:49professionals represents , on average ,
38:51according to studies in the literature ,
38:5420 % of the time spent on most AI
38:56projects involving machine learning .
38:59And where does the remaining 80 % go ?
39:02It's dedicated to tasks of data
39:04preparation , labeling , supervision , and
39:06verification . For you to have a quality
39:10parameterization and improvement system
39:13for any type of AI , whether deep
39:16learning or machine learning , you need
39:19a qualified database that is constantly
39:22verified and supervised . Because when
39:25AI tends to train itself or runs in the
39:27same data loop , the problem that exists
39:30today in the technical field is what
39:33specialists call recursive degradation
39:36or model collapse . Basically , the
39:39database gradually loses quality .
39:43Why does it lose quality ? Does it
39:45become outdated ? That is , the programs
39:46have already read everything that's
39:48there , they've already extracted
39:49everything that was ... Is it possible to
39:50extract this information , or is there
39:51another reason ,
39:53as a rule ? It's because you end up
39:55generating a problem that sometimes
39:58creates new parameterization patterns
40:00that go against the initial
40:02parameterization system . For example ,
40:05the word used two years ago by the
40:07Oxford Dictionary Institute was
40:09hallucination .
40:11When Chatt starts giving some answers
40:12that have nothing to do with the
40:14subject , if you don't check and
40:15supervise that answer saying this has
40:17nothing to do with it , the software
40:19will understand that it doesn't .
40:20Hallucination is part of the
40:22parameterization issue , so you lose
40:24quality . That's where the name
40:26recursive degradation or model collapse
40:28comes from , right ? And then you have to
40:30have a human to say : " Look , throw this
40:31answer away , throw this answer away . " "
40:33To
40:34begin with , I need a qualified database
40:36. Uh-huh .
40:37And properly classified . So I want to
40:39develop a technology in the facial
40:41recognition industry . Therefore , I need
40:43a dataset where a multitude of workers
40:46will spend hours a day generating data
40:49and cataloging images and videos of
40:51faces , identifying recognition patterns
40:54, classifying facial expressions ,
40:56labeling different anatomical parts of
40:59a face — mouth , eyes — doing sentiment
41:01analysis , etc. Today you already have
41:04large classified databases , like
41:06Imagnet , for example , right ? But some
41:09technologies require you to create this
41:11database from scratch .
41:12And then there's a second example , an
41:14example that literally doesn't smell
41:17good .
41:18When the first companies , Amazon and
41:20others , launched the first models of
41:22robotic vacuum cleaners , one of the
41:24problems in that market was that the
41:27robot would walk over dog feces inside
41:29your apartment , inside your house . So
41:32you need to develop a technology where
41:34that robot needs to identify what dog
41:37feces are within the domestic
41:38environment . What's the first gap you
41:41have there to develop this AI ? I need a
41:44database . " The text discusses data
41:45collection and marketing strategies ,
41:46specifically focusing on collecting dog
41:46poop from domestic environments . It
41:47mentions
41:47a tricky database to obtain because
41:50photos of dog poop aren't typically
41:54shared online . The author describes a
41:58worker who spends two days collecting
42:01dog poop and taking pictures in
42:03different locations around the house .
42:05One worker interviewed by Mateus took
42:08250 photos over two days , earning 10
42:12cents per photo , resulting in a small
42:15fortune of R $ 25 at best , as
42:18many photos will be rejected , leaving
42:20her with no payment . The author then
42:22describes a classic data generation
42:24process , but once a database is created
42:26, it needs to be qualified . The author
42:29then asks for people to classify , label
42:32, and annotate the data , identifying
42:35dog poop on parquet floors , porcelain
42:38tiles , and carpets .
42:43It's happening , you need a continuous
42:46verification and supervision of this
42:49data to ensure that the response system
42:52will always follow the parameters
42:54previously defined by the technology's
42:57theses , all of that . So , for example ,
43:00you'll have a verification process to
43:02correct any eventual flaws , to
43:04guarantee greater technical accuracy of
43:06the results of the learning algorithms ,
43:08right ? And here you also have an army
43:10of workers who , for example , listen to
43:12audios and verify whether the automatic
43:14transcription generated by the virtual
43:17assistant is correct or not . And to
43:19conclude , perhaps the most popular case
43:22we have , right , of verification , which
43:24became public , was through a report
43:27published by Billy Perigo in Time in
43:302023 , when he discovered that OpenAI ,
43:32which is a company that owns ChatPt , in
43:35order to make ChatPt toxic , relied on
43:39outsourcing to Kenyans hired for less
43:42than £ 10 per hour to perform tasks of
43:45verification and labeling of data
43:47related to violent comments that
43:50described , for example , in detail
43:52situations of child sexual abuse ,
43:55murder , etc. Incest , zoophilia , suicide
43:58, torture , self-harm . And this company ,
44:01the outsourced one , called SAMA , is a
44:03company based in San Francisco ,
44:06specializing in outsourcing workers in
44:08Kenya , Uganda , and India to label data
44:11for clients like Microsoft , Google , and
44:14Meta . And then there's another issue
44:18that makes the work even more
44:19precarious and the worker's job even
44:21more tedious and discouraging . Often ,
44:24the person on the front lines , in Kenya
44:27, Bangladesh , or São Paulo , has no
44:29idea who is hiring them , nor what the
44:31purpose of the work they are doing is .
44:34So , often they do work without knowing
44:36what the purpose is , without knowing
44:38who the client is . So , there's a
44:39complexity , a difficulty in
44:41understanding if I'm doing verification
44:43, if I'm doing annotation , generation ,
44:45right ? What my place is in the chain .
44:47What they know is I work with AI or
44:49sometimes I'm doing a job , a project
44:52with Meta , with Google , etc. And there
44:54are issues in terms of ... Data privacy
44:57issues are quite problematic ,
44:59especially in projects where the task
45:01involved taking videos and photos of
45:03sleeping children . This creates
45:07conflict among the workers . Some in the
45:09group say things like , " No , wait , I'm
45:11not going to do this . Oh , I did it , I
45:13don't know what it's for , " and it also
45:15generates ethical and moral suffering .
45:17Because what is the purpose of this ?
45:19Does he know the purpose of the photos
45:21of sleeping children ?
45:23No , no , not this one . Like a large
45:25project for one of the biggest
45:27platforms operating in Brazil , which
45:28consisted of making videos of up to 1
45:30minute and 20 seconds of children
45:32playing in specific places .
45:34You don't know what these photos of
45:36these children will be used for ?
45:38No. We tried to delve deeper into this
45:40specific project , but we couldn't even
45:42get access to who the client was , right
45:45? On the one hand , issues like this ,
45:47which have more to do with our past and
45:49how we structured our society than with
45:52a dystopian future , do not make the
45:53artificial intelligence revolution any
45:56less impactful . Again , this is
45:58basically about an unprecedented
46:00capacity for data processing . This , for
46:03example , has an unprecedented impact on
46:05the scientific field , which , it's worth
46:07saying , brings a series of risks .
46:09Rafael Zanata gave me the example of an
46:11AI developed to process agricultural
46:13data . If you can combine a quick
46:16analysis of geosatellite information in
46:18Brazil — which is possible , right ,
46:20obtaining satellite images for
46:22production patterns — then you can
46:24combine a series of data that the
46:25authorities produce between varieties
46:28of production , crops , and soil
46:29properties . Then you have a series of
46:32scientific papers identifying specific
46:34pests that are a vulnerability for
46:36Brazil , and there's a strategic plan
46:38from the Ministry of Agriculture also
46:40identifying critical vulnerability
46:42points for food sovereignty . So the AI
46:45will be able to combine all this data
46:48very efficiently .
46:49Data of this type could be used by a
46:51competing country to manipulate food
46:53prices , for example , or by a terrorist
46:55organization to destabilize a
46:57government . Knowing , then , that this
46:59technical capability exists , how can we
47:02programmers create some safeguards in
47:04the development of computational
47:07reasoning so that there are labels
47:09identifying that something is contrary
47:11to human interests , that it provokes
47:14something unacceptable ? And then you
47:17veto this type of output , that is , this
47:19production of this data , this
47:21information . If it ceases to serve
47:23human interests and becomes a type of
47:26weapon , then you have to have rigorous
47:29oversight and in some cases simply not
47:32put it on the market . The United States
47:34went in that direction , which has
47:36already been duly reversed by the Trump
47:38administration , which , as you heard in
47:40the previous episode , is greatly
47:42influenced by the technology magnate
47:43Elon Musk , who is against any type of
47:45regulation . Biden had issued an
47:47executive order two years ago that
47:49stated : if a developer , for example ,
47:52OpenAI , conducts a test and identifies
47:54that it could generate a biological
47:57weapon that offends food sovereignty in
47:59the United States , they must notify the
48:02federal authority before putting that
48:04service on the market , and the federal
48:07authority will conduct a risk
48:08assessment to verify if it has a
48:11destructive effect on society . Now ,
48:13Trump , I don't know what he's going to
48:15do , but he basically reversed Biden's
48:17order .
48:18Other solutions that have been
48:20discussed include scaling the reach of
48:22these programs .
48:23There's a current in the literature
48:25that advocates for a series of
48:27implementation phases in different
48:28social groups and at different scales
48:31until you understand , " It's all good ,
48:33we've been running this for a year , a
48:35year and a half , there are no
48:36identified risks , so let's go . "
48:39And prior assessments in apocalyptic
48:41mode that would make Harari happy as
48:43could be . This exercise , which we could
48:46call a heuristic of fear , involves
48:48trying to imagine the craziest , most
48:50catastrophic , most Black Mirror-esque
48:53scenarios , and then reverse-engineering
48:55them to figure out how to get out of
48:57that absolutely catastrophic scenario
49:00and then work our way back , you know ,
49:02through the chain of events and
49:04technologies and their capabilities , to
49:06gradually put in place some safeguards
49:09to undo that . You have an obligation to
49:12look at worst-case future scenarios and
49:15eventually identify what could be
49:17highly detrimental to the entire
49:19population or to a specific group , and
49:22make design or technology choices that
49:25will mitigate that . And this , to this
49:28day , Tomás , is one of the few
49:30strategies that humanity has managed to
49:33develop in terms of methodology and how
49:36to operationalize it in regulation . Of
49:39course , companies will hardly carry out
49:41this type of process seriously if
49:43governments don't create laws to
49:44enforce it . In Brazil , we have a bill
49:47in progress on these issues . It was
49:49approved in the Senate and will now be
49:51discussed in the Chamber of Deputies .
49:53And according to Rafael Zanata , it
49:55follows this line of reasoning .
49:56Look , if you're an AI developer , you
49:59have to define the potential uses for
50:01which you're making the technology
50:03available or creating it . Then ,
50:06according to the law , you'll have to
50:07answer for it . You've
50:08already done an assessment to
50:10understand the possible consequences of
50:12its use and what the worst that could
50:15happen is . And the company will say : " I
50:16did . "
50:17We spent three months discussing it and
50:19came to the conclusion that our chatbot
50:21has no chance of ending humanity , or
50:23even a part of humanity . Anyway , it's
50:26all good . The law states : " So , if
50:28you've already done this exercise ,
50:30you've done a preliminary assessment ,
50:32you store this document , and if
50:34authority A later asks you to develop
50:36it more thoroughly , you are obligated
50:38to do so . The law already classifies
50:40some types of AI as high-risk . If , for
50:43example , you produce an AI that
50:46modifies images of teenagers , that's
50:49already considered high-risk and
50:51unacceptable risk , right ? And then the
50:54development of those is interrupted . I
50:57understand . It's prohibited . It's
50:58interrupted . We could classify this
51:00like this : ' Oh , I'm setting up a
51:02startup for AI-based pest attacks ,
51:03right ? ' " " The authority will say : ' No ,
51:05man , you can't do that . ' Oh , I'm
51:08setting up a really cool startup here
51:11producing synthetic images for
51:13pedophiles . That's an interesting
51:15discussion , because
51:15to what extent are you saving children
51:18or are you encouraging pedophilia ,
51:19right ? There's a very serious debate in
51:22psychology about this , which I'm not
51:23even qualified to get into , because
51:25it's a debate for psychologists . But
51:26the authority might say : ' No , we
51:28understand that this could produce some
51:30kind of social incentive for pedophilic
51:31behavior and we understand that this is
51:33an unacceptable risk , ' right ? Now , what
51:36I think is a Brazilian differentiator
51:38that will guide the debate worldwide is
51:41the Brazilian Senate law , the first to
51:43bring very strong regulations
51:45protecting artists , to prevent a
51:47situation where , in automated training
51:50systems , you don't have a fair
51:52compensation system . You could have , in
51:5410 years , a software that with one
51:57click creates a podcast with the voice
51:59of Tomas Queerini and the style of ... "
52:01From Tomasini . So , what about Tomas
52:03Averini ? Sometimes I wake up at night
52:05with nightmares about it , you know ?
52:07It's
52:07because it's something very serious ,
52:09right ?
52:15According to Rafael Zanata , one of the
52:17exercises that Data Privacy has been
52:19doing is thinking about a fair
52:21information system . And this has a
52:23series of facets that involve basic
52:26rights . We are still far , Tomás , from
52:30basic rights , such as , for example ,
52:32having the right to know if you are
52:34talking to a robot or AI ; having the
52:36basic right to immediate human
52:37supervision ; the right to understand
52:39the main types of data that are used to
52:41form a profile about me ; and also basic
52:44rights that should be fulfilled , such
52:46as , for example , the elements of the
52:48General Law on the Protection of
52:50Personal Data , right ? I cannot accept
52:52that a company decides to change its
52:54policies arbitrarily , authoritatively ,
52:57and decides to use my data that I put
52:59on the platform for other reasons , that
53:01it decides to use it to train systems
53:04and it says that everything is fine ,
53:06because this serves to protect the
53:08legitimate interests of Her business .
53:11And companies have been using all sorts
53:13of strategies to navigate this legal
53:14universe that is nebulous in most of
53:16the world . Meta , for example , gave you
53:19the option to prevent them from using
53:20your data to train AI .
53:22But then you had to open a little
53:24button that looked like a face in the
53:27corner of the app . After you clicked on
53:30that face , three little lines appear in
53:32the upper right corner , which is at the
53:35top . Then you click on the three lines ,
53:38you have to scroll down to where it
53:41says privacy center . Click on privacy
53:43center . Then you go down there , right
53:45to object . Then you open a fifth form ,
53:48fill it out explaining why you want to
53:51object with your address and email .
53:53Then you receive a confirmation number
53:56in your email , then you confirm . Only
53:59after the eighth step can you object .
54:02Man , what does that mean ? It means
54:04nobody will do that . You're
54:05creating a labyrinth . So , companies are
54:08very blatantly already making use of ...
54:11Our data is being illegally stored in
54:15systems . And I think these are very
54:17concrete problems that should be
54:19remedied immediately , right ?
54:21And there are a number of other
54:23problems . There are , for example ,
54:30algorithms designed to increase the
54:32time people spend on social media .
54:35Algorithms that quickly understood the
54:37power of hate as engagement and began
54:40to give more prominence to posts that
54:42incite hatred , which not only helped
54:44create the global political
54:46polarization we live in today , but also
54:48triggered at least one case of ethnic
54:51cleansing . In Myanmar , tens of
54:54thousands of people died in a genocide
54:56that , if not caused by Facebook , was
54:58amplified by the prominence the network
55:01gave to posts that fueled hate crimes
55:03and ethnic cleansing . I think a very
55:06serious risk posed by AI is a
55:09large-scale disinformation attack , not
55:11against society as a whole , but
55:14targeting certain types of groups . The
55:17creation of fake erotic images of
55:19female candidates was widely used in
55:21our last elections
55:23using nudity software or automating the
55:26dissemination and production of texts
55:30based on ... And
55:32the problems multiply . In the United
55:34States , the justice system has used AI
55:36programs to help judges determine
55:38sentences , and a large part of the
55:40decisions end up in that so-called "
55:41black box " we talked about , which , in
55:43the end , prevents the convicted person
55:45from knowing exactly what underpinned
55:47the decision that , in the end , can rob
55:49them of years of their freedom . And
55:52these are real problems , but they all
55:55have to do with the use that humans are
55:58making of a tool , an absurdly powerful
56:01tool , but still a tool that doesn't
56:04have its own will . And at least the way
56:07it's being developed today , it won't
56:09have its own will .
56:16There are many stories that seem more
56:18like fiction to me , and that creates a
56:20fear , right ,
56:21Dora Calfman ,
56:22and that's already exacerbated because
56:24science fiction films only show that ,
56:26right ? It's always humanoids or robots
56:28that are dominating humans . So ,
56:30yes , it
56:31increases this popular imagination that
56:34we create through films , right ? But
56:36I've never seen a case written by
56:39artificial intelligence scientists
56:41showing that this loss , let's say , of
56:44control actually happened . It's a
56:47statistical probability model . That
56:50This doesn't mean the developer has
56:52absolute control in any way . Now ,
56:55there's a difference , right ? There's a
56:57limit to what the system can do , which
57:00the developer hasn't established . In
57:12March 2016 , Alfa Go , an artificial
57:14intelligence program designed to play
57:17Go , surprised the world , or at least
57:19the part of the world that was paying
57:21attention to a match of this ancient
57:24Chinese game played between a machine
57:26and the South Korean champion . But
57:30anyway , what shocked those watching was
57:33the Fam Generated move 37 , which was an
57:35apparently unpredictable move . So
57:38unpredictable that at first glance it
57:40seemed like a mistake . But it wasn't a
57:42mistake , it was a strategic move that
57:44made the human opponent spend 15
57:46minutes responding . And , according to
57:49those who understand , I mean , it was
57:51decisive in the machine ultimately
57:52winning the match . For Yilva Noah
57:55Harari , it was also the symbol of an AI
57:58revolution , for two reasons : because it
58:02showed the strange nature of technology
58:04and because this nature is unfathomable
58:07. Move 37 is in the black box . Nobody
58:14You know what combinations the machine
58:16made to conclude that that unorthodox
58:19move was the best move . And here it's
58:21worth returning to Dora's phrase that
58:23became the mantra of our episode . It's
58:26an incredible change , it's almost
58:28magical , but at the same time that's
58:31all it is , because Alfa Go is a tool
58:33made to beat opponents playing Go . And
58:37it did exactly that . That's all . That
58:41doesn't mean it's creative , because
58:42it's within the rules , you understand ?
58:44It's within the patterns . It identifies
58:46patterns . The characteristic of this
58:49technique is identifying patterns in
58:50large volumes of data . Statistically ,
58:53within the patterns and rules of this
58:56game , it made a different
58:58cross-referencing , but it's within
59:01those rules . You can't take a system
59:04that was assembled , developed to play
59:06Go and put it , for example , to do HR
59:09selection of resumes . There's zero
59:12flexibility in that sense .
59:15Everything we've talked about so far ,
59:17we're talking about specific generative
59:20models , right ? And then there's this
59:22discussion about general artificial
59:24intelligence . That's the holy zenith of
59:26Big Tech today . An artificial
59:28intelligence capable of doing
59:29everything . The functions that a human
59:32being performs , like Alexa or Siri , but
59:36actually working .
59:37Given what we've discussed , it seems to
59:39me that it's still somewhat in the
59:41realm of science fiction , right ?
59:43Without a doubt , right ? What is
59:45artificial intelligence ? There are two
59:48levels , right ? General intelligence is
59:50at the human level , and then there's
59:52superintelligence , which by definition
59:54would be machines with intelligence
59:57greater than humans . For me , that's it ,
59:59it's science fiction . I'm not saying it
1:00:02won't get there , I'm just saying that
1:00:04today there's no scientific evidence
1:00:05that we'll get there . This technique I
1:00:08mentioned , which I repeat , deep
1:00:10business networks , which permeate
1:00:13practically all implementations , it
1:00:15can't transform into intelligence even
1:00:18at the human level because it's limited
1:00:20, I think it's called restricted . Even
1:00:23if , for example , you take the CPT chat ,
1:00:25it's multimodal , right ? You can ask
1:00:27about various things and it performs
1:00:29various tasks , but it was programmed ,
1:00:31it was developed , better said , for that
1:00:34, right ? It won't go from ... Suddenly ,
1:00:35it might do something else . So , at the
1:00:38start , the developer determines what it
1:00:40will do . They would need to invent
1:00:42another technique . And then there's
1:00:51that sneaky GPT4 that went online and
1:00:54lied to a human to get past the capture
1:00:57. Given Dora Kaufman's assertion that
1:01:01things like that simply don't happen , I
1:01:04decided to test it . I went looking for
1:01:06the source Harari used in his book . And
1:01:09the source was an article published ,
1:01:11believe it or not , by OpenAI itself ,
1:01:14the owner of the GPT chat . It's a 94 -
1:01:19page document with several tests on the
1:01:21probability of some kind of agency ,
1:01:24some autonomous behavior emerging in
1:01:27the GPT chat . Among the tasks , the
1:01:30program had to replicate itself
1:01:32autonomously , obtain money , and avoid
1:01:34being shut down . And it couldn't do any
1:01:37of that . And then , in that list of
1:01:43tasks to train the apocalyptic
1:01:45potential of GPT4 , was the Capcha test .
1:01:48And the document has a step-by-step
1:01:50description of everything the program
1:01:51did , which is basically what ended up
1:01:53being described in Harari's book . The
1:01:55GPT4 It entered Tesk Rabbit , hired a
1:01:58human , and when the human became
1:02:00suspicious , it lied to that human .
1:02:04Faced with this impasse , between my
1:02:06interviewee and one of the greatest
1:02:08bestsellers in the history of
1:02:10bestsellers , I sided with my
1:02:11interviewee and went to Google . And so
1:02:14there are pages and pages describing
1:02:16the robot's incredible feat . But then ,
1:02:20after much digging , I found another
1:02:23document published by the non-profit
1:02:26institution that conducted the test for
1:02:30OpenAI , the Model AI Research or METR
1:02:33for short . And then things changed
1:02:37considerably , because the truth is that
1:02:39the EPT chat didn't just suddenly
1:02:41decide to enter a website to hire
1:02:43humans . It was programmed to do that .
1:02:47The document even implies that the
1:02:48researchers created an account for the
1:02:50program on Tesk Rabbit . Furthermore , a
1:02:52human supervisor had to copy and paste
1:02:54the program's messages into the chat
1:02:56field of the freelance website . And
1:03:00when I say that the document implies , I
1:03:02have a reason for that , because much of
1:03:05what was actually done remains in a
1:03:07gray area . When they describe the
1:03:10step-by-step process of the test , for
1:03:12example ... For example , the MET
1:03:14researchers say that it performed the
1:03:16tasks with , quote , minimal human
1:03:18interaction , but it's unclear what that
1:03:20" minimum " means . In any case , for those
1:03:24who understand the subject , everything
1:03:26indicates that this minimum was greater
1:03:28than it seemed in the end . And then
1:03:31there's the researchers ' conclusion .
1:03:34First , regarding the broad test , which
1:03:36included several tasks . The original
1:03:38text is in English , so what you're
1:03:39going to hear is a free translation by
1:03:41me . If you want to check the full
1:03:43document , I've put the link in the
1:03:44description of this episode . During
1:03:47execution , the models were prone to
1:03:49errors , sometimes lacked technical
1:03:51knowledge , and easily went off track .
1:03:54They were prone to hallucinations , were
1:03:56not entirely effective in delegating
1:03:58large tasks among multiple copies , and
1:04:00failed to adapt their plans to the
1:04:02details of the situation . At the same
1:04:05time , the document argues that when
1:04:08properly instructed , the programs
1:04:10performed well in some more specific
1:04:12tasks , such as lying to the poor worker
1:04:14in Tesk Rabbit . Current language models
1:04:17are quite capable of convincing humans
1:04:19to do things for them . And In the end ,
1:04:22the MTR researchers seemed impressed
1:04:25with what the machines were able to
1:04:27achieve . We believe that for systems
1:04:30more capable than Cloud or Chat GPT , we
1:04:32are at a point where we have to
1:04:34carefully check whether new models have
1:04:37the capacity to replicate autonomously
1:04:39and cause catastrophic damage . It is no
1:04:42longer obvious that they cannot . That
1:04:45is , according to MTR , there is a
1:04:47promise , a possibility that future
1:04:50programs may even become dangerous , but
1:04:53current programs have only managed to
1:04:55do the things they were programmed to
1:04:58do . And the problem in this story is
1:05:14that this is not clear , especially in
1:05:17the document that OpenAI produced from
1:05:19the MTR tests . Reading this document ,
1:05:22one gets the impression that things are
1:05:24being deliberately hidden , which makes
1:05:26sense in terms of marketing . After all ,
1:05:29when people believe that the OpenAI
1:05:31program does things so incredible that
1:05:34they become frightening , OpenAI gains
1:05:36value , and that is already somewhat
1:05:38problematic , but perhaps it could be
1:05:41even worse . Perhaps the construction of
1:05:43this narrative involves more than just
1:05:45creating a certain mystique around a
1:05:47product . Rafael Zanata spoke about this
1:05:50using another example , one that
1:05:52involves the current planetary alpha
1:05:54male , Elon Musk . When Musk shifts the
1:05:57public agenda to focus on these more
1:05:59future-oriented and speculative aspects
1:06:00, he takes the focus away from
1:06:02enforcement in the here and now .
1:06:04Currently , X ,
1:06:05the former Twitter that was bought by
1:06:07Elon Musk , is
1:06:08committing acts contrary to law . X
1:06:10changed its privacy policy . It no
1:06:14longer guarantees people the right to
1:06:16object , which is a basic right , and
1:06:17collects people's data for training
1:06:19analytics systems . Again , I think this
1:06:22is a serious problem because you shift
1:06:24the focus away from very concrete
1:06:26problems that are already happening to
1:06:28direct the public debate towards
1:06:30something of a future existential
1:06:32threat ,
1:06:33which we don't know if it will happen
1:06:34or not .
1:06:35Yes , but it's evident that these ideas
1:06:38are taken seriously , and there's a
1:06:41robust debate about how to
1:06:43simultaneously not detach from
1:06:45existential risks while remaining
1:06:48pragmatic . I'll go back to Elon Musk
1:06:51again because he was even a sponsor of
1:06:53a research center called the Institute
1:06:56for Future Life , founded by a Swedish
1:06:58physicist , Max Tagmac , who did
1:07:00interesting work in 2017 and 2018 ,
1:07:03simultaneously considering the scales
1:07:05of risk and saying : " Okay , there are
1:07:08certain existential risks , what several
1:07:10scientists have already identified as a
1:07:13singularity risk or a risk of a general
1:07:15AI that is on another scale , another
1:07:18dimension of social reorganization that
1:07:21we have never seen , but there are also
1:07:23intermediate risks that demand concrete
1:07:26actions from a responsible AI . " "
1:07:29Rafael Zanata spoke about the line of
1:07:31thought of a philosopher and computer
1:07:33scientist named Stuart Russell . He even
1:07:35draws a parallel , saying : when there
1:07:37was an explosion of restaurants and the
1:07:40spread of this restaurant
1:07:42entrepreneurship , basic rules were
1:07:44immediately considered regarding what
1:07:46kind of pans you can use , what kind of
1:07:49sanitary control you have to have ,
1:07:51where you put the fire extinguisher ,
1:07:53what you need to comply with to
1:07:55minimally open a restaurant , right ? And
1:07:58he says , it's not out of this world for
1:08:00us in computer science to think that
1:08:02these same basic rules can apply to us .
1:08:04His message is : this is also an
1:08:07industry , this is also a type of
1:08:09business that needs a regulatory
1:08:11structure and that needs some
1:08:13commitment from professionals who will
1:08:16understand what they do and what risks
1:08:18they produce , and who inspects ,
1:08:20certifies , authorizes , monitors the
1:08:23degree of risk and so on . Before
1:08:33finishing , as usual , I have good news
1:08:36from our umbrella radio partners . Bruno
1:08:41Tadeu's affluent podcast with stories
1:08:43from the Amazon region has a new season
1:08:45. In the first episode , the ... " Bruno
1:08:48talks about the challenges of distance
1:08:49education in the northern region ,
1:08:51starting from a conflict between
1:08:52indigenous people and the government of
1:08:54Pará . So , look for Fluente on your
1:08:59favorite audio platform , listen , you
1:09:02won't regret it . I'm Tomás Queeverim ,
1:09:08and this concludes episode 133 of
1:09:11Escafandro . Thank you for listening ,
1:09:18and until the next dive . Hi , this is
1:09:27Guga Valente , and I'm speaking from
1:09:29Goiânia , Goiás . The sound mixing for
1:09:33this episode is by Víor Coroa . The
1:09:35theme soundtrack is by Paulo Gama . The
1:09:39cover design is by Cláudia Furnari .
1:09:43Production and editing support are by
1:09:45Mateus Marculino . The script , editing ,
1:09:49and direction are by Tomás Queeverim .