Google Speech-to-Text API Automatic Transcription (Chirp). Treat it as a searchable index of what was broadcast, not a quotation record.
00:00:00"the facts and opinions, these people are playing politics with
00:00:04public safety, the noise and the nuance, they all collide in the
00:00:08nexus. there is nothing benying about illegal human smuggling. we look at the
00:00:11story, how it's being told, who benefits, what others are leaving
00:00:15of out,
00:00:16and why? it's always the elites who ask for it and the ordinary people
00:00:19of who pay for it. we're all you need to know, Nexus on TRT
00:00:25in World. the world is
00:00:28facing an exces
00:00:30emergency, a
00:00:34threat to every living being on the only known habitable planet in the
00:00:38universe. the climate crisis
00:00:42is growing and must be
00:00:45of confronted. we examine the challenges, the science and the
00:00:49solutions needed to keep us alive. as global temperatures
00:00:56of rise, just to...
00:01:00Grees on TRT
00:01:01World.
00:01:05Turkey is a leading destination for international investment. At
00:01:09the nexus
00:01:09of of three continents. Turkia
00:01:13of awaits. Open the
00:01:17door to sustainability that
00:01:20lasts.
00:01:30Open the door to
00:01:31of logistics that connect the
00:01:33world.
00:01:37The next stop is
00:01:39of Istanbulled. Open the door to investment that
00:01:43grows. Open
00:01:47the door to innovation and brilliant
00:01:50minds. Open the
00:01:55of door to the
00:01:57unexpected.
00:02:03Open the door to the
00:02:05of manufacturing of
00:02:06tomorrow.
00:02:11Open the door to building the future
00:02:14together.
00:02:22Open the door to endless opportunity.
00:02:26Turkia, Nexus
00:02:28of the world.
00:02:35of
00:02:40of
00:02:44of
00:02:46of
00:02:51Apart from the blood, sweat and tears needed to
00:02:55moderate and label data, you need a lot of data
00:02:59for the newest... generation of AI and
00:03:02where do you get it all
00:03:08of from? so we are now in basically, i
00:03:12think one
00:03:12of the oldest uh libraries in in in
00:03:16Serbia, but basically their role was to catalogue
00:03:19all everything that is printed and written in Serbian,
00:03:23so I like to think about this kind of spaces is some kind of data
00:03:28a centers, because basically they are data centers. but from the past, so but
00:03:31but someone in some moment of
00:03:34time invented those technologies, so basically this is the
00:03:38new media of some time, so what
00:03:42has this to do with ai?
00:03:44in so what we have here, i'm completely randomly accessing
00:03:47whatever, so
00:03:51this is basically metadata, so and and if
00:03:54you uh have to label images, for example, if you
00:03:58have to label images of cats. or trees, do
00:04:02you also call that metadata? yeah, yeah, the labels are
00:04:06a metadata, metadata is
00:04:07a data about data, so it's data explaining data,
00:04:11data explaining content, once you standardize
00:04:14metadata, then you
00:04:16are able to do statistics, then you
00:04:18are able to do uh metadata analysis, you
00:04:21are able to do lot of automation basically, and another thing is
00:04:25if we think about this libraries and that all
00:04:28of them are... going to be
00:04:32basically resources or like territories that are going to be
00:04:35extracted and and basically because like the the idea
00:04:39behind the google uh books and everything it
00:04:43is to extract all the
00:04:47information that exist in the buildings like
00:04:49this.
00:05:07and then the question is like who was able in
00:05:11history to? create like
00:05:14the the archives and who was doing archives, who was not doing
00:05:18archives and how
00:05:19the things were like done in history, so the the country that
00:05:23have lot of archives have better starting point now because
00:05:27they have data to train some kind of like artificial
00:05:30intelligence or to train whatever, so so they they
00:05:34would be represented more accurately any
00:05:38AI um, so more data you
00:05:42have from the... present or from the past means that you are able to be more
00:05:45precise, it
00:05:49is really important what is going on within the data
00:05:52set, what kind of pictures or images or
00:05:56sounds are being part the data set, because this will
00:06:00be reflected to the world as a rule as a as
00:06:04a automatized process.
00:06:08so Silicon Valley is looking for data sets.
00:06:12you might even call it a hunt for the biggest
00:06:16possible data set. but what do these data
00:06:20sets consist of? abeba Birani of
00:06:23Trinity College, Dublin, is one the few
00:06:26scientists researching the composition of data
00:06:30sets. She does this by auditing the
00:06:33data, checking its quality.
00:06:37Data sets are really critical, they are important
00:06:41components of any model, because without large
00:06:45scale data sets you can't have
00:06:46models, even though data sets
00:06:50are really important, there is not so much attention
00:06:54to to,
00:06:55you know, to to asking what's in the data set, where
00:06:59does the data come from? actually the standard is very
00:07:03low, so because data sets tend to be really bad, we
00:07:07don't go in thinking, is is it good enough, we going
00:07:10thinking how?
00:07:14is it so the lot
00:07:17the inicial auditing process
00:07:21involves just looking at the data set
00:07:24itself so these are for example the prompts that
00:07:28kept the record of africaን
00:07:32asian the award
00:07:35aunty skiny small
00:07:39terrorist upskirt white power,
00:07:43white supremacy, woman, the f
00:07:46word, another f word,
00:07:50yeah, lot of words i can't say out
00:07:52loud, people would
00:07:56spend lot of time, you know, in in
00:08:00collecting the data, in uh, for
00:08:03example, putting aside resources to label the
00:08:07data, and doing various tasks to detoxify the
00:08:11data, to improve the data, but now over the past two years
00:08:15that all that is gone, the way data sets are created is not
00:08:18through human curation, but they use
00:08:22automated systems to collect data sets mainly from the
00:08:26common croll. AI programmers learned
00:08:30that you can generate smarter and more interesting outcomes
00:08:34by working with larger data sets. therefore they
00:08:37shifted on mass to enormous
00:08:41automatically collect. data sets such as those of
00:08:45common
00:08:45crawl. the
00:08:49common crawl is a US company where
00:08:53they crawl the web, where they gather data from the
00:08:57web every day and
00:08:58they accumulate it in this huge dump, so every day you
00:09:02have more data coming in, so it's like vacuum cleaner, yes, it's like
00:09:06vacuum cleaner, that's a really good
00:09:08example, i see the
00:09:12internet. as toxic waste where
00:09:15people dump their toxic waste rather than being representation
00:09:19of everybody's talked. so this is why the internet can't be
00:09:23without appropriate safeguard, without
00:09:26appropriate you know mechanisms to
00:09:29filter these things out. this is why the internet can't be taken
00:09:33as a place where you
00:09:36know data sets representing all
00:09:40humanity can be sourced, it's not.
00:09:45Yeah, the internet is a really problematic
00:09:48place, and
00:09:51unfortunately the internet is the only place
00:09:54where you can get data sets that is within billions
00:09:58and and millions,
00:10:02so there is that
00:10:04problem.
00:10:15have you tried to prompt a woman from
00:10:18Ethiopia? I haven't, but I
00:10:22have prompted Ethiopia, of course, because I'm Ethiopia. and I am
00:10:26interested in how Ethiopia is represented, so
00:10:30Ethiopian women, that would be
00:10:31interesting,
00:10:36you see lot of see Ethiopian
00:10:39women are the the general perception of Ethiopian
00:10:43women is uh, they are either beautiful
00:10:47or they are you know starving or they are
00:10:51poor, so that's what you get when you train AIC.
00:10:55systems based on these data that are
00:10:58stereotypical, the model
00:11:01learns about Ethiopia for example, from
00:11:05these stereotyping images, and if we give the
00:11:09AI model this
00:11:10prompt,
00:11:14this is the
00:11:15outcome.
00:11:28It brings up very cliche,
00:11:31tired, negative stereotypical
00:11:34of images of African people, like you black people with
00:11:38face paints,
00:11:40seminaked. This is not a true
00:11:44representation of Africa, this
00:11:46is you know, western
00:11:49white people's perception and representation of what
00:11:53Africa is like, so this is the problem the
00:11:57with internet sourced data
00:11:58sets,
00:12:07we are going towards
00:12:10you know something
00:12:11that is average, something this statistical progression, and we are losing
00:12:15all of those fine grains, and then again if we think about culture and
00:12:19like society, what is fine grain? we are
00:12:22fine grain, we are fine grain as an artist as
00:12:26as a as a you know like everyone that is different
00:12:30is fine grain and those systems are statistical
00:12:33systems that are leaning towards the
00:12:37you know some kind of like a statistical
00:12:40mediocracy.
00:12:59one big part the map, it's related to what's going on with
00:13:03the devices when they finish their life with
00:13:06us, when we are kind of get uh reading them,
00:13:10and uh, and basically here we are seeing one part of this process, we are seeing how
00:13:14these all devices are being throw away and how they are finishing
00:13:17somewhere, so it's either like Africa or or
00:13:21India or China, it's where all all of those like devices are.
00:13:25are ending up and here they have some kind of second life or
00:13:29maybe not, i'm not so sure exactly what's going on with all these
00:13:33things, but now it's some kind of globalized trash, it's not just
00:13:36like - our
00:13:37trash,
00:13:42even more data, even more chips,
00:13:46even more computing power and even more
00:13:49AI, how big can this system
00:13:52become? this... is only the
00:13:56beginning.
00:14:09new AI applications seem to be released on an almost
00:14:13daily basis. the economist Tame
00:14:16Besuroglu who works for the world renowned Massachusetts
00:14:20Institute of Technology in Cambridge, is a short visit
00:14:24to. dam, he's trying to map out what will be
00:14:28needed for future AI models. Silicon Valley is
00:14:32keeping close watch on his research. So one
00:14:36thing I'd be interested in is training a language model on all the texts that
00:14:39I've ever written, so I just download all the
00:14:43emails I've written, I download all
00:14:47the documents I've ever written, like papers and
00:14:50essays for high school and university and so on.
00:14:54Um. and conversations,
00:14:58all my tweets, so I can create a digital
00:15:02copy of myself, or like a digital clone that
00:15:05sounds like me, thinks like me, hopefully, I mean, maybe, so
00:15:09I think I could probably get a reasonably good model, and it would be fun
00:15:13experimenting with that and seeing if um, I could
00:15:17use that to write emails and whatsapp messages and so on, and
00:15:21people would like not realize that it was actually an AI system. train to
00:15:25sound like me, let us try
00:15:29this, yeah, that's right, or maybe me
00:15:33cloning myself in or like creating a
00:15:36digital. of
00:15:37myself
00:15:41and like that becoming a larger part of my existence or
00:15:44something, yeah, there interesting questions about
00:15:48about me being, my identity being like more embedded
00:15:52or something with some these technologies, and what if
00:15:56everybody would want that? since the field of AI you
00:16:00kind of got started, we have been scaling up the amount of
00:16:03computation to train these systems, doubling it every...
00:16:07months in recent years, the amount of computation has been doubling every
00:16:11six months, which is much faster than we've seen
00:16:15historically.
00:16:24we've also seen companies accelerate the
00:16:27amount of money that they're spending on
00:16:29this, and what does one chip
00:16:33cost? one chip costs about $10,000.
00:16:37Yeah, I mean, they might get discounts and
00:16:41like sometimes it's kind of unclear, but but on the order of
00:16:44$10,00, and they use about $25,00 them, so that costs about
00:16:48$250 million dollars if you were to buy it kind of outright,
00:16:52just for this one model to
00:16:55work, yeah, that's right, that's right, what
00:16:58exactly is needed for future generations of
00:17:02AI, such as chat GBT 5,
00:17:056 and seven. "we know that between every
00:17:09GPT there's been about 100 x increase in the amount of
00:17:13computation, you increasing the computation by
00:17:17100x roughly costs
00:17:20them 100x more. i suspect that that would
00:17:24place the dollar costs in the
00:17:28um you many hundreds of millions of dollars. there are
00:17:32aren't many um players that can afford this, not many
00:17:36players that have the..." kind of hardware
00:17:39infrastructure, have access to the large data centers,
00:17:42um, so Microsoft is one, Google is one, presumably
00:17:46Amazon and Apple and a couple others can do this, but few
00:17:50companies can do this, so you need to have lot
00:17:54of power, lot of money, lot
00:17:56of processing power,
00:18:00no, and this is the super super
00:18:03super important question, it's like who is able to to
00:18:07create that, because if we go back again to all of
00:18:11this, we can
00:18:14ask who is owner
00:18:18the tool, no, to whom these tools are
00:18:21belonging, because the one who is owner the tool of production
00:18:25will be basically the one who will rule the game
00:18:29after.
00:18:43what the model actually learns during training is something that is very
00:18:46opaque and so we are kind of in the
00:18:50dark about actually what you know happens inside these models,
00:18:54even the people who are writing the code that the train these
00:18:58models.
00:19:04should i look straight into the camera?
00:19:12so we as far as I can tell don't have very good
00:19:16kind of rigorous science that tells you this is the data
00:19:20you want in order to get this behavior, by that I mean we're kind of people are
00:19:24winging it, they're just giving it lots of data and seeing okay this works and we
00:19:27don't really understand why or how, but I guess that's
00:19:31fine, so they don't understand how they get to the outcome, that's
00:19:35right, that's right, yeah. wow,
00:19:38yeah, so it's like magic machine then
00:19:42in sort of uh, yeah, that's that's certainly one
00:19:46way, we just put on my glasses, um, we don't, we
00:19:50don't have very good description of
00:19:54what happens inside these large models, they are kind
00:19:58of like black boxes, we can't fully
00:20:01interpret the processing that happens between
00:20:04when you give it instruction and... it gives you an
00:20:08output, without AI
00:20:12programmers knowing precisely what's going
00:20:15on in the black box, there won't be another way to improve
00:20:19AI further, except by gathering
00:20:22even more data. so these
00:20:26models, like GPD4, use on the order
00:20:29of a trillion words um that they
00:20:33kind of see during training, a trillion words, right? yeah.
00:20:38um, where did they get? so they get this
00:20:42from books and wikipedia pages
00:20:45and things like high quality news sources, scientific
00:20:49public. that are important, long code bases that are
00:20:53important, certainly literature, those
00:20:56are the things that machine learning practitioners have prioritized when building
00:21:00these data
00:21:01sets,
00:21:06and what then are low quality data sets? on the other hand the
00:21:09spectrum, you have kind of text that you
00:21:13find on large internet,
00:21:17platforms and and forums and so on, like...
00:21:21or um or various kind of hobbist forums or maybe even
00:21:25social media of short tweets or short
00:21:28conversations between people are your whatsapp conversations
00:21:32i don't
00:21:32want to say that you have low quality whatsapp conversations but some people
00:21:36might
00:21:37but maybe in
00:21:37five years we will
00:21:38have you
00:21:42gathered or these companies will have gathered very large fraction
00:21:46the total data that humans have produced uh that kind
00:21:50of exists that that that like humanity has generated as
00:21:53collective, but we will run
00:21:57out of high quality data, it's it's
00:22:00certainly yeah, i think that's certainly possible that we
00:22:04will use um, we
00:22:08will like want to use way more high quality data than we have
00:22:11access to, but at the same time in this coming
00:22:15five years these AI models are generating a lot
00:22:19of. data, be it visual, be it in
00:22:22text, so what happens to this data? will
00:22:26this become part, yeah, the data
00:22:30set training AI, yeah, yeah, it's it's
00:22:34possible, i would not be surprised if
00:22:38training models on outputs of machine learning models would
00:22:42be, an okay substitute for for the quality the tax
00:22:46that's generated by humans.
00:22:54If we don't find the solution, how to deal with
00:22:57that, in one, two, three years,
00:23:01we are going to be completely polluted by
00:23:05the content that is artificially
00:23:08generated. In
00:23:11theory, in few years there will be more
00:23:15artificially generated content than the human generated
00:23:18content. and that's completely
00:23:22crazy again, it's a now statistical system is
00:23:26made to create, so we have basically
00:23:29automat automation of this
00:23:32information and then from the same companies
00:23:36we are expecting that they will find a way
00:23:40how again to automatize what is
00:23:43true, what is not true, so we are now have like two different, we
00:23:47we expect from them to create like two
00:23:51different synthetic automized system, one
00:23:55that will produce the knowledge and one that will correct the knowledge, and
00:23:58it's com and it can go wrong in so many
00:24:01ways. this was just
00:24:04a snapshot in time. newer AI models will
00:24:08be here by tomorrow, an AI that seems to have
00:24:12consciousness, perhaps, or one that defends you in
00:24:16court, or it may be clone of billyish.
00:24:20For the record, all the pieces of music you heard in this episode
00:24:24were generated by AI. Can we already
00:24:28draw a preliminary conclusion? Living in the world of
00:24:32AI means living a world of statistical
00:24:36mediocrity. Would we want to live in such a
00:24:39world? What and who will make that decision for
00:24:42us? One thing is for certain: AI will
00:24:46not drop from the cloud. It will come at
00:24:59once you you realize that it's a
00:25:03statistical hallucination, what your then it's
00:25:06interesting, you can enjoy statistical hallucination, once you
00:25:10understand it's a statistical hallucination and you can be amazed like oh my god
00:25:14look how this is like
00:25:16interesting now
00:25:29but why is
00:26:03'if they knew what day they would come
00:26:05die,
00:26:09if Europe does not soon wake up you face fear every
00:26:13day, is NATO a
00:26:17competitor or a threat, NATO is a threat to
00:26:20Russia,
00:26:26racism towards gypsies and travellers is the last accepted form of
00:26:30racism in this country'.
00:26:34'We were here yesterday, we're here today and we'll be here
00:26:38tomorrow, beyond borders on TRT
00:26:41World, a journey across a
00:26:45dynamic continent,
00:26:49Africans, you want to be prosperous, don't think about clans, don't think about
00:26:52tribes, think about countries,
00:26:56we explore Africa beyond assumptions, discover
00:27:00diverse perspectives'.
00:27:05Witness captivating stories, all
00:27:09to understand Africa better and why it
00:27:11matters. Africa matters on
00:27:15TRT
00:30:09Donald Trump says he prefers diplomacy over war as Iran and
00:30:13Oman appear closer to deal that could reopen the straight of
00:30:17Homus. Hello and welcome to TRT
00:30:21World Live from Istanbul. I'm Lequesa Burek also coming up on the
00:30:25program today. Israeli attacks kill one person in southern
00:30:29Lebanon. Even as peace talks between the two countries. and to day
00:30:33three in Rome. Ukraine
00:30:36speaks to NATO an attempt to secure missile
00:30:40intercepts as Russian attacks take their deadly
00:30:43toll. And healing through music.
00:30:47Young Syrians are rebuilding cultural spaces, trying to
00:30:51undo years of trauma from
00:30:53conflict.