Latest / Chief in Tech Deep Dive Shortcast / Chief in Tech S1E35 - Wendy Gonzalez : Changing the World Through AI and Training Data

Transcript

Chief in Tech S1E35 - Wendy Gonzalez : Changing the World Through AI and Training Data · Chief in Tech Deep Dive Shortcast

• 0:00
  I want you to imagine, just for a second, that
• 0:02
  you are sitting in the passenger seat of a brand
• 0:04
  new self -driving car. OK, I'm picturing it.
• 0:07
  Right. And you are cruising up to a busy, completely
• 0:10
  chaotic city intersection. You've got... Pedestrians
• 0:15
  stepping off the curb while looking at their
• 0:17
  phones. You've got a person on a bicycle weaving
• 0:21
  aggressively through traffic, and the sun is
• 0:23
  setting, casting these long, weird shadows from
• 0:26
  the buildings above. Basically a nightmare scenario
• 0:29
  for any driver. Exactly. Now, the car stops perfectly.
• 0:33
  It waits for the bike, ignores those confusing
• 0:36
  shadows, and then smoothly makes the turn. Which
• 0:39
  is... Pretty incredible. It is. Most of us see
• 0:41
  that and we just think, wow, technology is basically
• 0:44
  magic now. But how does that car actually know
• 0:46
  the difference between a real person, a bicycle
• 0:49
  and like a dark patch of asphalt? Right. It's
• 0:52
  definitely not magic. Not at all. Yeah. It's
• 0:54
  actually the result of a massive, incredibly
• 0:56
  tedious and entirely hidden human effort. I'm
• 1:00
  your host, by the way, and I'm thrilled to get
• 1:01
  into this. And I'm your resident expert for today.
• 1:05
  We definitely have this illusion that artificial
• 1:07
  intelligence just spontaneously understands the
• 1:12
  physical world. Like it just wakes up one day.
• 1:14
  Yeah, exactly. Like it wakes up on a server farm
• 1:16
  somewhere, downloads the internet, and suddenly
• 1:19
  knows what a stop sign looks like in a snowstorm.
• 1:22
  It's just so funny to think about. But the reality
• 1:24
  is far more grounded. It's built on a foundation
• 1:28
  of human labor that is much more intensive than
• 1:32
  most people ever realize. And that hidden foundation
• 1:35
  is exactly what we are diving into today. This
• 1:38
  is our deep dive into the invisible human workforce
• 1:41
  powering global artificial intelligence. It's
• 1:44
  a huge topic. It really is. Specifically, we're
• 1:46
  looking at the world of ethical AI training data.
• 1:49
  Our source material for this is a fascinating
• 1:51
  presentation delivered by Wendy Gonzalez. Oh,
• 1:53
  she's brilliant. She is. She's the acting CEO
• 1:55
  and president of an organization called Samisource.
• 1:58
  Right, sometimes referred to as SOO in the materials.
• 2:01
  Yes, exactly. And she gave this talk at a Women
• 2:04
  in Tech Network conference. Her mission in this
• 2:07
  presentation really struck me. We are going to
• 2:10
  explore how the literal act of teaching an AI
• 2:13
  to see and think is, simultaneously, being used
• 2:17
  to lift tens of thousands of people out of extreme
• 2:20
  poverty. This is a really powerful story. It
• 2:22
  is. And along the way, we'll see why this model
• 2:24
  completely upends what we assume about the tech
• 2:27
  industry's supply chain. Okay, let's unpack this
• 2:30
  yet to truly grasp the sheer scale of what Wendy
• 2:34
  Gonzalez is talking about We have to strip AI
• 2:36
  down to its its fundamental building blocks,
• 2:40
  right? The absolute basics exactly before we
• 2:42
  even talk about the people sitting at the computers
• 2:44
  doing this work We have to look at why this specific
• 2:47
  type of labor is so Unbelievably valuable to
• 2:50
  the largest companies on the planet like Google
• 2:52
  and Walmart, right? We're talking about training
• 2:54
  data, because you literally can't have artificial
• 2:57
  intelligence without it. Right. Before a machine
• 2:58
  can see, hear, or speak, it has to be taught
• 3:01
  what it's looking at. Yep. And it is taught using
• 3:04
  labeled data. So going back to our self -driving
• 3:07
  car at the intersection, it doesn't intuitively
• 3:09
  know what a pedestrian looks like. Not at all.
• 3:12
  Someone, an actual human being, had to sit at
• 3:15
  a screen. look at an image from a street camera,
• 3:18
  draw a digital box around a person, and tag it
• 3:21
  like this is a pedestrian. And they had to do
• 3:23
  that millions and millions of times. From every
• 3:26
  conceivable angle. Yes. What's fascinating here
• 3:29
  is that this manual repetition isn't just a small
• 3:31
  part of the process. It is the single biggest
• 3:33
  bottleneck in the entire tech industry. Wow.
• 3:36
  Really? Yeah. Gonzalez points out in her presentation
• 3:39
  that by 2023, nearly $100 billion was projected
• 3:43
  to be invested in artificial intelligence. A
• 3:46
  hundred billion? That's insane. It's massive.
• 3:49
  But here is the massive catch that most people
• 3:51
  outside of Silicon Valley completely miss. Eighty
• 3:55
  percent of the time consumed in most AI and machine
• 3:58
  learning projects is spent just dealing with
• 4:00
  this training data. You mean preparing it, cleaning
• 4:03
  it, labeling it? Exactly. 80%. That statistic
• 4:06
  is just wild to me. 80 % of the time. It makes
• 4:10
  me think of training data as basically flashcards
• 4:13
  for a billion dollar toddler. Oh, yes. That's
• 4:16
  exactly what it is. I mean, you have this incredibly
• 4:18
  powerful, super expensive AI system, but it's
• 4:22
  fundamentally naive. It knows absolutely nothing
• 4:25
  about the physical world until a human sits down
• 4:27
  with it and holds up a million flashcards saying,
• 4:30
  you know, car, tree, road sign. It really is
• 4:33
  shocking when you put it like that. It is. It's
• 4:35
  honestly shocking that the most cutting edge
• 4:37
  technology in human history is entirely dependent
• 4:40
  on manual human repetition. A billion -dollar
• 4:44
  toggler is the perfect way to look at it because
• 4:46
  the algorithms themselves, the brilliant complex
• 4:49
  math that the engineers write, they are functionally
• 4:52
  useless without the ground truth provided by
• 4:55
  that human labeling. Ah, ground truth. Yeah.
• 4:57
  If the data is bad, if the human incorrectly
• 5:00
  labels a pedestrian as a shadow, that $100 billion
• 5:02
  industry grinds to an absolute halt. The algorithm
• 5:06
  learns the wrong lesson. Exactly. And then the
• 5:08
  self -driving car makes a fatal error. Man, it's
• 5:11
  wild to think that these hundred billion dollar
• 5:14
  algorithms are entirely at the mercy of whoever's
• 5:17
  clicking the mouse to draw those boxes. It really
• 5:19
  puts it into perspective. It does, which makes
• 5:22
  me wonder how these massive companies actually
• 5:24
  source that labor. If the tech world has this
• 5:27
  bottomless, desperate hunger for flawlessly labeled
• 5:31
  data, who is getting paid to sit there and do
• 5:34
  it? Well, traditionally, the tech industry has
• 5:36
  treated data labeling as low level gig work.
• 5:39
  Outsourcing. Yeah, exactly. It's outsourced to
• 5:42
  massive click farms where workers are paid like
• 5:44
  pennies per task with zero job security, right?
• 5:48
  It's treated as highly expendable temporary labor.
• 5:52
  But that brings us directly to Sama sources radical
• 5:54
  philosophy, which completely flips that traditional
• 5:57
  model on its head. I love this part. Yeah, to
• 5:59
  understand their approach, you really have to
• 6:01
  look at their late founder, Leila Janna. She
• 6:04
  founded the organization in 2008 with one very
• 6:07
  specific driving ethos. Talent is distributed
• 6:10
  equally while opportunity is not. Yes, exactly
• 6:12
  that. That is such a fascinating lens to view
• 6:15
  the global economy through, because it totally
• 6:18
  rejects the idea that a lack of wealth equals
• 6:21
  a lack of capability. It does. Jaina believed
• 6:24
  that traditional charity wasn't the answer to
• 6:26
  systemic poverty. Her goal was giving work instead
• 6:29
  of giving aid. Which is a huge paradigm shift.
• 6:32
  Right. So Samasource's demographic target for
• 6:35
  hiring is highly specific and very unusual for
• 6:38
  a tech company. They intentionally recruit people
• 6:41
  who are living below the World Bank average poverty
• 6:44
  line. Wow. And crucially, they require absolutely
• 6:47
  no prior education, no college degree, and no
• 6:51
  formal job experience. Which brings us to that
• 6:53
  incredibly vivid example from the presentation,
• 6:56
  the story of a man named Ken Kihara. Oh, Ken's
• 6:59
  story is amazing. It really is. So Ken lived
• 7:01
  in Mathar, which is one of the largest, most
• 7:03
  densely populated slums in Kenya. And Gonzalez
• 7:06
  notes that in Mathar, unemployment for youth
• 7:08
  and women hovers around 70%. Or even higher,
• 7:11
  yeah. Right. Now Ken had amazing grades coming
• 7:13
  out of high school. He clearly had the raw talent.
• 7:16
  But no opportunity. Exactly. He had absolutely
• 7:18
  no network, no access to capital, and no way
• 7:21
  to pay for a college degree. He was entirely
• 7:24
  locked out of the formal economy. And you know,
• 7:26
  when someone with that much potential is locked
• 7:28
  out of the formal economy, they are pushed into
• 7:31
  incredibly dangerous survival situations. Just
• 7:34
  to get by. Yes. Because he couldn't find safe
• 7:37
  formal work, Ken was forced into the informal
• 7:40
  economy just to feed himself. The presentation
• 7:43
  details how he became a producer of Chancho.
• 7:45
  Right, which is a local moonshine. But Gonzales
• 7:48
  specifies this gritty, terrifying detail about
• 7:51
  the environment Ken was working in. The jet fuel
• 7:53
  detail. Yes. She points out a photo during her
• 7:56
  presentation explaining that the specific batch
• 7:59
  of Chandra Ken was making was literally brewed
• 8:02
  with jet fuel. It's hard to even comprehend.
• 8:04
  Just let that sink in for a second, right? A
• 8:06
  young guy with top high school grades is pushed
• 8:08
  into brewing highly toxic combustible moonshine
• 8:11
  in a slum entirely because he was born in a neighborhood
• 8:14
  with a 70 % unemployment rate. Yeah. That is
• 8:17
  exactly where Samasour stepped in, offering him
• 8:19
  training and his first formal job in data annotation.
• 8:23
  And it completely changes his trajectory. He
• 8:25
  goes from mixing toxic chemicals to survive to
• 8:29
  training the algorithms of Fortune 50 companies.
• 8:32
  OK, but I want to look at this from Google or
• 8:33
  Walmart's perspective for a second. Sure. Wait.
• 8:36
  If companies like Walmart and Google need flawlessly
• 8:39
  labeled data for critical systems, isn't it a
• 8:42
  massive risk to hire people with absolutely zero
• 8:45
  tech background or college degrees? Sounds risky,
• 8:48
  right? I mean, you are building a multi -billion
• 8:51
  dollar self -driving car system or a massive
• 8:54
  global search engine. Are you really going to
• 8:57
  entrust the foundational safety data, the ground
• 9:00
  truth of your algorithm, to someone whose last
• 9:02
  job was brewing jet fuel moonshine in a slum?
• 9:05
  That is the exact skepticism traditional business
• 9:08
  leaders throw at this model. I don't guarantee
• 9:10
  quality. Right, but that skepticism fundamentally
• 9:12
  misunderstands the genius of what Samisource
• 9:14
  built. Let's go back to how the traditional tech
• 9:17
  industry handles this. gig work. Right. The click
• 9:19
  farms. Exactly. When you treat data labelers
• 9:22
  as disposable temporary gig workers, they have
• 9:25
  absolutely no incentive to care about the quality
• 9:27
  of the data. They just want to get paid. Yeah.
• 9:29
  They just want to click as fast as possible to
• 9:30
  make their pennies. The result is incredibly
• 9:33
  high turnover, burnout, and terrible quality
• 9:37
  control. Right. If you treat someone like a disposable
• 9:39
  cog, they are going to give you disposable work.
• 9:42
  Exactly. Samisource recognize that to get premium
• 9:45
  data, you actually need stability. So they provide
• 9:49
  full -time employment, not gig contracts. It's
• 9:52
  just huge. And they provide living wages that
• 9:55
  are intentionally calculated to support a family
• 9:58
  of four. They provide real health benefits. So
• 10:01
  it's a real career. It is. By taking someone
• 10:04
  like Ken, who has previously only known economic
• 10:07
  chaos. and placing him in an environment of profound
• 10:10
  professional stability, they breed incredible
• 10:13
  loyalty and focus. That makes total sense. Right.
• 10:15
  This social mission actually solves the tech
• 10:17
  world's biggest quality problem. Here's where
• 10:20
  it gets really interesting. Samasaurs isn't just
• 10:23
  a charity project that these tech giants tolerate
• 10:25
  for some good PR. Not at all. Their social impact
• 10:28
  mission is precisely what gives them their competitive
• 10:30
  advantage in the marketplace. By treating people
• 10:33
  as long -term investments rather than short -term
• 10:35
  expenses, they end up with better, cleaner data
• 10:38
  than the traditional gig economy tech companies.
• 10:41
  And the retention numbers completely validate
• 10:43
  that approach. Gonzalez notes that Samisource
• 10:46
  boasts industry -leading retention rates. They
• 10:49
  have less than 10 % annualized attrition. Wait,
• 10:52
  less than 10 %? Yes. In the tech support and
• 10:55
  data labeling world, a sub -10 % turnover rate
• 10:58
  is practically unheard of. I believe it. Because
• 11:02
  these jobs are so deeply meaningful to the workers,
• 11:05
  they stay. And because they stay year after year,
• 11:08
  they transform from absolute beginners into deeply
• 11:11
  specialized, world -class labeling experts. And
• 11:14
  the stakes of that acquired expertise are literally
• 11:17
  life and death. Going back to our self -driving
• 11:18
  car example, Gazalos points out that an AI must
• 11:21
  be trained to perfectly distinguish between a
• 11:23
  motorcycle and a bicycle. Or it must be able
• 11:26
  to instantly read specific vital road signs while
• 11:29
  ignoring irrelevant billboards. What they call
• 11:32
  the last mile of quality. Yes, the last mile
• 11:35
  of quality. You cannot get that from a burned
• 11:37
  out gig worker who started yesterday and is rushing
• 11:40
  to make minimum wage. You need a dedicated expert.
• 11:44
  And it goes beyond autonomous vehicles, too.
• 11:46
  Think about the Walmart example she provided.
• 11:48
  OK. Walmart has the world's largest catalog of
• 11:51
  products, and that catalog is shifting by the
• 11:53
  minute. Oh, I'm sure. Millions of items. Millions.
• 11:57
  Yes. Items being added, removed, changing seasons.
• 12:00
  If Walmart introduces a new seasonal aisle, the
• 12:03
  AI might miscategorize a plastic Halloween pumpkin
• 12:07
  basket as fresh produce. That would be a mess.
• 12:11
  Right. If their product search algorithms aren't
• 12:14
  constantly being audited and retrained by human
• 12:16
  experts who understand nuance, the search function
• 12:19
  breaks down entirely. And consumers wouldn't
• 12:21
  be able to find what they're looking for. Which
• 12:23
  means lost revenue. It requires a highly accurate,
• 12:26
  ongoing, human -in -the -loop system. Which brings
• 12:29
  up the reality of how this work is actually done
• 12:31
  on the ground. During the presentation, she showed
• 12:33
  a picture of their delivery center in Kampala,
• 12:36
  Uganda. Oh, that facility is stunning. It really
• 12:38
  is, and I think this completely shatters the
• 12:40
  preconceived notions a lot of people might have.
• 12:43
  This is not some dark, crowded sweatshop. No,
• 12:46
  definitely not. You walk off a busy street in
• 12:48
  Kampala and enter a highly secure center with
• 12:51
  biometric access, top tier technology and facilities
• 12:54
  built with B Corp certified sustainability. It
• 12:58
  looks like it belongs in Cupertino. If we connect
• 13:01
  this to the bigger picture, what we are seeing
• 13:03
  is that investing heavily in human infrastructure
• 13:06
  is the prerequisite. for superior machine infrastructure.
• 13:10
  That's a great way to put it. Yeah. And Gonzalez
• 13:12
  talks about this concept of machine learning
• 13:15
  assisted annotation. I'm glad you brought that
• 13:17
  up, actually, because the phrase machine learning
• 13:19
  assisted annotation sounds like deep corporate
• 13:21
  jargon. How does that actually work in practice?
• 13:24
  It's a fascinating feedback loop, really. The
• 13:27
  basic idea is that as the software platform gets
• 13:29
  somewhat smart, it takes over the easy, repetitive
• 13:32
  parts of the image. Like identifying the sky
• 13:34
  or the road. Exactly. And this frees up the human
• 13:37
  experts to handle the highly nuanced, uniquely
• 13:40
  human edge cases. Give me an example of an edge
• 13:43
  case. OK, imagine a pedestrian. walking down
• 13:46
  the street carrying a large, highly reflective
• 13:48
  mirror. An early AI system might look at that
• 13:52
  reflection of the street and think, there are
• 13:54
  two pedestrians here. Or worse, there is a hole
• 13:57
  in reality. Because it doesn't understand context.
• 14:00
  It lacks human context. The human expert in Kampala
• 14:04
  spots that error. draws the correct boundary
• 14:06
  around the pedestrian in the mirror, and tags
• 14:09
  it properly. So they aren't just fixing a single
• 14:11
  photo. No, they are actually codifying their
• 14:14
  human knowledge into the system. When the expert
• 14:17
  corrects that mirror reflection, that feedback
• 14:20
  alters the neural network's weights. Wow. The
• 14:23
  platform gets smarter over time, specifically
• 14:25
  because the humans training it are treated as
• 14:28
  respected experts capable of nuanced judgment
• 14:30
  rather than just clicking machines. So the technology
• 14:33
  elevates the human's capabilities and the human's
• 14:36
  nuanced judgment elevates the technology. Exactly.
• 14:39
  It's a brilliant synergy. And when you look at
• 14:41
  the leadership team driving this, it becomes
• 14:43
  clear that this philosophy of valuing overlooked
• 14:45
  talent isn't just a happy accident. No, it's
• 14:48
  very intentional. It is built into the structural
• 14:51
  DNA of the entire organization. Diversity and
• 14:54
  inclusion are so often used as like hollow buzzwords
• 14:58
  in the tech industry, you know, usually just
• 15:00
  a slide in an HR presentation. Right, checking
• 15:02
  a box. But at Samisource, it is a structural
• 15:05
  mandate. They purposefully mandate that 50 %
• 15:08
  of their hires are women. Which is amazing. And
• 15:11
  Gonzalez was very explicit in her talk that this
• 15:14
  isn't just about padding the stats at the entry
• 15:16
  level data labeling positions. Oh, really? Yeah,
• 15:19
  this 50 % mandate extends upward through middle
• 15:22
  management, executive leadership, and all the
• 15:25
  way to the board of directors. That structural
• 15:27
  commitment was really put to the test during
• 15:29
  their 2019 Series A funding round. Yes, that
• 15:32
  was a pivotal moment. Gonzales shared a story
• 15:34
  about how she and Leila John were out raising
• 15:36
  $14 .8 million. And she made a really poignant
• 15:40
  observation about what it's like to be in those
• 15:42
  high stakes venture capital meetings. Oh, I remember
• 15:44
  this part. Yeah. As two minority women executives
• 15:47
  seeking multimillion dollar funding for a software
• 15:50
  company, they were very often the only women
• 15:52
  in the room. That can be an incredibly daunting
• 15:55
  environment. Especially when pitching a business
• 15:57
  model that, quite frankly, defies conventional
• 16:01
  Silicon Valley logic. Right, pitching social
• 16:03
  impact to traditional VCs. Exactly. Yet they
• 16:06
  found investors who didn't just passively tolerate
• 16:09
  their diversity or their social mission, but
• 16:12
  actively valued it as a business strategy. Which
• 16:14
  is so rare. It is. They found financial backers
• 16:18
  who understood their technical AI expertise,
• 16:21
  but also recognized that their diverse workforce
• 16:24
  and deeply ethical supply chain were massive
• 16:27
  strategic assets. It's impossible to talk about
• 16:30
  that grand vision without acknowledging the deeply
• 16:32
  tragic part of this story. Yeah, Leila's passing.
• 16:35
  Leila Janna, the founder who pioneered this entire
• 16:38
  concept of using a tech business as a direct
• 16:41
  force for social change, passed away from a rare
• 16:43
  aggressive cancer in January 2020. It was a huge
• 16:47
  loss. It really was. But Gonzalez makes it very
• 16:50
  clear that her legacy, her fundamental DNA, remains
• 16:53
  entirely embedded in how the company operates
• 16:55
  today. You see Leila's legacy most clearly when
• 16:58
  you look at the individual lives that have been
• 17:00
  transformed by this model. The presentation highlighted
• 17:03
  another worker named Martha. whose story is just
• 17:06
  as compelling as Cannes. Right, Martha. Martha
• 17:08
  is a young woman who was completely locked out
• 17:10
  of the workforce. She had the drive, but she
• 17:13
  couldn't get a job simply because of her gender
• 17:16
  and because she lived in a marginalized slum
• 17:17
  area. She faced compounding barriers. But then
• 17:20
  she was recruited by Samisource. And the trajectory
• 17:24
  of her life didn't just change slowly, it changed
• 17:27
  dramatically. It did. After joining and earning
• 17:30
  a stable living wage, she was able to move entirely
• 17:34
  out of the slums and lease a formal apartment.
• 17:36
  Which is life -changing on its own. Completely.
• 17:39
  But the story doesn't stop there. Gonzalez mentioned
• 17:41
  running into her in Nairobi years later, and
• 17:44
  Martha had become a leader in an online hospitality
• 17:47
  company. All right. The training data job wasn't
• 17:50
  just a temporary gig to pay rent. It was a launch
• 17:53
  pad into a permanent professional tech career.
• 17:56
  When you aggregate stories like Martha's and
• 17:58
  Ken's across over a decade of operations, the
• 18:02
  overall impact is staggering. Let's hear the
• 18:04
  numbers. To date, this model has moved over 50
• 18:07
  ,000 people out of extreme poverty. 50 ,000?
• 18:10
  Yeah. That metric includes over 11 ,000 direct
• 18:14
  workers, plus all of their immediate dependents
• 18:17
  whose lives are stabilized by that income. That
• 18:19
  is just incredible. And perhaps the most impressive
• 18:22
  metric is their internal tracking. It shows that
• 18:25
  three years after leaving Samosaurs, former workers
• 18:28
  are earning on average, four times their previous
• 18:32
  earnings. Four times. So what does this all mean?
• 18:35
  When you synthesize Martha's launchpad into hospitality,
• 18:38
  Ken's escape from the informal economy, and that
• 18:41
  cutting edge delivery center in Kampala, it forces
• 18:44
  you to completely rethink what a tech hub actually
• 18:47
  looks like. It really does. We are so conditioned
• 18:50
  to think of tech hubs as shiny, sprawling campuses
• 18:53
  in Silicon Valley with free cafeterias. And ping
• 18:56
  pong tables. Exactly. But this deep dive suggests
• 18:59
  that a secure facility in East Africa, filled
• 19:01
  with brilliant people who were previously ignored
• 19:03
  by the global economy, is just as vital to the
• 19:06
  future of technology as any engineering office
• 19:08
  in California. This raises an important cushion
• 19:11
  for the broader tech industry, especially as
• 19:13
  we enter an era dominated by artificial intelligence.
• 19:16
  Oh, absolutely. Logical leadership cannot just
• 19:19
  be about writing the fastest algorithm or securing
• 19:21
  the largest venture capital check. It has to
• 19:24
  be about leading through action and operational
• 19:26
  ethics, exactly as Leila Gianna envisioned. Because
• 19:29
  the proof is right there. Right. If samisores
• 19:32
  can provide enterprise -grade, flawlessly accurate
• 19:36
  training data to a quarter of the Fortune 50
• 19:38
  companies by using a vertically integrated, deeply
• 19:42
  ethical supply chain, why can't the rest of the
• 19:46
  industry? That is the million dollar question.
• 19:49
  It proves definitively that exploiting gig labor
• 19:52
  is a business choice, not a technological necessity.
• 19:56
  It really is a choice. And that is what I want
• 19:58
  you, the listener, to take away from this exploration
• 20:01
  today. Think about your own daily routines. Yeah,
• 20:04
  the things we take for granted. Right. The next
• 20:06
  time you use a major search engine to find a
• 20:08
  product, or you see a highly targeted recommendation
• 20:11
  pop up on your feed, or maybe you find yourself
• 20:13
  riding in a smart, semi -autonomous car. Remember
• 20:17
  that the machine didn't just figure it out on
• 20:19
  its own. Exactly. There is a very high probability
• 20:23
  that a human being, perhaps sitting at a monitor
• 20:26
  in Kenya or Uganda, carefully and deliberately
• 20:29
  taught that specific machine how to serve you.
• 20:32
  It's humbling, really. And in the process of
• 20:34
  teaching that machine, they fundamentally changed
• 20:37
  the economic trajectory of their own family.
• 20:39
  It creates a really profound, invisible connection
• 20:42
  between the everyday end user and the unseen
• 20:46
  creator across the globe. It does. And it leaves
• 20:49
  me with one final thought, something to really
• 20:51
  chew on that goes even a step further than the
• 20:53
  presentation. We've established that human beings
• 20:56
  are the ones acting as the teachers for these
• 20:58
  incredibly powerful AI systems. Right. Providing
• 21:00
  the ground truth. Yes, they are the ones defining
• 21:03
  the ground truth, teaching the AI how to see
• 21:06
  and understand our chaotic world. The human beings
• 21:08
  aren't perfectly neutral robots. We all have
• 21:11
  context. We all have a lens through which we
• 21:13
  view the world. We do. So how might the unique
• 21:16
  cultural perspectives, the diverse environments,
• 21:18
  and the lived experiences of these data experts
• 21:20
  in East Africa be subtly shaping the global worldview
• 21:24
  of the AI systems we all rely on? Oh, that's
• 21:27
  a fascinating point. If the human teacher always
• 21:29
  leaves a subtle fingerprint on the digital student,
• 21:33
  What does it mean that the foundational worldview
• 21:36
  of tomorrow's AI is being shaped by people who
• 21:39
  understand resilience, community, and resourcefulness
• 21:41
  in a way that most of Silicon Valley never could?
• 21:44
  It's a completely different paradigm. It definitely
• 21:46
  makes you look at that self -driving car at the
• 21:49
  intersection a whole lot differently.