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

WEBVTT
00:00:00.000 --> 00:00:02.259
I want you to imagine, just for a second, that
00:00:02.259 --> 00:00:04.519
you are sitting in the passenger seat of a brand
00:00:04.519 --> 00:00:07.040
new self -driving car. OK, I'm picturing it.
00:00:07.259 --> 00:00:10.919
Right. And you are cruising up to a busy, completely
00:00:10.919 --> 00:00:15.720
chaotic city intersection. You've got... Pedestrians
00:00:15.720 --> 00:00:17.820
stepping off the curb while looking at their
00:00:17.820 --> 00:00:21.120
phones. You've got a person on a bicycle weaving
00:00:21.120 --> 00:00:23.760
aggressively through traffic, and the sun is
00:00:23.760 --> 00:00:26.960
setting, casting these long, weird shadows from
00:00:26.960 --> 00:00:29.239
the buildings above. Basically a nightmare scenario
00:00:29.239 --> 00:00:33.280
for any driver. Exactly. Now, the car stops perfectly.
00:00:33.799 --> 00:00:36.659
It waits for the bike, ignores those confusing
00:00:36.659 --> 00:00:39.299
shadows, and then smoothly makes the turn. Which
00:00:39.299 --> 00:00:41.899
is... Pretty incredible. It is. Most of us see
00:00:41.899 --> 00:00:44.020
that and we just think, wow, technology is basically
00:00:44.020 --> 00:00:46.679
magic now. But how does that car actually know
00:00:46.679 --> 00:00:49.579
the difference between a real person, a bicycle
00:00:49.579 --> 00:00:52.659
and like a dark patch of asphalt? Right. It's
00:00:52.659 --> 00:00:54.740
definitely not magic. Not at all. Yeah. It's
00:00:54.740 --> 00:00:56.939
actually the result of a massive, incredibly
00:00:56.939 --> 00:01:00.340
tedious and entirely hidden human effort. I'm
00:01:00.340 --> 00:01:01.899
your host, by the way, and I'm thrilled to get
00:01:01.899 --> 00:01:04.680
into this. And I'm your resident expert for today.
00:01:05.480 --> 00:01:07.959
We definitely have this illusion that artificial
00:01:07.959 --> 00:01:12.260
intelligence just spontaneously understands the
00:01:12.260 --> 00:01:14.280
physical world. Like it just wakes up one day.
00:01:14.739 --> 00:01:16.920
Yeah, exactly. Like it wakes up on a server farm
00:01:16.920 --> 00:01:19.700
somewhere, downloads the internet, and suddenly
00:01:19.700 --> 00:01:21.780
knows what a stop sign looks like in a snowstorm.
00:01:22.019 --> 00:01:24.659
It's just so funny to think about. But the reality
00:01:24.659 --> 00:01:28.459
is far more grounded. It's built on a foundation
00:01:28.459 --> 00:01:32.040
of human labor that is much more intensive than
00:01:32.040 --> 00:01:35.680
most people ever realize. And that hidden foundation
00:01:35.680 --> 00:01:38.340
is exactly what we are diving into today. This
00:01:38.340 --> 00:01:41.359
is our deep dive into the invisible human workforce
00:01:41.359 --> 00:01:44.079
powering global artificial intelligence. It's
00:01:44.079 --> 00:01:46.799
a huge topic. It really is. Specifically, we're
00:01:46.799 --> 00:01:49.079
looking at the world of ethical AI training data.
00:01:49.519 --> 00:01:51.299
Our source material for this is a fascinating
00:01:51.299 --> 00:01:53.239
presentation delivered by Wendy Gonzalez. Oh,
00:01:53.239 --> 00:01:55.700
she's brilliant. She is. She's the acting CEO
00:01:55.700 --> 00:01:58.280
and president of an organization called Samisource.
00:01:58.379 --> 00:02:00.760
Right, sometimes referred to as SOO in the materials.
00:02:01.219 --> 00:02:04.840
Yes, exactly. And she gave this talk at a Women
00:02:04.840 --> 00:02:07.760
in Tech Network conference. Her mission in this
00:02:07.760 --> 00:02:10.560
presentation really struck me. We are going to
00:02:10.560 --> 00:02:13.539
explore how the literal act of teaching an AI
00:02:13.539 --> 00:02:17.099
to see and think is, simultaneously, being used
00:02:17.099 --> 00:02:20.060
to lift tens of thousands of people out of extreme
00:02:20.060 --> 00:02:22.419
poverty. This is a really powerful story. It
00:02:22.419 --> 00:02:24.919
is. And along the way, we'll see why this model
00:02:24.919 --> 00:02:27.639
completely upends what we assume about the tech
00:02:27.639 --> 00:02:30.659
industry's supply chain. Okay, let's unpack this
00:02:30.659 --> 00:02:34.020
yet to truly grasp the sheer scale of what Wendy
00:02:34.020 --> 00:02:36.759
Gonzalez is talking about We have to strip AI
00:02:36.759 --> 00:02:40.039
down to its its fundamental building blocks,
00:02:40.039 --> 00:02:42.639
right? The absolute basics exactly before we
00:02:42.639 --> 00:02:44.419
even talk about the people sitting at the computers
00:02:44.419 --> 00:02:47.159
doing this work We have to look at why this specific
00:02:47.159 --> 00:02:50.340
type of labor is so Unbelievably valuable to
00:02:50.340 --> 00:02:52.259
the largest companies on the planet like Google
00:02:52.259 --> 00:02:54.479
and Walmart, right? We're talking about training
00:02:54.479 --> 00:02:57.020
data, because you literally can't have artificial
00:02:57.020 --> 00:02:58.939
intelligence without it. Right. Before a machine
00:02:58.939 --> 00:03:01.580
can see, hear, or speak, it has to be taught
00:03:01.580 --> 00:03:04.340
what it's looking at. Yep. And it is taught using
00:03:04.340 --> 00:03:07.159
labeled data. So going back to our self -driving
00:03:07.159 --> 00:03:09.599
car at the intersection, it doesn't intuitively
00:03:09.599 --> 00:03:11.840
know what a pedestrian looks like. Not at all.
00:03:12.159 --> 00:03:15.199
Someone, an actual human being, had to sit at
00:03:15.199 --> 00:03:18.020
a screen. look at an image from a street camera,
00:03:18.319 --> 00:03:21.900
draw a digital box around a person, and tag it
00:03:21.900 --> 00:03:23.840
like this is a pedestrian. And they had to do
00:03:23.840 --> 00:03:26.379
that millions and millions of times. From every
00:03:26.379 --> 00:03:29.240
conceivable angle. Yes. What's fascinating here
00:03:29.240 --> 00:03:31.500
is that this manual repetition isn't just a small
00:03:31.500 --> 00:03:33.639
part of the process. It is the single biggest
00:03:33.639 --> 00:03:36.740
bottleneck in the entire tech industry. Wow.
00:03:36.979 --> 00:03:39.680
Really? Yeah. Gonzalez points out in her presentation
00:03:39.680 --> 00:03:43.800
that by 2023, nearly $100 billion was projected
00:03:43.800 --> 00:03:46.020
to be invested in artificial intelligence. A
00:03:46.020 --> 00:03:49.240
hundred billion? That's insane. It's massive.
00:03:49.719 --> 00:03:51.979
But here is the massive catch that most people
00:03:51.979 --> 00:03:55.300
outside of Silicon Valley completely miss. Eighty
00:03:55.300 --> 00:03:58.259
percent of the time consumed in most AI and machine
00:03:58.259 --> 00:04:00.780
learning projects is spent just dealing with
00:04:00.780 --> 00:04:03.330
this training data. You mean preparing it, cleaning
00:04:03.330 --> 00:04:06.210
it, labeling it? Exactly. 80%. That statistic
00:04:06.210 --> 00:04:10.150
is just wild to me. 80 % of the time. It makes
00:04:10.150 --> 00:04:13.189
me think of training data as basically flashcards
00:04:13.189 --> 00:04:16.459
for a billion dollar toddler. Oh, yes. That's
00:04:16.459 --> 00:04:18.839
exactly what it is. I mean, you have this incredibly
00:04:18.839 --> 00:04:22.480
powerful, super expensive AI system, but it's
00:04:22.480 --> 00:04:25.319
fundamentally naive. It knows absolutely nothing
00:04:25.319 --> 00:04:27.839
about the physical world until a human sits down
00:04:27.839 --> 00:04:30.060
with it and holds up a million flashcards saying,
00:04:30.060 --> 00:04:33.060
you know, car, tree, road sign. It really is
00:04:33.060 --> 00:04:35.060
shocking when you put it like that. It is. It's
00:04:35.060 --> 00:04:37.000
honestly shocking that the most cutting edge
00:04:37.000 --> 00:04:40.779
technology in human history is entirely dependent
00:04:40.779 --> 00:04:44.199
on manual human repetition. A billion -dollar
00:04:44.199 --> 00:04:46.339
toggler is the perfect way to look at it because
00:04:46.339 --> 00:04:49.800
the algorithms themselves, the brilliant complex
00:04:49.800 --> 00:04:52.660
math that the engineers write, they are functionally
00:04:52.660 --> 00:04:55.019
useless without the ground truth provided by
00:04:55.019 --> 00:04:57.480
that human labeling. Ah, ground truth. Yeah.
00:04:57.759 --> 00:05:00.040
If the data is bad, if the human incorrectly
00:05:00.040 --> 00:05:02.839
labels a pedestrian as a shadow, that $100 billion
00:05:02.839 --> 00:05:06.120
industry grinds to an absolute halt. The algorithm
00:05:06.120 --> 00:05:08.660
learns the wrong lesson. Exactly. And then the
00:05:08.660 --> 00:05:11.100
self -driving car makes a fatal error. Man, it's
00:05:11.100 --> 00:05:14.480
wild to think that these hundred billion dollar
00:05:14.480 --> 00:05:17.779
algorithms are entirely at the mercy of whoever's
00:05:17.779 --> 00:05:19.819
clicking the mouse to draw those boxes. It really
00:05:19.819 --> 00:05:22.079
puts it into perspective. It does, which makes
00:05:22.079 --> 00:05:24.399
me wonder how these massive companies actually
00:05:24.399 --> 00:05:27.759
source that labor. If the tech world has this
00:05:27.759 --> 00:05:31.480
bottomless, desperate hunger for flawlessly labeled
00:05:31.480 --> 00:05:34.319
data, who is getting paid to sit there and do
00:05:34.319 --> 00:05:36.439
it? Well, traditionally, the tech industry has
00:05:36.439 --> 00:05:39.319
treated data labeling as low level gig work.
00:05:39.600 --> 00:05:42.279
Outsourcing. Yeah, exactly. It's outsourced to
00:05:42.279 --> 00:05:44.980
massive click farms where workers are paid like
00:05:44.980 --> 00:05:48.120
pennies per task with zero job security, right?
00:05:48.319 --> 00:05:51.360
It's treated as highly expendable temporary labor.
00:05:52.139 --> 00:05:54.660
But that brings us directly to Sama sources radical
00:05:54.660 --> 00:05:57.259
philosophy, which completely flips that traditional
00:05:57.259 --> 00:05:59.540
model on its head. I love this part. Yeah, to
00:05:59.540 --> 00:06:01.360
understand their approach, you really have to
00:06:01.360 --> 00:06:04.110
look at their late founder, Leila Janna. She
00:06:04.110 --> 00:06:07.170
founded the organization in 2008 with one very
00:06:07.170 --> 00:06:10.449
specific driving ethos. Talent is distributed
00:06:10.449 --> 00:06:12.949
equally while opportunity is not. Yes, exactly
00:06:12.949 --> 00:06:15.730
that. That is such a fascinating lens to view
00:06:15.730 --> 00:06:18.509
the global economy through, because it totally
00:06:18.509 --> 00:06:21.209
rejects the idea that a lack of wealth equals
00:06:21.209 --> 00:06:24.250
a lack of capability. It does. Jaina believed
00:06:24.250 --> 00:06:26.629
that traditional charity wasn't the answer to
00:06:26.629 --> 00:06:29.949
systemic poverty. Her goal was giving work instead
00:06:29.949 --> 00:06:32.370
of giving aid. Which is a huge paradigm shift.
00:06:32.569 --> 00:06:35.310
Right. So Samasource's demographic target for
00:06:35.310 --> 00:06:38.910
hiring is highly specific and very unusual for
00:06:38.910 --> 00:06:41.769
a tech company. They intentionally recruit people
00:06:41.769 --> 00:06:44.360
who are living below the World Bank average poverty
00:06:44.360 --> 00:06:47.579
line. Wow. And crucially, they require absolutely
00:06:47.579 --> 00:06:51.319
no prior education, no college degree, and no
00:06:51.319 --> 00:06:53.800
formal job experience. Which brings us to that
00:06:53.800 --> 00:06:55.959
incredibly vivid example from the presentation,
00:06:56.279 --> 00:06:59.279
the story of a man named Ken Kihara. Oh, Ken's
00:06:59.279 --> 00:07:01.259
story is amazing. It really is. So Ken lived
00:07:01.259 --> 00:07:03.300
in Mathar, which is one of the largest, most
00:07:03.300 --> 00:07:06.439
densely populated slums in Kenya. And Gonzalez
00:07:06.439 --> 00:07:08.500
notes that in Mathar, unemployment for youth
00:07:08.500 --> 00:07:11.100
and women hovers around 70%. Or even higher,
00:07:11.180 --> 00:07:13.540
yeah. Right. Now Ken had amazing grades coming
00:07:13.540 --> 00:07:16.100
out of high school. He clearly had the raw talent.
00:07:16.199 --> 00:07:18.800
But no opportunity. Exactly. He had absolutely
00:07:18.800 --> 00:07:21.740
no network, no access to capital, and no way
00:07:21.740 --> 00:07:24.230
to pay for a college degree. He was entirely
00:07:24.230 --> 00:07:26.389
locked out of the formal economy. And you know,
00:07:26.470 --> 00:07:28.829
when someone with that much potential is locked
00:07:28.829 --> 00:07:31.529
out of the formal economy, they are pushed into
00:07:31.529 --> 00:07:34.370
incredibly dangerous survival situations. Just
00:07:34.370 --> 00:07:37.509
to get by. Yes. Because he couldn't find safe
00:07:37.509 --> 00:07:40.449
formal work, Ken was forced into the informal
00:07:40.449 --> 00:07:43.910
economy just to feed himself. The presentation
00:07:43.910 --> 00:07:45.889
details how he became a producer of Chancho.
00:07:45.930 --> 00:07:48.500
Right, which is a local moonshine. But Gonzales
00:07:48.500 --> 00:07:51.160
specifies this gritty, terrifying detail about
00:07:51.160 --> 00:07:53.879
the environment Ken was working in. The jet fuel
00:07:53.879 --> 00:07:56.620
detail. Yes. She points out a photo during her
00:07:56.620 --> 00:07:59.199
presentation explaining that the specific batch
00:07:59.199 --> 00:08:02.060
of Chandra Ken was making was literally brewed
00:08:02.060 --> 00:08:04.079
with jet fuel. It's hard to even comprehend.
00:08:04.120 --> 00:08:06.279
Just let that sink in for a second, right? A
00:08:06.279 --> 00:08:08.399
young guy with top high school grades is pushed
00:08:08.399 --> 00:08:11.379
into brewing highly toxic combustible moonshine
00:08:11.379 --> 00:08:14.199
in a slum entirely because he was born in a neighborhood
00:08:14.199 --> 00:08:17.180
with a 70 % unemployment rate. Yeah. That is
00:08:17.180 --> 00:08:19.439
exactly where Samasour stepped in, offering him
00:08:19.439 --> 00:08:22.680
training and his first formal job in data annotation.
00:08:23.540 --> 00:08:25.660
And it completely changes his trajectory. He
00:08:25.660 --> 00:08:29.319
goes from mixing toxic chemicals to survive to
00:08:29.319 --> 00:08:31.860
training the algorithms of Fortune 50 companies.
00:08:32.139 --> 00:08:33.879
OK, but I want to look at this from Google or
00:08:33.879 --> 00:08:35.940
Walmart's perspective for a second. Sure. Wait.
00:08:36.080 --> 00:08:39.399
If companies like Walmart and Google need flawlessly
00:08:39.399 --> 00:08:42.360
labeled data for critical systems, isn't it a
00:08:42.360 --> 00:08:45.279
massive risk to hire people with absolutely zero
00:08:45.279 --> 00:08:48.240
tech background or college degrees? Sounds risky,
00:08:48.419 --> 00:08:51.000
right? I mean, you are building a multi -billion
00:08:51.000 --> 00:08:54.519
dollar self -driving car system or a massive
00:08:54.519 --> 00:08:57.750
global search engine. Are you really going to
00:08:57.750 --> 00:09:00.269
entrust the foundational safety data, the ground
00:09:00.269 --> 00:09:02.590
truth of your algorithm, to someone whose last
00:09:02.590 --> 00:09:05.269
job was brewing jet fuel moonshine in a slum?
00:09:05.610 --> 00:09:08.149
That is the exact skepticism traditional business
00:09:08.149 --> 00:09:10.009
leaders throw at this model. I don't guarantee
00:09:10.009 --> 00:09:12.570
quality. Right, but that skepticism fundamentally
00:09:12.570 --> 00:09:14.990
misunderstands the genius of what Samisource
00:09:14.990 --> 00:09:17.129
built. Let's go back to how the traditional tech
00:09:17.129 --> 00:09:19.629
industry handles this. gig work. Right. The click
00:09:19.629 --> 00:09:22.429
farms. Exactly. When you treat data labelers
00:09:22.429 --> 00:09:25.370
as disposable temporary gig workers, they have
00:09:25.370 --> 00:09:27.169
absolutely no incentive to care about the quality
00:09:27.169 --> 00:09:29.149
of the data. They just want to get paid. Yeah.
00:09:29.250 --> 00:09:30.970
They just want to click as fast as possible to
00:09:30.970 --> 00:09:33.649
make their pennies. The result is incredibly
00:09:33.649 --> 00:09:37.309
high turnover, burnout, and terrible quality
00:09:37.309 --> 00:09:39.970
control. Right. If you treat someone like a disposable
00:09:39.970 --> 00:09:42.590
cog, they are going to give you disposable work.
00:09:42.909 --> 00:09:45.929
Exactly. Samisource recognize that to get premium
00:09:45.929 --> 00:09:49.509
data, you actually need stability. So they provide
00:09:49.509 --> 00:09:52.610
full -time employment, not gig contracts. It's
00:09:52.610 --> 00:09:55.870
just huge. And they provide living wages that
00:09:55.870 --> 00:09:58.309
are intentionally calculated to support a family
00:09:58.309 --> 00:10:01.490
of four. They provide real health benefits. So
00:10:01.490 --> 00:10:04.710
it's a real career. It is. By taking someone
00:10:04.710 --> 00:10:07.330
like Ken, who has previously only known economic
00:10:07.330 --> 00:10:10.519
chaos. and placing him in an environment of profound
00:10:10.519 --> 00:10:13.000
professional stability, they breed incredible
00:10:13.000 --> 00:10:15.299
loyalty and focus. That makes total sense. Right.
00:10:15.460 --> 00:10:17.919
This social mission actually solves the tech
00:10:17.919 --> 00:10:20.029
world's biggest quality problem. Here's where
00:10:20.029 --> 00:10:23.029
it gets really interesting. Samasaurs isn't just
00:10:23.029 --> 00:10:25.230
a charity project that these tech giants tolerate
00:10:25.230 --> 00:10:28.450
for some good PR. Not at all. Their social impact
00:10:28.450 --> 00:10:30.970
mission is precisely what gives them their competitive
00:10:30.970 --> 00:10:33.029
advantage in the marketplace. By treating people
00:10:33.029 --> 00:10:35.490
as long -term investments rather than short -term
00:10:35.490 --> 00:10:38.149
expenses, they end up with better, cleaner data
00:10:38.149 --> 00:10:40.730
than the traditional gig economy tech companies.
00:10:41.070 --> 00:10:43.470
And the retention numbers completely validate
00:10:43.470 --> 00:10:46.399
that approach. Gonzalez notes that Samisource
00:10:46.399 --> 00:10:49.559
boasts industry -leading retention rates. They
00:10:49.559 --> 00:10:52.580
have less than 10 % annualized attrition. Wait,
00:10:52.580 --> 00:10:55.580
less than 10 %? Yes. In the tech support and
00:10:55.580 --> 00:10:58.840
data labeling world, a sub -10 % turnover rate
00:10:58.840 --> 00:11:02.019
is practically unheard of. I believe it. Because
00:11:02.019 --> 00:11:04.659
these jobs are so deeply meaningful to the workers,
00:11:05.139 --> 00:11:07.799
they stay. And because they stay year after year,
00:11:08.139 --> 00:11:11.200
they transform from absolute beginners into deeply
00:11:11.200 --> 00:11:14.399
specialized, world -class labeling experts. And
00:11:14.399 --> 00:11:17.000
the stakes of that acquired expertise are literally
00:11:17.000 --> 00:11:18.700
life and death. Going back to our self -driving
00:11:18.700 --> 00:11:21.500
car example, Gazalos points out that an AI must
00:11:21.500 --> 00:11:23.820
be trained to perfectly distinguish between a
00:11:23.820 --> 00:11:26.759
motorcycle and a bicycle. Or it must be able
00:11:26.759 --> 00:11:29.899
to instantly read specific vital road signs while
00:11:29.899 --> 00:11:32.059
ignoring irrelevant billboards. What they call
00:11:32.059 --> 00:11:35.220
the last mile of quality. Yes, the last mile
00:11:35.220 --> 00:11:37.960
of quality. You cannot get that from a burned
00:11:37.960 --> 00:11:40.379
out gig worker who started yesterday and is rushing
00:11:40.379 --> 00:11:43.899
to make minimum wage. You need a dedicated expert.
00:11:44.159 --> 00:11:46.179
And it goes beyond autonomous vehicles, too.
00:11:46.480 --> 00:11:48.419
Think about the Walmart example she provided.
00:11:48.639 --> 00:11:51.200
OK. Walmart has the world's largest catalog of
00:11:51.200 --> 00:11:53.580
products, and that catalog is shifting by the
00:11:53.580 --> 00:11:57.019
minute. Oh, I'm sure. Millions of items. Millions.
00:11:57.080 --> 00:11:59.879
Yes. Items being added, removed, changing seasons.
00:12:00.340 --> 00:12:03.440
If Walmart introduces a new seasonal aisle, the
00:12:03.440 --> 00:12:07.059
AI might miscategorize a plastic Halloween pumpkin
00:12:07.059 --> 00:12:11.710
basket as fresh produce. That would be a mess.
00:12:11.950 --> 00:12:14.110
Right. If their product search algorithms aren't
00:12:14.110 --> 00:12:16.570
constantly being audited and retrained by human
00:12:16.570 --> 00:12:19.309
experts who understand nuance, the search function
00:12:19.309 --> 00:12:21.309
breaks down entirely. And consumers wouldn't
00:12:21.309 --> 00:12:23.230
be able to find what they're looking for. Which
00:12:23.230 --> 00:12:26.190
means lost revenue. It requires a highly accurate,
00:12:26.190 --> 00:12:29.029
ongoing, human -in -the -loop system. Which brings
00:12:29.029 --> 00:12:31.049
up the reality of how this work is actually done
00:12:31.049 --> 00:12:33.769
on the ground. During the presentation, she showed
00:12:33.769 --> 00:12:35.990
a picture of their delivery center in Kampala,
00:12:36.070 --> 00:12:38.740
Uganda. Oh, that facility is stunning. It really
00:12:38.740 --> 00:12:40.899
is, and I think this completely shatters the
00:12:40.899 --> 00:12:42.759
preconceived notions a lot of people might have.
00:12:43.320 --> 00:12:46.299
This is not some dark, crowded sweatshop. No,
00:12:46.399 --> 00:12:48.840
definitely not. You walk off a busy street in
00:12:48.840 --> 00:12:51.519
Kampala and enter a highly secure center with
00:12:51.519 --> 00:12:54.960
biometric access, top tier technology and facilities
00:12:54.960 --> 00:12:58.480
built with B Corp certified sustainability. It
00:12:58.480 --> 00:13:01.360
looks like it belongs in Cupertino. If we connect
00:13:01.360 --> 00:13:03.879
this to the bigger picture, what we are seeing
00:13:03.879 --> 00:13:06.240
is that investing heavily in human infrastructure
00:13:06.240 --> 00:13:10.509
is the prerequisite. for superior machine infrastructure.
00:13:10.809 --> 00:13:12.889
That's a great way to put it. Yeah. And Gonzalez
00:13:12.889 --> 00:13:15.190
talks about this concept of machine learning
00:13:15.190 --> 00:13:17.419
assisted annotation. I'm glad you brought that
00:13:17.419 --> 00:13:19.419
up, actually, because the phrase machine learning
00:13:19.419 --> 00:13:21.899
assisted annotation sounds like deep corporate
00:13:21.899 --> 00:13:24.419
jargon. How does that actually work in practice?
00:13:24.779 --> 00:13:27.000
It's a fascinating feedback loop, really. The
00:13:27.000 --> 00:13:29.440
basic idea is that as the software platform gets
00:13:29.440 --> 00:13:32.620
somewhat smart, it takes over the easy, repetitive
00:13:32.620 --> 00:13:34.779
parts of the image. Like identifying the sky
00:13:34.779 --> 00:13:37.860
or the road. Exactly. And this frees up the human
00:13:37.860 --> 00:13:40.899
experts to handle the highly nuanced, uniquely
00:13:40.899 --> 00:13:43.360
human edge cases. Give me an example of an edge
00:13:43.360 --> 00:13:46.190
case. OK, imagine a pedestrian. walking down
00:13:46.190 --> 00:13:48.730
the street carrying a large, highly reflective
00:13:48.730 --> 00:13:52.070
mirror. An early AI system might look at that
00:13:52.070 --> 00:13:54.049
reflection of the street and think, there are
00:13:54.049 --> 00:13:57.490
two pedestrians here. Or worse, there is a hole
00:13:57.490 --> 00:14:00.210
in reality. Because it doesn't understand context.
00:14:00.669 --> 00:14:04.409
It lacks human context. The human expert in Kampala
00:14:04.409 --> 00:14:06.929
spots that error. draws the correct boundary
00:14:06.929 --> 00:14:09.370
around the pedestrian in the mirror, and tags
00:14:09.370 --> 00:14:11.950
it properly. So they aren't just fixing a single
00:14:11.950 --> 00:14:14.850
photo. No, they are actually codifying their
00:14:14.850 --> 00:14:17.090
human knowledge into the system. When the expert
00:14:17.090 --> 00:14:20.110
corrects that mirror reflection, that feedback
00:14:20.110 --> 00:14:23.129
alters the neural network's weights. Wow. The
00:14:23.129 --> 00:14:25.710
platform gets smarter over time, specifically
00:14:25.710 --> 00:14:28.289
because the humans training it are treated as
00:14:28.289 --> 00:14:30.750
respected experts capable of nuanced judgment
00:14:30.750 --> 00:14:33.690
rather than just clicking machines. So the technology
00:14:33.690 --> 00:14:36.309
elevates the human's capabilities and the human's
00:14:36.309 --> 00:14:38.990
nuanced judgment elevates the technology. Exactly.
00:14:39.090 --> 00:14:41.289
It's a brilliant synergy. And when you look at
00:14:41.289 --> 00:14:43.250
the leadership team driving this, it becomes
00:14:43.250 --> 00:14:45.950
clear that this philosophy of valuing overlooked
00:14:45.950 --> 00:14:48.850
talent isn't just a happy accident. No, it's
00:14:48.850 --> 00:14:51.470
very intentional. It is built into the structural
00:14:51.470 --> 00:14:54.470
DNA of the entire organization. Diversity and
00:14:54.470 --> 00:14:58.190
inclusion are so often used as like hollow buzzwords
00:14:58.190 --> 00:15:00.429
in the tech industry, you know, usually just
00:15:00.429 --> 00:15:02.570
a slide in an HR presentation. Right, checking
00:15:02.570 --> 00:15:05.389
a box. But at Samisource, it is a structural
00:15:05.389 --> 00:15:08.409
mandate. They purposefully mandate that 50 %
00:15:08.409 --> 00:15:11.470
of their hires are women. Which is amazing. And
00:15:11.470 --> 00:15:14.090
Gonzalez was very explicit in her talk that this
00:15:14.090 --> 00:15:16.690
isn't just about padding the stats at the entry
00:15:16.690 --> 00:15:19.210
level data labeling positions. Oh, really? Yeah,
00:15:19.309 --> 00:15:22.850
this 50 % mandate extends upward through middle
00:15:22.850 --> 00:15:25.070
management, executive leadership, and all the
00:15:25.070 --> 00:15:27.200
way to the board of directors. That structural
00:15:27.200 --> 00:15:29.059
commitment was really put to the test during
00:15:29.059 --> 00:15:32.179
their 2019 Series A funding round. Yes, that
00:15:32.179 --> 00:15:34.700
was a pivotal moment. Gonzales shared a story
00:15:34.700 --> 00:15:36.700
about how she and Leila John were out raising
00:15:36.700 --> 00:15:40.220
$14 .8 million. And she made a really poignant
00:15:40.220 --> 00:15:42.299
observation about what it's like to be in those
00:15:42.299 --> 00:15:44.639
high stakes venture capital meetings. Oh, I remember
00:15:44.639 --> 00:15:47.480
this part. Yeah. As two minority women executives
00:15:47.480 --> 00:15:50.080
seeking multimillion dollar funding for a software
00:15:50.080 --> 00:15:52.639
company, they were very often the only women
00:15:52.639 --> 00:15:55.240
in the room. That can be an incredibly daunting
00:15:55.240 --> 00:15:57.879
environment. Especially when pitching a business
00:15:57.879 --> 00:16:01.000
model that, quite frankly, defies conventional
00:16:01.000 --> 00:16:03.559
Silicon Valley logic. Right, pitching social
00:16:03.559 --> 00:16:06.769
impact to traditional VCs. Exactly. Yet they
00:16:06.769 --> 00:16:09.250
found investors who didn't just passively tolerate
00:16:09.250 --> 00:16:12.269
their diversity or their social mission, but
00:16:12.269 --> 00:16:14.769
actively valued it as a business strategy. Which
00:16:14.769 --> 00:16:18.590
is so rare. It is. They found financial backers
00:16:18.590 --> 00:16:21.610
who understood their technical AI expertise,
00:16:21.610 --> 00:16:24.090
but also recognized that their diverse workforce
00:16:24.090 --> 00:16:27.570
and deeply ethical supply chain were massive
00:16:27.570 --> 00:16:30.090
strategic assets. It's impossible to talk about
00:16:30.090 --> 00:16:32.450
that grand vision without acknowledging the deeply
00:16:32.450 --> 00:16:35.090
tragic part of this story. Yeah, Leila's passing.
00:16:35.289 --> 00:16:38.330
Leila Janna, the founder who pioneered this entire
00:16:38.330 --> 00:16:41.029
concept of using a tech business as a direct
00:16:41.029 --> 00:16:43.909
force for social change, passed away from a rare
00:16:43.909 --> 00:16:47.169
aggressive cancer in January 2020. It was a huge
00:16:47.169 --> 00:16:50.070
loss. It really was. But Gonzalez makes it very
00:16:50.070 --> 00:16:53.210
clear that her legacy, her fundamental DNA, remains
00:16:53.210 --> 00:16:55.570
entirely embedded in how the company operates
00:16:55.570 --> 00:16:58.710
today. You see Leila's legacy most clearly when
00:16:58.710 --> 00:17:00.710
you look at the individual lives that have been
00:17:00.710 --> 00:17:03.509
transformed by this model. The presentation highlighted
00:17:03.509 --> 00:17:06.089
another worker named Martha. whose story is just
00:17:06.089 --> 00:17:08.549
as compelling as Cannes. Right, Martha. Martha
00:17:08.549 --> 00:17:10.869
is a young woman who was completely locked out
00:17:10.869 --> 00:17:13.730
of the workforce. She had the drive, but she
00:17:13.730 --> 00:17:16.089
couldn't get a job simply because of her gender
00:17:16.089 --> 00:17:17.970
and because she lived in a marginalized slum
00:17:17.970 --> 00:17:20.750
area. She faced compounding barriers. But then
00:17:20.750 --> 00:17:24.240
she was recruited by Samisource. And the trajectory
00:17:24.240 --> 00:17:27.940
of her life didn't just change slowly, it changed
00:17:27.940 --> 00:17:30.799
dramatically. It did. After joining and earning
00:17:30.799 --> 00:17:34.259
a stable living wage, she was able to move entirely
00:17:34.259 --> 00:17:36.519
out of the slums and lease a formal apartment.
00:17:36.799 --> 00:17:38.920
Which is life -changing on its own. Completely.
00:17:39.279 --> 00:17:41.700
But the story doesn't stop there. Gonzalez mentioned
00:17:41.700 --> 00:17:44.480
running into her in Nairobi years later, and
00:17:44.480 --> 00:17:47.460
Martha had become a leader in an online hospitality
00:17:47.460 --> 00:17:50.400
company. All right. The training data job wasn't
00:17:50.400 --> 00:17:53.500
just a temporary gig to pay rent. It was a launch
00:17:53.500 --> 00:17:56.319
pad into a permanent professional tech career.
00:17:56.819 --> 00:17:58.859
When you aggregate stories like Martha's and
00:17:58.859 --> 00:18:02.359
Ken's across over a decade of operations, the
00:18:02.359 --> 00:18:04.380
overall impact is staggering. Let's hear the
00:18:04.380 --> 00:18:07.740
numbers. To date, this model has moved over 50
00:18:07.740 --> 00:18:10.359
,000 people out of extreme poverty. 50 ,000?
00:18:10.440 --> 00:18:14.029
Yeah. That metric includes over 11 ,000 direct
00:18:14.029 --> 00:18:17.289
workers, plus all of their immediate dependents
00:18:17.289 --> 00:18:19.829
whose lives are stabilized by that income. That
00:18:19.829 --> 00:18:22.529
is just incredible. And perhaps the most impressive
00:18:22.529 --> 00:18:25.730
metric is their internal tracking. It shows that
00:18:25.730 --> 00:18:28.789
three years after leaving Samosaurs, former workers
00:18:28.789 --> 00:18:32.190
are earning on average, four times their previous
00:18:32.190 --> 00:18:34.930
earnings. Four times. So what does this all mean?
00:18:35.390 --> 00:18:38.269
When you synthesize Martha's launchpad into hospitality,
00:18:38.869 --> 00:18:41.309
Ken's escape from the informal economy, and that
00:18:41.309 --> 00:18:44.269
cutting edge delivery center in Kampala, it forces
00:18:44.269 --> 00:18:47.569
you to completely rethink what a tech hub actually
00:18:47.569 --> 00:18:50.480
looks like. It really does. We are so conditioned
00:18:50.480 --> 00:18:53.720
to think of tech hubs as shiny, sprawling campuses
00:18:53.720 --> 00:18:56.619
in Silicon Valley with free cafeterias. And ping
00:18:56.619 --> 00:18:59.400
pong tables. Exactly. But this deep dive suggests
00:18:59.400 --> 00:19:01.559
that a secure facility in East Africa, filled
00:19:01.559 --> 00:19:03.640
with brilliant people who were previously ignored
00:19:03.640 --> 00:19:06.380
by the global economy, is just as vital to the
00:19:06.380 --> 00:19:08.819
future of technology as any engineering office
00:19:08.819 --> 00:19:11.140
in California. This raises an important cushion
00:19:11.140 --> 00:19:13.579
for the broader tech industry, especially as
00:19:13.579 --> 00:19:16.259
we enter an era dominated by artificial intelligence.
00:19:16.319 --> 00:19:19.160
Oh, absolutely. Logical leadership cannot just
00:19:19.160 --> 00:19:21.980
be about writing the fastest algorithm or securing
00:19:21.980 --> 00:19:24.660
the largest venture capital check. It has to
00:19:24.660 --> 00:19:26.859
be about leading through action and operational
00:19:26.859 --> 00:19:29.839
ethics, exactly as Leila Gianna envisioned. Because
00:19:29.839 --> 00:19:32.619
the proof is right there. Right. If samisores
00:19:32.619 --> 00:19:36.099
can provide enterprise -grade, flawlessly accurate
00:19:36.099 --> 00:19:38.940
training data to a quarter of the Fortune 50
00:19:38.940 --> 00:19:42.500
companies by using a vertically integrated, deeply
00:19:42.500 --> 00:19:46.359
ethical supply chain, why can't the rest of the
00:19:46.359 --> 00:19:48.930
industry? That is the million dollar question.
00:19:49.150 --> 00:19:52.569
It proves definitively that exploiting gig labor
00:19:52.569 --> 00:19:56.289
is a business choice, not a technological necessity.
00:19:56.450 --> 00:19:58.789
It really is a choice. And that is what I want
00:19:58.789 --> 00:20:01.490
you, the listener, to take away from this exploration
00:20:01.490 --> 00:20:04.009
today. Think about your own daily routines. Yeah,
00:20:04.029 --> 00:20:06.589
the things we take for granted. Right. The next
00:20:06.589 --> 00:20:08.769
time you use a major search engine to find a
00:20:08.769 --> 00:20:11.250
product, or you see a highly targeted recommendation
00:20:11.250 --> 00:20:13.509
pop up on your feed, or maybe you find yourself
00:20:13.509 --> 00:20:17.480
riding in a smart, semi -autonomous car. Remember
00:20:17.480 --> 00:20:19.220
that the machine didn't just figure it out on
00:20:19.220 --> 00:20:23.059
its own. Exactly. There is a very high probability
00:20:23.059 --> 00:20:26.599
that a human being, perhaps sitting at a monitor
00:20:26.599 --> 00:20:29.799
in Kenya or Uganda, carefully and deliberately
00:20:29.799 --> 00:20:32.359
taught that specific machine how to serve you.
00:20:32.819 --> 00:20:34.819
It's humbling, really. And in the process of
00:20:34.819 --> 00:20:37.339
teaching that machine, they fundamentally changed
00:20:37.339 --> 00:20:39.579
the economic trajectory of their own family.
00:20:39.839 --> 00:20:42.720
It creates a really profound, invisible connection
00:20:42.720 --> 00:20:46.559
between the everyday end user and the unseen
00:20:46.559 --> 00:20:49.200
creator across the globe. It does. And it leaves
00:20:49.200 --> 00:20:51.200
me with one final thought, something to really
00:20:51.200 --> 00:20:53.200
chew on that goes even a step further than the
00:20:53.200 --> 00:20:56.079
presentation. We've established that human beings
00:20:56.079 --> 00:20:58.380
are the ones acting as the teachers for these
00:20:58.380 --> 00:21:00.920
incredibly powerful AI systems. Right. Providing
00:21:00.920 --> 00:21:03.819
the ground truth. Yes, they are the ones defining
00:21:03.819 --> 00:21:06.019
the ground truth, teaching the AI how to see
00:21:06.019 --> 00:21:08.960
and understand our chaotic world. The human beings
00:21:08.960 --> 00:21:11.519
aren't perfectly neutral robots. We all have
00:21:11.519 --> 00:21:13.759
context. We all have a lens through which we
00:21:13.759 --> 00:21:16.140
view the world. We do. So how might the unique
00:21:16.140 --> 00:21:18.500
cultural perspectives, the diverse environments,
00:21:18.680 --> 00:21:20.960
and the lived experiences of these data experts
00:21:20.960 --> 00:21:24.259
in East Africa be subtly shaping the global worldview
00:21:24.259 --> 00:21:27.000
of the AI systems we all rely on? Oh, that's
00:21:27.000 --> 00:21:29.880
a fascinating point. If the human teacher always
00:21:29.880 --> 00:21:32.839
leaves a subtle fingerprint on the digital student,
00:21:33.380 --> 00:21:36.059
What does it mean that the foundational worldview
00:21:36.059 --> 00:21:39.099
of tomorrow's AI is being shaped by people who
00:21:39.099 --> 00:21:41.680
understand resilience, community, and resourcefulness
00:21:41.680 --> 00:21:44.740
in a way that most of Silicon Valley never could?
00:21:44.960 --> 00:21:46.900
It's a completely different paradigm. It definitely
00:21:46.900 --> 00:21:49.119
makes you look at that self -driving car at the
00:21:49.119 --> 00:21:50.460
intersection a whole lot differently.