WEBVTT

00:00:00.160 --> 00:00:03.160
The discussion about

00:00:03.440 --> 00:00:05.166
what is the 
current impact

00:00:05.220 --> 00:00:07.913
of artificial intelligence
in the modern day

00:00:07.987 --> 00:00:09.427
is very important,

00:00:10.560 --> 00:00:14.520
because it has been one 
of the probably worst

00:00:14.639 --> 00:00:17.639
innovations in technology in the last

00:00:17.700 --> 00:00:19.846
30 years, 
in my opinion.

00:00:20.400 --> 00:00:22.200
Because it has 
everything,

00:00:22.200 --> 00:00:25.920
it's all that is wrong with 
the modern, capitalist technology.

00:00:26.400 --> 00:00:30.000
It's a system that obstructs 
information, obscures it,

00:00:31.000 --> 00:00:33.040
makes communication
more difficult.

00:00:33.040 --> 00:00:35.166
It increases 
misinformation.

00:00:35.400 --> 00:00:38.360
It has incredible 
impacts

00:00:38.400 --> 00:00:41.566
on the infrastructure 
of the internet,

00:00:41.627 --> 00:00:43.940
just because of how much 
computing it uses.

00:00:44.080 --> 00:00:45.880
It has horrible,

00:00:46.840 --> 00:00:48.640
environmental impacts.

00:00:48.640 --> 00:00:51.640
All these new data centers
that have been proposed to be built.

00:00:51.873 --> 00:00:54.973
The idea of having company
owned nuclear reactors

00:00:55.013 --> 00:00:58.013
to power data 
centers is genuinely,...

00:00:58.546 --> 00:00:59.660
makes me sick.

00:01:01.253 --> 00:01:02.760
This process 
is not new.

00:01:02.840 --> 00:01:06.000
This is very similar to all 
the centralization and monopolies.

00:01:06.000 --> 00:01:08.000
that modern 
technology has,

00:01:08.067 --> 00:01:10.133
just exploded 
in a way that is,

00:01:11.320 --> 00:01:13.246
frankly, ridiculous

00:01:14.280 --> 00:01:16.160
And it's very 
mysterious.

00:01:16.160 --> 00:01:18.560
As with a lot of 
technological systems,

00:01:18.660 --> 00:01:21.640
people see it and they 
lack the knowledge

00:01:21.640 --> 00:01:24.640
or maybe don't know that 
they can even know how it works.

00:01:24.920 --> 00:01:27.640
They see this mysterious 
chat bot you talk to,

00:01:27.640 --> 00:01:30.640
you type and it replies, 
looking like a human.

00:01:31.440 --> 00:01:34.440
I can understand why, if you 
don't have the knowledge,

00:01:34.553 --> 00:01:39.720
you would just abstract it in some way
as this magical box that you can talk to.

00:01:40.320 --> 00:01:41.520
So that's where it came from.

00:01:41.626 --> 00:01:45.360
And then there's the extra layer
of the idea of labour

00:01:45.360 --> 00:01:50.240
and reproduction of labour,
of how these AI systems

00:01:50.240 --> 00:01:53.640
have a big impact on labour
and exploitation of labour.

00:02:00.680 --> 00:02:03.560
A lot of the outputs

00:02:03.560 --> 00:02:06.560
of a modern artificial 
intelligence tool

00:02:07.040 --> 00:02:09.233
is actually not,

00:02:09.480 --> 00:02:14.280
from some magical technological
advancements that we have.

00:02:15.180 --> 00:02:17.060
The technology 
has existed

00:02:17.133 --> 00:02:20.160
 for the last 
15 years or so.

00:02:20.260 --> 00:02:23.160
The modern artificial 
intelligence, at least.

00:02:23.719 --> 00:02:26.719
What it is, is just 
fancy auto-complete.

00:02:26.820 --> 00:02:29.900
So, whenever you type 
a text on your phone

00:02:29.960 --> 00:02:32.580
and you see the box that says 
what the next word would be

00:02:32.646 --> 00:02:34.560
- that is pretty 
much it!

00:02:34.633 --> 00:02:36.806
It's just fancier 
and bigger.

00:02:36.853 --> 00:02:39.846
The kind of key feature 
of modern AI,

00:02:39.920 --> 00:02:44.040
it's knowledge and 
how it is able to talk

00:02:44.093 --> 00:02:47.093
in such an uncanny 
way, like a human

00:02:47.167 --> 00:02:48.767
comes from training.

00:02:49.560 --> 00:02:52.560
What is training 
from an AI?

00:02:53.640 --> 00:02:56.240
It's not like learning 
like a human.

00:02:56.333 --> 00:03:00.320
AI is not really... there's 
nothing intelligent about it.

00:03:00.880 --> 00:03:03.960
What it does,
it takes an enormous amount of data.

00:03:04.320 --> 00:03:08.453
We're talking to the scale 
of every book ever written,

00:03:09.040 --> 00:03:10.760
the whole 
of Wikipedia,

00:03:10.760 --> 00:03:14.400
every single post message 
ever posted on any forum.

00:03:14.500 --> 00:03:17.320
since the 
internet existed,

00:03:17.393 --> 00:03:20.500
all fed to this
gigantic algorithm,

00:03:20.940 --> 00:03:22.440
that starts matching:

00:03:22.440 --> 00:03:24.440
Oh, this word goes 
again after this word,

00:03:24.440 --> 00:03:27.600
if that word was behind that.
- That's pretty much it.

00:03:27.840 --> 00:03:30.960
The reason it can talk like 
a human, is because it has read

00:03:31.433 --> 00:03:33.980
billions of human 
messages.

00:03:34.640 --> 00:03:38.600
So the question comes to a point of:
Okay, you've trained the system.

00:03:38.600 --> 00:03:42.720
How do you actually ensure
that it produces good outputs?

00:03:43.120 --> 00:03:45.433
What do I mean 
by "good outputs"?

00:03:46.386 --> 00:03:50.640
Modern AI companies want 
the so-called "agents"

00:03:50.640 --> 00:03:54.480
to have a, what they call 
a "persona", a personality.

00:03:54.960 --> 00:03:59.560
To be friendly, always 
happy, very obsequious.

00:03:59.560 --> 00:04:01.320
Like if you ask it 
to do something,

00:04:01.320 --> 00:04:04.320
it will always do it
without asking questions.

00:04:04.440 --> 00:04:06.360
To be very useful, 
to be very productive.

00:04:07.640 --> 00:04:09.306
And also to not

00:04:09.413 --> 00:04:11.493
say anything 
that is dangerous.

00:04:11.593 --> 00:04:14.906
If you ask ChatGPT: "Hey, can 
you help me make a bomb?"

00:04:15.360 --> 00:04:17.753
It will say: "I cannot 
help you with that."

00:04:17.866 --> 00:04:19.286
That doesn't come from training.

00:04:19.326 --> 00:04:22.160
That comes from an 
extra layer on top of it.

00:04:22.680 --> 00:04:24.140
How do you do that?

00:04:24.920 --> 00:04:29.000
You cannot just have 
a program doing it.

00:04:29.080 --> 00:04:32.480
At some point, you need to have a human
that looks at the outputs of the text bots

00:04:32.480 --> 00:04:35.640
and says: "Okay, this 
output looks good.

00:04:35.680 --> 00:04:36.840
It's not harmful. Okay?

00:04:36.840 --> 00:04:39.680
This output is very harmful.
Don't do that."

00:04:39.680 --> 00:04:43.840
And this system is called 
reinforcement learning.

00:04:44.200 --> 00:04:45.800
So what do you do
is kind of like

00:04:45.872 --> 00:04:47.106
when you train 
your dog.

00:04:47.180 --> 00:04:49.360
You have positive feedback.
You have negative feedback.

00:04:49.360 --> 00:04:52.360
If you read the message 
and the bot is,

00:04:52.506 --> 00:04:54.920
on the question: "Hey,
can you help me build a bomb?"

00:04:54.920 --> 00:04:57.720
and just  tells you the ingredients,
- that's bad feedback.

00:04:57.720 --> 00:05:01.640
A human needs to press the "bad" 
button, and then it goes ahead.

00:05:03.840 --> 00:05:06.960
Now, as you may imagine, 
a lot of this work,

00:05:07.960 --> 00:05:10.880
and it's a lot of work
because this needs to be done

00:05:10.880 --> 00:05:14.400
not just 1 to 10 times needs to be done
thousands and thousands

00:05:14.400 --> 00:05:18.720
and thousands of times, to get the model
trained enough so it's able to do that.

00:05:19.440 --> 00:05:20.760
And that is not 
only with text.

00:05:20.760 --> 00:05:24.000
It has to be every video 
that it's ever ingested,

00:05:24.360 --> 00:05:27.120
every image, everything,
every piece of audio.

00:05:27.120 --> 00:05:31.280
You have to: one, see if the output 
is good, but also the inputs.

00:05:31.280 --> 00:05:33.480
Sometimes,

00:05:33.480 --> 00:05:36.480
companies get every single 
movie ever made

00:05:37.000 --> 00:05:40.099
to train an AI 
to generate video.

00:05:41.040 --> 00:05:44.440
You have to tell the
AI that this video is, let's say,

00:05:44.680 --> 00:05:47.680
oh, this video is,

00:05:48.240 --> 00:05:51.746
a movie about
...something.

00:05:51.859 --> 00:05:54.386
A human has to see 
the video and label it.

00:05:55.440 --> 00:05:58.906
This process labeling and reinforcement
learning has to be done by humans.

00:05:59.040 --> 00:06:05.080
And what companies do is, 
they outsource this labour

00:06:05.286 --> 00:06:07.840
to most often 
the global South.

00:06:08.520 --> 00:06:11.040
So you have these 
companies, they hire

00:06:11.040 --> 00:06:14.440
hundreds and hundreds 
and hundreds of people

00:06:14.440 --> 00:06:17.440
to be pretty much becoming cog,
a cog in the training machine.

00:06:17.600 --> 00:06:22.393
So their job is to sit up a computer
day after day and look at thousands

00:06:22.440 --> 00:06:26.280
upon thousands 
of pieces of text, video

00:06:26.880 --> 00:06:31.360
and images to label them 
or to train the machine.

00:06:32.000 --> 00:06:36.066
As you may imagine, this work 
is extremely exploitative,

00:06:36.166 --> 00:06:38.760
because it's the global South,
so people don't get paid much.

00:06:39.480 --> 00:06:43.200
It has a mental and emotional toll
that is very high, because

00:06:43.440 --> 00:06:44.960
most of the time,

00:06:44.960 --> 00:06:49.520
the training is to ensure
that the AI doesn't produce bad output.

00:06:50.160 --> 00:06:52.600
So you have to read

00:06:52.600 --> 00:06:55.600
the horrifying texts 
of how on the AI is...

00:06:56.106 --> 00:06:58.800
The question is: "Hey, can you 
tell me how to bury a body?"

00:06:58.860 --> 00:07:03.520
And you have to look at 
facts of how that is done.

00:07:03.560 --> 00:07:05.873
That is a lot 
of mental toll.

00:07:06.720 --> 00:07:07.960
And this is just 
one side.

00:07:07.960 --> 00:07:10.493
There are many other 
systems that are needed

00:07:10.566 --> 00:07:12.533
to make
modern AI work

00:07:12.613 --> 00:07:16.273
that are pretty much hidden 
from the end user.

00:07:16.480 --> 00:07:20.040
As with all capital, the 
means of production

00:07:20.040 --> 00:07:23.680
are alienated from the worker, but
also the product alienated on itself.

00:07:23.680 --> 00:07:27.560
You don't know even 
how we got to the point

00:07:27.640 --> 00:07:31.840
where you can tell a chat bot: "Hey,
can you help me with cooking?"

00:07:33.080 --> 00:07:36.480
There was such a big line of 
the person that originally wrote

00:07:36.480 --> 00:07:40.480
the article on how to cook that food
that then went to the internet.

00:07:40.880 --> 00:07:43.680
Then there was 
a programmer

00:07:43.720 --> 00:07:46.926
that built the system to 
illegally steal that data.

00:07:47.280 --> 00:07:50.520
There was other humans in 
the global South having to,

00:07:51.147 --> 00:07:54.187
label that outputs. - 
And that is just hidden.

00:07:55.739 --> 00:07:57.859
You don't even 
know it's there.

00:07:58.000 --> 00:08:01.380
And there's been 
reporting about this,

00:08:02.160 --> 00:08:05.160
the labeling and 
training by humans

00:08:05.160 --> 00:08:06.906
from the global South.

00:08:08.280 --> 00:08:11.360
There hasn't been 
much reaction to it,

00:08:11.840 --> 00:08:14.920
because it's very similar
to how other modern internet systems

00:08:14.920 --> 00:08:17.920
that are using exploitative work
from the global South are.

00:08:19.500 --> 00:08:22.140
It is very similar to
how de-localization works.

00:08:22.280 --> 00:08:25.800
You have to translate a video game
or a movie into 15 different languages?

00:08:26.280 --> 00:08:29.000
You hire someone from Lionsgate
that is somewhere

00:08:29.000 --> 00:08:32.560
in Vietnam, gets 
paid 30 Cent every hour.