April 1, 2023
#80 Tackling Misinformation with AI: Insights from Alex Fink Founder of Otherweb
In this episode, we speak with Alex Fink, founder of Otherweb, an app that uses AI to filter out "junk" content and combat misinformation online. Alex shares his insights on the challenges facing the current information ecosystem, the role of AI in detecting and stopping misinformation, and the technical aspects of Otherweb's content filtering. He also provides practical tips for navigating the complex landscape of online information and discusses the future of the information ecosystem.
https://otherweb.com/
1
00:00:06,500 --> 00:00:09,400
Hello, and welcome to a new
episode of the CTO show with
2
00:00:09,400 --> 00:00:10,800
Mohammed.
My name is Mohammed.
3
00:00:11,200 --> 00:00:13,900
And as, you know, in each
episode, I share some
4
00:00:13,900 --> 00:00:17,800
information about the latest
trends inside strategies, and
5
00:00:17,800 --> 00:00:22,000
topic, like cybersecurity
digital transformation, emerging
6
00:00:22,000 --> 00:00:25,000
Tech, startups and
Entrepreneurship and sometime.
7
00:00:25,000 --> 00:00:28,300
I have thought leaders, and
innovators and entrepreneurs to
8
00:00:28,400 --> 00:00:33,400
let them share with us, you
know, their expertise in the
9
00:00:33,400 --> 00:00:35,700
industry.
They are in and how they are
10
00:00:35,700 --> 00:00:37,900
making.
Impact in the world of tech and
11
00:00:37,900 --> 00:00:40,000
business.
And today I'm very pleased to
12
00:00:40,000 --> 00:00:44,200
have with me on the show.
Alex, Alex is joining me from
13
00:00:44,200 --> 00:00:49,100
the US Alex.
He is the an entrepreneur for
14
00:00:49,100 --> 00:00:51,300
long time.
I will leave him to introduce
15
00:00:51,500 --> 00:00:55,400
himself and what he does.
Alright, so thank you so much
16
00:00:55,400 --> 00:00:57,200
for having me.
It's great to be on the show.
17
00:00:57,500 --> 00:01:00,600
My name is Alex Fink.
I'm the founder and CEO of the
18
00:01:00,600 --> 00:01:03,000
other web or other web ink,
right?
19
00:01:03,700 --> 00:01:07,800
We are a company that is
dedicated to Trying to fix the
20
00:01:07,800 --> 00:01:11,900
world's information ecosystem or
in other words, if we make it a
21
00:01:11,900 --> 00:01:14,400
little more practical, we want
to help people consume higher
22
00:01:14,400 --> 00:01:17,200
quality information.
Now, the way that works is first
23
00:01:17,200 --> 00:01:20,800
we've developed a set of filters
that evaluate information and
24
00:01:20,800 --> 00:01:22,700
figure out what is junk and
throw it out.
25
00:01:22,700 --> 00:01:26,100
And then we created an
information platform that
26
00:01:26,100 --> 00:01:31,300
Aggregates content from all over
the web news commentary podcasts
27
00:01:31,300 --> 00:01:34,900
research, studies, everything we
can get our Crawlers on and then
28
00:01:34,900 --> 00:01:37,400
we filter the junk out and we
Create a feed for people that
29
00:01:37,400 --> 00:01:40,700
they can customize so that they
decide what they consume instead
30
00:01:40,700 --> 00:01:42,600
of social media deciding for
them.
31
00:01:43,700 --> 00:01:46,900
Great great.
What was your opinion motive
32
00:01:46,900 --> 00:01:50,700
Alex to start this?
What how it all started?
33
00:01:51,200 --> 00:01:55,100
So it's been brewing in my mind
for a very long time, right?
34
00:01:55,100 --> 00:01:57,300
And I've been an information
junkie for a long time.
35
00:01:57,300 --> 00:02:01,800
So I've been watching this
ecosystem deteriorate basically
36
00:02:01,800 --> 00:02:04,800
throughout my life and almost
every single outlet.
37
00:02:04,800 --> 00:02:08,100
If you look at it today if it's
This the 20 years ago it was
38
00:02:08,100 --> 00:02:12,000
better 20 years ago, they are
all becoming worse and so I was
39
00:02:12,000 --> 00:02:14,600
trying to figure out why for a
long time and trying to figure
40
00:02:14,600 --> 00:02:18,100
out what incentives, drive them
towards all becoming clickbait
41
00:02:18,100 --> 00:02:21,500
centrally.
Now on a personal level, I've
42
00:02:21,500 --> 00:02:24,600
been building perception systems
for 15 years, so cameras,
43
00:02:24,600 --> 00:02:25,800
computer vision, that sort of
thing.
44
00:02:26,100 --> 00:02:29,500
And so, sometime last year, I
had this crisis of conscience of
45
00:02:30,200 --> 00:02:33,800
does the world need more cameras
and I decided it doesn't.
46
00:02:33,800 --> 00:02:36,000
And I need to do something more
useful with my life.
47
00:02:36,200 --> 00:02:39,500
To avoid having this disconnect
where I built cameras to make
48
00:02:39,500 --> 00:02:41,700
money.
But I keep thinking that the
49
00:02:41,708 --> 00:02:44,000
biggest problem is that people
are consuming junk.
50
00:02:44,200 --> 00:02:48,200
So, I decided to connect those
two and actually start devoting
51
00:02:48,200 --> 00:02:52,100
my time towards fixing this big
problem that I see in the world
52
00:02:52,100 --> 00:02:54,700
that everything around us is
becoming junk and that's what
53
00:02:54,700 --> 00:02:58,200
people can seal.
That's, I mean, it's very
54
00:02:59,500 --> 00:03:02,800
inspirational.
I would say, especially, you
55
00:03:02,800 --> 00:03:05,900
know, like everyone has, you
said they try to make money, but
56
00:03:05,900 --> 00:03:08,200
for you, It's like you go out on
a mission.
57
00:03:08,500 --> 00:03:12,600
Now, if I can, you know, ask you
this.
58
00:03:12,700 --> 00:03:15,700
So it's all about Miss
information, right?
59
00:03:15,700 --> 00:03:18,300
And all these things.
So, how do you think artificial
60
00:03:18,300 --> 00:03:22,300
intelligence can play a role in
fighting against?
61
00:03:22,300 --> 00:03:26,700
You know, this misinformation
propaganda and how you at other
62
00:03:26,700 --> 00:03:29,800
web you are leveraging this in
to filter the content.
63
00:03:29,900 --> 00:03:32,700
So first of all, let me say that
it can play a role in both
64
00:03:32,700 --> 00:03:35,300
directions, right?
And especially now that you're
65
00:03:35,300 --> 00:03:38,300
seeing these And of big advances
in generative AI.
66
00:03:38,700 --> 00:03:42,800
They can play a role in making
things worse, because if you
67
00:03:42,800 --> 00:03:45,500
think about what it took 10
years ago to write the clickbait
68
00:03:45,500 --> 00:03:49,400
article, you have to a be test
various headlines.
69
00:03:49,400 --> 00:03:53,300
You have to actually compose
reasonably coherent text.
70
00:03:54,400 --> 00:03:57,100
You don't have to do any of that
stuff anymore, judge if he can
71
00:03:57,100 --> 00:03:59,500
do that for you.
And so right now that same
72
00:03:59,500 --> 00:04:02,300
person generating Jung called a
can generate 10 times, the
73
00:04:02,308 --> 00:04:05,300
amount of junk that he used to
generate a year ago.
74
00:04:05,700 --> 00:04:09,500
So If left unchecked to its own
devices, this entire ecosystem
75
00:04:09,500 --> 00:04:13,300
will get worse because of a.
I now, the way we view AI is
76
00:04:13,300 --> 00:04:17,200
thinking, okay?
What is needed to clean?
77
00:04:17,200 --> 00:04:21,000
Things up, probably more
editors, and we need those
78
00:04:21,000 --> 00:04:24,400
editors to be more consistent.
We need them to never have a bad
79
00:04:24,400 --> 00:04:26,200
day.
We need them to have great
80
00:04:26,200 --> 00:04:29,200
throughput.
So essentially, we are training
81
00:04:29,200 --> 00:04:32,500
models that behave as newspaper
editors, right?
82
00:04:32,700 --> 00:04:36,000
So we take some particular
trait, like, checking the
83
00:04:36,200 --> 00:04:38,500
Headline, to make sure it
matches the body of the article.
84
00:04:38,700 --> 00:04:42,400
And we trained about a model
that does just that, right?
85
00:04:42,400 --> 00:04:45,300
And it does it almost as good as
a human editor, but it never has
86
00:04:45,300 --> 00:04:50,300
a bad day and never and it can
be completely transparent.
87
00:04:50,300 --> 00:04:54,500
So we also open the source of
all our models made them Source
88
00:04:54,500 --> 00:04:57,800
available because we want people
to be able to actually look at
89
00:04:57,800 --> 00:04:59,600
them and see.
There is nothing hidden in
90
00:04:59,600 --> 00:05:01,200
there.
We're not just randomly
91
00:05:01,200 --> 00:05:03,800
penalizing everybody who's right
wing or everybody who's left
92
00:05:03,800 --> 00:05:06,000
wing, we're just trying to train
it on.
93
00:05:06,300 --> 00:05:10,000
Academic data sets or data sets
that we scrambled together to
94
00:05:10,000 --> 00:05:12,500
make them as balanced and clear
as possible.
95
00:05:12,500 --> 00:05:18,000
So that's where a I can do some
good but ultimately it's a
96
00:05:18,000 --> 00:05:20,100
kitchen knife.
You can make a salad or you can
97
00:05:20,100 --> 00:05:25,100
stab someone it's up to you.
Yeah, so if if I understood
98
00:05:25,100 --> 00:05:30,400
right, so should we say that
your platform would get better
99
00:05:30,400 --> 00:05:33,200
with time because the you know,
it's a matter of fact, the
100
00:05:33,200 --> 00:05:35,900
better you train a model, the
battery.
101
00:05:36,100 --> 00:05:38,600
Becomes later.
So is this like what we should
102
00:05:38,600 --> 00:05:42,600
expect from from other web.
Absolutely, it gets better and
103
00:05:42,600 --> 00:05:45,600
time because we gather more data
and we feed that back and
104
00:05:45,600 --> 00:05:48,100
retrain the model.
But also because we get feedback
105
00:05:48,100 --> 00:05:51,000
from our users.
So the initial set of models we
106
00:05:51,000 --> 00:05:56,100
did our entire feedback was we
showed some examples of the
107
00:05:56,100 --> 00:05:58,900
results to 5 friends that I
have.
108
00:05:58,900 --> 00:06:01,600
And they told me whether it kind
of conforms to their Intuition
109
00:06:01,600 --> 00:06:05,000
or not.
So that's a very low bar for
110
00:06:05,100 --> 00:06:07,900
like, the last I was just
testing your model against
111
00:06:07,900 --> 00:06:10,100
reality, right?
We certainly didn't have the
112
00:06:10,100 --> 00:06:13,600
budget that GPT forehead now
where they had thousands of
113
00:06:13,600 --> 00:06:16,700
people working for nine months
if I understood correctly.
114
00:06:16,700 --> 00:06:20,100
Right?
Tragic, tragic jet GPT for was
115
00:06:20,100 --> 00:06:23,300
done in summer.
They released it now right?
116
00:06:23,300 --> 00:06:26,700
So this entire time they've been
doing human feedback essential
117
00:06:27,200 --> 00:06:30,400
so now we have natural human
feedback, we have a quarter
118
00:06:30,400 --> 00:06:33,600
million users that actually read
the news and they flag stuff
119
00:06:33,600 --> 00:06:36,000
that seems off right?
That's it.
120
00:06:36,100 --> 00:06:38,300
Big piece of feedback for us,
right?
121
00:06:38,300 --> 00:06:43,100
We can also contact them and ask
them and now we also rolled out
122
00:06:43,500 --> 00:06:47,700
a version of the app that
essentially forces, the users to
123
00:06:48,300 --> 00:06:50,700
interact with content to certain
way, right?
124
00:06:50,700 --> 00:06:53,100
They have to swipe an item to
get to the next one.
125
00:06:53,300 --> 00:06:56,400
And the direction of their swipe
is feedback for us about how
126
00:06:56,400 --> 00:06:59,300
good this item was.
And so over time, we will learn
127
00:06:59,300 --> 00:07:03,100
to incorporate that real human
feedback as well and to making
128
00:07:03,100 --> 00:07:05,600
our models better.
And I'm talking here about the
129
00:07:05,600 --> 00:07:08,400
filtering Girls, I'm talking
about the Sorting of the feed
130
00:07:08,800 --> 00:07:11,600
and I'm also talking about
things like summarizes because
131
00:07:11,600 --> 00:07:15,500
we show users a summary of the
article that we create instead
132
00:07:15,500 --> 00:07:17,600
of the original excerpt from the
article.
133
00:07:18,400 --> 00:07:22,500
Yeah.
Like maybe this is should be
134
00:07:22,500 --> 00:07:27,300
also related but like you
started this so people can get
135
00:07:27,500 --> 00:07:33,100
as you said like more clean and
remove the junk from from their
136
00:07:33,100 --> 00:07:36,000
feet but it will like what do
you think?
137
00:07:36,100 --> 00:07:39,500
Think, you know, people should
really care about especially in
138
00:07:39,500 --> 00:07:43,400
this, you know, we are living
now in a, in a world where you
139
00:07:43,408 --> 00:07:45,600
have a lot of things coming in,
something comes from the
140
00:07:45,600 --> 00:07:48,800
mainstream media.
Something comes from digital
141
00:07:49,000 --> 00:07:51,200
media.
Something comes from even, you
142
00:07:51,200 --> 00:07:54,800
know, companies where they work.
So what do you think, you know,
143
00:07:54,800 --> 00:08:00,800
they can do to navigate better
this complex landscape of online
144
00:08:00,800 --> 00:08:03,600
information?
So, I think the best analogy to
145
00:08:03,600 --> 00:08:07,200
keep in mind is the way we Food,
right?
146
00:08:07,900 --> 00:08:09,600
Information is food for the
brain.
147
00:08:09,600 --> 00:08:13,500
So the way that we look at
improving your normal diet is
148
00:08:13,500 --> 00:08:16,000
probably the way that we should
look at improving our digital
149
00:08:16,000 --> 00:08:18,100
day.
So when it comes to normal food,
150
00:08:18,100 --> 00:08:20,300
if you go to the doctor, doctor
tells you you're unhealthy, you
151
00:08:20,300 --> 00:08:22,500
need to eat better food, you
come back home?
152
00:08:22,500 --> 00:08:25,200
What's your next step?
The first step is probably going
153
00:08:25,200 --> 00:08:28,600
to be to cut out the things that
are obviously Just Junk.
154
00:08:28,600 --> 00:08:30,600
They're terrible, write the
empty calories.
155
00:08:30,900 --> 00:08:33,299
I don't put sugar in what your
drink.
156
00:08:33,500 --> 00:08:35,900
Don't eat cookies all day,
remove the obvious.
157
00:08:36,100 --> 00:08:40,000
We bad stuff, that's step one.
So we try to help with that.
158
00:08:40,100 --> 00:08:43,799
But obviously users still need
to pay some attention to other
159
00:08:43,799 --> 00:08:45,700
sources of empty calories that
they have.
160
00:08:45,800 --> 00:08:48,400
Now, once you do that, your job
is not over.
161
00:08:48,700 --> 00:08:51,600
You still have to figure out how
to balance the things that you
162
00:08:51,600 --> 00:08:56,200
do it in a reasonable way.
And I've seen some pretty crazy
163
00:08:56,200 --> 00:08:59,400
studies about what happens when
you overdose on one type of
164
00:08:59,400 --> 00:09:02,900
content.
There's a study from 2013 where
165
00:09:02,900 --> 00:09:07,800
researchers compared people who
were at the Marathon bombing and
166
00:09:07,800 --> 00:09:11,100
actually witnessed the bombing
to people who watched six hours
167
00:09:11,100 --> 00:09:15,300
of news coverage about the
bombing the group that watch the
168
00:09:15,300 --> 00:09:19,100
news coverage had higher chances
of PTSD than the group that what
169
00:09:19,100 --> 00:09:23,400
that was in the event.
So news is obviously good.
170
00:09:23,400 --> 00:09:25,100
You want to know what's going on
in the world?
171
00:09:25,400 --> 00:09:29,500
But 15 minutes is good 6 hours
on a single bombing is probably
172
00:09:29,500 --> 00:09:32,800
really bad, right?
And so you have to learn to
173
00:09:32,800 --> 00:09:35,800
balance what you have and
there's many different sources
174
00:09:35,800 --> 00:09:36,500
of information.
Information.
175
00:09:36,500 --> 00:09:39,700
We could consume could be Dance
videos on Tick.
176
00:09:39,700 --> 00:09:44,000
Tock could be serious.
News coverage could be books
177
00:09:44,000 --> 00:09:47,200
written by somebody. 2,000 years
ago about philosophy, right?
178
00:09:47,200 --> 00:09:50,500
You want to balance all of these
out in some reasonable way?
179
00:09:51,500 --> 00:09:53,300
What is it?
That depends on you?
180
00:09:53,300 --> 00:09:55,000
There's probably different
preferences.
181
00:09:55,000 --> 00:09:58,000
Just like with diet, see of the
keto people, and the vegan
182
00:09:58,000 --> 00:10:01,400
people in the Mediterranean, all
of these right there, all better
183
00:10:01,400 --> 00:10:03,800
than just mindlessly consuming.
What's in front of you?
184
00:10:04,000 --> 00:10:07,600
So I think it's the same thing
with You should just sit down
185
00:10:07,800 --> 00:10:10,400
maybe track a little bit.
What you're consuming right now.
186
00:10:10,600 --> 00:10:12,400
Yeah.
And try to determine what you
187
00:10:12,400 --> 00:10:14,500
want to do.
I want to mention one more thing
188
00:10:14,500 --> 00:10:20,200
with regards to how we handle it
and the other web and that the
189
00:10:20,200 --> 00:10:23,500
vast majority of platforms out
there, try to maximize your time
190
00:10:23,500 --> 00:10:25,900
on site because then they can
show you more ads.
191
00:10:25,900 --> 00:10:28,700
That's their incentive.
But if you think about your
192
00:10:28,700 --> 00:10:33,500
incentive as a human, you don't
want to maximize any single food
193
00:10:33,500 --> 00:10:35,900
group.
Why would you want to spend?
194
00:10:36,000 --> 00:10:38,800
And too much time on any
particular site, you want to get
195
00:10:38,800 --> 00:10:42,700
as much value from this site in
as little time as possible,
196
00:10:42,700 --> 00:10:44,600
right?
You don't want to spend as much
197
00:10:44,600 --> 00:10:48,400
time as possible on Technical
and so, but the average person
198
00:10:48,400 --> 00:10:51,700
spends an hour and a half, a day
of data, which is pretty bad if
199
00:10:51,700 --> 00:10:54,300
you think about it.
So our approach to it, as we
200
00:10:54,300 --> 00:10:56,500
want to help you to get as much
value as you can.
201
00:10:56,500 --> 00:11:00,000
And then stop, there is no
reason for you to just continue
202
00:11:00,000 --> 00:11:01,400
consuming more and more and
more.
203
00:11:01,900 --> 00:11:05,400
We think that if you tell us you
want to read 50 summaries of
204
00:11:05,400 --> 00:11:09,400
articles, We're going to stop
you at 50 because you don't want
205
00:11:09,400 --> 00:11:11,600
to read more 50 is your quarter,
right?
206
00:11:12,000 --> 00:11:14,300
It's counterintuitive.
It goes proudly against
207
00:11:14,300 --> 00:11:17,700
Financial incentives.
It's why we had to register as a
208
00:11:17,700 --> 00:11:20,900
public benefit Corporation
because I think as a c chord, we
209
00:11:20,900 --> 00:11:23,000
would not be able to do this
very long, right.
210
00:11:23,000 --> 00:11:27,900
Our investors would force us to
maximize time on site like
211
00:11:27,900 --> 00:11:31,300
everybody else does.
So how do you monetize just out
212
00:11:31,300 --> 00:11:33,500
of curiosity?
We haven't started yet.
213
00:11:33,900 --> 00:11:37,700
If we have some idea, Ideas on
what we want to experiment with
214
00:11:37,900 --> 00:11:39,600
but we want to start those
experiments.
215
00:11:39,600 --> 00:11:42,400
When there is enough users that
we can actually see what value
216
00:11:42,400 --> 00:11:45,400
they get out of the platform and
how we can monetize without
217
00:11:45,400 --> 00:11:48,600
interfering with that.
One of the interesting
218
00:11:49,000 --> 00:11:55,400
hypotheses we have is that we
think when users consume content
219
00:11:55,400 --> 00:11:59,600
and kind of a passive way like a
feed, eventually they notice
220
00:11:59,600 --> 00:12:00,800
something that Sparks their
interest.
221
00:12:00,800 --> 00:12:03,500
Like, you see a phrase in an
article that you want to know
222
00:12:03,500 --> 00:12:08,000
more about if this happens in a
And then you typically just ask
223
00:12:08,000 --> 00:12:11,200
what's that, right?
But if this happens online, we
224
00:12:11,200 --> 00:12:13,700
have this weird pattern where
people open the browser window
225
00:12:13,700 --> 00:12:16,700
type something into the address
bar, click enter select the
226
00:12:16,700 --> 00:12:19,500
result and then what they see is
usually unrelated to what they
227
00:12:19,500 --> 00:12:21,700
searched because the entire
complex was lost and all these
228
00:12:21,700 --> 00:12:25,200
transitions, right?
So we implemented a what's that
229
00:12:25,200 --> 00:12:27,600
into the platform?
And we see that the average
230
00:12:27,600 --> 00:12:30,900
newsreader on our platform uses
that two or three times per
231
00:12:30,900 --> 00:12:35,300
session and so and when they do,
we show them something that
232
00:12:35,300 --> 00:12:39,100
looks a lot like Search results.
And we know pretty well how to
233
00:12:39,100 --> 00:12:42,800
put ads on search results in a
way that doesn't interfere with
234
00:12:42,800 --> 00:12:46,300
the users experience.
So our hope, and our hypothesis
235
00:12:46,300 --> 00:12:48,700
right now is ads on search
results are enough.
236
00:12:49,100 --> 00:12:53,000
And we never need to put ads on
the feed because that's annoying
237
00:12:53,000 --> 00:12:56,600
and we don't want to do that.
But ultimately, we have to test
238
00:12:56,600 --> 00:12:59,800
and see how users react and make
sure users are happy.
239
00:13:00,100 --> 00:13:01,700
Yeah, that's really interesting
model.
240
00:13:01,700 --> 00:13:05,300
I would say now.
My understanding.
241
00:13:05,300 --> 00:13:10,200
And what we discussed is you
know, like having this garbage
242
00:13:10,200 --> 00:13:13,700
cleaning I would say of read
content.
243
00:13:13,700 --> 00:13:19,500
Right now, something I brought
many times on, on the, on the
244
00:13:19,500 --> 00:13:23,100
show is about the fake where I
talked myself, you know, about,
245
00:13:23,200 --> 00:13:26,500
you know, the fake and what's
happening and even II tried to
246
00:13:26,500 --> 00:13:31,600
do a few want to call it a proof
of concept where I used multiple
247
00:13:31,700 --> 00:13:35,500
AI tools and, you know, I Myself
a video, of course it was like
248
00:13:35,500 --> 00:13:39,700
still a better version of that
software, but we've seen
249
00:13:39,700 --> 00:13:43,900
recently with the generative AI
really thinks that are scary.
250
00:13:45,100 --> 00:13:48,500
How do you think this can be
treated by mean from the
251
00:13:48,500 --> 00:13:50,600
mission?
You are in now because you're
252
00:13:50,600 --> 00:13:53,700
trying to remove right?
Like faking use you're trying to
253
00:13:53,700 --> 00:13:55,600
to clean something, which
nonsense.
254
00:13:56,000 --> 00:13:58,200
How do you think you can tackle
this?
255
00:13:59,100 --> 00:14:02,600
So it's an interesting problem
and there will be a sort of an
256
00:14:02,600 --> 00:14:05,900
arms race between the fakers and
the guys who try to spot the
257
00:14:05,900 --> 00:14:08,800
fakers, right?
The way that I would view this,
258
00:14:08,800 --> 00:14:11,700
as first of all, if you look at
all the different types of bad
259
00:14:11,700 --> 00:14:18,300
content out there content that
is faked on purpose with effort
260
00:14:18,300 --> 00:14:22,400
from the faker is actually the
minority most of the bad content
261
00:14:22,400 --> 00:14:26,300
out there is not fake news.
It's just not news at all,
262
00:14:26,800 --> 00:14:27,700
right?
It's CNN.
263
00:14:27,700 --> 00:14:30,600
Publishing an article with a
title like stop what you're
264
00:14:30,600 --> 00:14:33,300
doing and watch this elephant
play with bubbles, that's an
265
00:14:33,300 --> 00:14:39,000
actual CNN article, right?
So, I think that to actually
266
00:14:39,000 --> 00:14:43,600
focus on Thinking, critically
about content and trying to
267
00:14:43,600 --> 00:14:47,500
notice the fakes we first have
to get rid of the obviously
268
00:14:47,500 --> 00:14:49,900
about stuff.
The stuff that is bad, just
269
00:14:49,900 --> 00:14:53,000
based on form without even
looking at the substance, and
270
00:14:53,000 --> 00:14:55,300
that is the majority.
So once you clean all of that
271
00:14:55,300 --> 00:15:00,600
out, you will be left with the
more stubborn cases of Russian
272
00:15:00,600 --> 00:15:06,000
propaganda or of the Democrats,
trying to create actual fake
273
00:15:06,000 --> 00:15:08,600
stories about Republicans or
Republicans trying to create
274
00:15:08,600 --> 00:15:10,100
actual fake stories about
Democrats.
275
00:15:10,100 --> 00:15:13,900
Those will be Harder to spot and
I don't think you can ever clean
276
00:15:13,900 --> 00:15:17,400
it out completely.
So even if you look at science,
277
00:15:17,700 --> 00:15:19,400
right, you look at scientific
peer-reviewed.
278
00:15:19,400 --> 00:15:24,000
Publishing the last figure I saw
is that in almost every decade,
279
00:15:24,000 --> 00:15:28,500
this figure is kind of fixed the
number of completely fake
280
00:15:29,600 --> 00:15:34,100
experiments published or papers
published was about 14%, right?
281
00:15:34,100 --> 00:15:35,800
So, fourteen percent of
everything.
282
00:15:35,800 --> 00:15:39,000
You see that this published in
peer-reviewed journals is
283
00:15:39,000 --> 00:15:41,000
usually a paper that describes
an experiment.
284
00:15:41,000 --> 00:15:43,900
That They're didn't happen or
the result of the experiment is
285
00:15:43,900 --> 00:15:47,800
not what the paper says it.
Right, it's hard to spot.
286
00:15:48,500 --> 00:15:51,400
You can't ever, if you're a peer
reviewer, you cannot just drive
287
00:15:51,400 --> 00:15:53,900
to somebody's lab and start
counting the mice, right?
288
00:15:54,300 --> 00:15:56,600
You have to believe them, when
they tell you, I started with
289
00:15:56,600 --> 00:15:59,000
1713 survive.
Therefore, this is the death
290
00:15:59,000 --> 00:16:02,100
rate, right?
So, typically peer reviewers,
291
00:16:02,100 --> 00:16:05,100
focus on foreign they don't
focus on going to the lab and
292
00:16:05,100 --> 00:16:07,700
double-checking the number of
miles they just see, okay,
293
00:16:07,700 --> 00:16:11,700
structured correctly, described
correctly headline, You subtract
294
00:16:11,700 --> 00:16:13,500
matches body?
Correct references.
295
00:16:13,600 --> 00:16:14,800
All right.
Clear to publish.
296
00:16:17,400 --> 00:16:19,100
Detecting, whether something was
fake.
297
00:16:19,100 --> 00:16:22,200
All along, might take years, you
will not find it in real time.
298
00:16:22,700 --> 00:16:26,000
So let's start from the easy
problems, then the intermediate
299
00:16:26,000 --> 00:16:29,300
problems and maybe there will be
some number of intractable
300
00:16:29,300 --> 00:16:32,200
problems that we cannot fix.
Now, when you're talking about
301
00:16:32,200 --> 00:16:36,900
deep deep Faith videos, that's
an interesting conundrum in that
302
00:16:37,900 --> 00:16:40,900
You will see those people will
create those videos, then a
303
00:16:40,900 --> 00:16:43,900
journalist will write about it
and then we will take it up in
304
00:16:43,900 --> 00:16:48,600
our aggregator.
So to some extent, the best we
305
00:16:48,600 --> 00:16:50,700
can do is cross-reference
between the journals.
306
00:16:50,700 --> 00:16:54,200
Because again, we cannot be the
guys who verified the video
307
00:16:54,200 --> 00:16:57,700
itself.
We are writing about the person
308
00:16:57,700 --> 00:17:01,400
who wrote about the video in
some sense and so the best we
309
00:17:01,400 --> 00:17:03,300
can do is come up with some
heuristics on.
310
00:17:03,300 --> 00:17:08,599
How do you cross reference Maybe
if left-wing Outlets publish
311
00:17:08,599 --> 00:17:11,099
something, right?
Wing Outlets say that it's fake,
312
00:17:11,900 --> 00:17:15,300
then it's only half true until
both sides, agree on it, right?
313
00:17:16,500 --> 00:17:18,900
That's actually a heuristic.
I use for myself.
314
00:17:19,000 --> 00:17:22,400
Generally speaking things that
are obviously true typically
315
00:17:22,400 --> 00:17:25,700
things that people who disagree
on everything still agree on,
316
00:17:26,300 --> 00:17:28,300
right?
Otherwise, if one side says it's
317
00:17:28,300 --> 00:17:30,000
true in the other one says it's
not.
318
00:17:30,000 --> 00:17:32,200
It's probably only half true it
best.
319
00:17:32,800 --> 00:17:36,400
So I don't know, I don't think
it's entirely solvable.
320
00:17:36,400 --> 00:17:39,600
We will continue to develop
better tools, the fakers will
321
00:17:39,600 --> 00:17:42,600
continue to develop better tools
as well, so it will be an arms
322
00:17:42,600 --> 00:17:44,600
race.
Yeah, it's similar to, you know,
323
00:17:44,600 --> 00:17:47,500
because I cover cyber security
as well at the same case as
324
00:17:47,500 --> 00:17:49,500
right?
So always you have the
325
00:17:50,400 --> 00:17:53,900
researchers trying to, you know,
close the gap and then you see
326
00:17:53,900 --> 00:17:56,900
the other side, the bad guys.
They are also using, you know,
327
00:17:56,900 --> 00:18:00,200
latest Technologies and so it's
like an air is that it's
328
00:18:00,200 --> 00:18:03,300
never-ending.
But one thing which is also
329
00:18:03,300 --> 00:18:07,200
like, Very interesting in what
you're trying to do.
330
00:18:07,400 --> 00:18:12,500
So usually you know you try to
keep people and you said this,
331
00:18:12,500 --> 00:18:15,900
you try to keep people on your
platform as much as possible and
332
00:18:15,900 --> 00:18:18,400
now you're doing the complete
opposite.
333
00:18:18,500 --> 00:18:22,200
I mean in a sense that you're
telling people enough right?
334
00:18:23,400 --> 00:18:27,000
But from user Behavior
perspective, do you think like
335
00:18:27,000 --> 00:18:32,200
this will be something that the
end user will be fine with it
336
00:18:32,200 --> 00:18:35,400
like or not?
You've done some, some, I mean,
337
00:18:35,400 --> 00:18:39,700
testing on that.
So testing is ongoing so far.
338
00:18:39,700 --> 00:18:41,900
The number of users is growing
pretty fast, right.
339
00:18:41,900 --> 00:18:45,400
But we'll see at what point we
need to cross the chasm, right?
340
00:18:45,400 --> 00:18:48,200
We were going to probably hit
the ceiling at some point with
341
00:18:48,200 --> 00:18:53,100
early adopters and then have to
change what we provide to fit
342
00:18:53,300 --> 00:18:54,800
more normal users.
Right?
343
00:18:55,300 --> 00:18:57,500
I don't know when that
transition point is coming, but
344
00:18:57,500 --> 00:19:01,000
it's coming at some point.
So we're still looking at that
345
00:19:01,000 --> 00:19:02,400
feedback.
It's not that we're telling
346
00:19:02,400 --> 00:19:04,700
users enough, right?
I think what we actually
347
00:19:04,700 --> 00:19:07,100
implemented is a
congratulations.
348
00:19:07,400 --> 00:19:10,100
You're now well-informed you
reach their quota for the day,
349
00:19:10,300 --> 00:19:11,500
right?
If they want to continue, they
350
00:19:11,500 --> 00:19:13,200
can continue.
But at least they got this.
351
00:19:13,200 --> 00:19:16,100
Reminder, this was your quarter,
you're done.
352
00:19:16,700 --> 00:19:19,100
It's really want to keep going,
right, right.
353
00:19:19,200 --> 00:19:21,300
But it's also phrased as a
reward, which gives them
354
00:19:21,300 --> 00:19:24,400
positive emotion, instead of
negative emotion, right?
355
00:19:25,100 --> 00:19:26,600
But at least that's the
hypothesis.
356
00:19:26,600 --> 00:19:30,400
Was we still have to test all of
these nudges that we've built
357
00:19:30,400 --> 00:19:33,200
into our system?
So I don't know.
358
00:19:33,200 --> 00:19:35,500
I'm more Read about the
different angle tool.
359
00:19:35,500 --> 00:19:40,800
This, my real concern is that
what every other tool does?
360
00:19:40,800 --> 00:19:43,500
It's not just that they're
trying to maximize time on site,
361
00:19:43,900 --> 00:19:46,800
but they're trying to maximize
your time on site by using the
362
00:19:46,800 --> 00:19:50,800
tools of addiction by variable
rewards to everything that you
363
00:19:50,800 --> 00:19:53,200
see by triggering strong
emotions.
364
00:19:53,400 --> 00:19:57,400
So that, even if that emotion is
- you crave strong, emotion,
365
00:19:57,400 --> 00:20:00,600
next time you see a headline,
right, and people have gotten so
366
00:20:00,600 --> 00:20:03,200
accustomed to these oversized
emotions.
367
00:20:03,200 --> 00:20:07,200
Whenever they Read the news that
now when we give them a clean
368
00:20:07,200 --> 00:20:11,000
version of the news that doesn't
have any of that hate or
369
00:20:11,000 --> 00:20:13,500
agitation or anxiety or that
kind of stuff.
370
00:20:14,000 --> 00:20:16,300
They might be missing something,
right?
371
00:20:16,300 --> 00:20:19,400
They might have been so
accustomed to a know the
372
00:20:19,400 --> 00:20:22,300
sugar-laden version of a cookie
that now that we give them a
373
00:20:22,308 --> 00:20:24,100
real organic cookie.
They don't know what to do with
374
00:20:24,100 --> 00:20:26,400
it anymore, or at least a lot of
people want.
375
00:20:26,500 --> 00:20:30,400
And so, that's my big concern
and that's where most of our
376
00:20:30,700 --> 00:20:32,900
growth hacking experiments are
focusing.
377
00:20:33,100 --> 00:20:37,300
Its what can we substitute?
The bad stuff with once we
378
00:20:37,300 --> 00:20:41,700
remove it because it seems like
just removing might not work.
379
00:20:41,700 --> 00:20:45,800
It's again, just imagine that
some company decided to remove
380
00:20:45,900 --> 00:20:49,200
all the sugar out of older
cookies, people learn going to
381
00:20:49,200 --> 00:20:51,600
buy those cookies, right?
You have to put at least some
382
00:20:51,600 --> 00:20:54,300
erythritol back.
Something to make it resemble,
383
00:20:54,300 --> 00:20:56,200
what people are used to eat,
right?
384
00:20:56,400 --> 00:20:58,800
So that's kind of our big
challenge.
385
00:20:58,800 --> 00:21:03,400
I think the challenge of
reminding people that Their
386
00:21:04,000 --> 00:21:06,200
quota for the day was 50 and
they're about 50.
387
00:21:06,200 --> 00:21:09,000
Now, I don't think people are
going to revolt against that,
388
00:21:09,000 --> 00:21:11,700
that sounds like just the pure
benefit to them, right?
389
00:21:12,100 --> 00:21:17,800
But removing the team sport Vibe
of just hate the other guy.
390
00:21:18,800 --> 00:21:21,700
People might revolt against that
because they're enjoying the
391
00:21:21,700 --> 00:21:24,900
team sport or at least they
think they do because they're so
392
00:21:24,900 --> 00:21:26,200
accustomed to it.
Right.
393
00:21:26,300 --> 00:21:28,700
Yeah.
One thing also that came to my
394
00:21:28,700 --> 00:21:32,100
mind, how you can deal with, I
mean, do you have anything?
395
00:21:32,200 --> 00:21:35,800
Thing that can, you know,
overcome gated content.
396
00:21:36,800 --> 00:21:40,700
So right now we try to avoid it
because we respect the fact that
397
00:21:40,700 --> 00:21:44,600
content is Gator, right?
The reality is most of the Gated
398
00:21:44,600 --> 00:21:47,800
content is actually gated in a
way that only stops humans and
399
00:21:47,800 --> 00:21:50,600
does not stop crawlers at all
because they still want to be
400
00:21:50,600 --> 00:21:52,500
indexed by every search engine
of the right.
401
00:21:52,700 --> 00:21:57,300
And so if I just point my
crawler to a link from Financial
402
00:21:57,300 --> 00:21:59,700
Times, it's going to scrape the
full article from Financial
403
00:21:59,700 --> 00:22:01,900
Times.
Do I want to publish even
404
00:22:02,200 --> 00:22:04,100
Headline or the summary of that
article.
405
00:22:04,100 --> 00:22:06,200
No.
Because it seems to me like if
406
00:22:06,200 --> 00:22:09,500
they put the gate in front of
Financial Times, they're telling
407
00:22:09,500 --> 00:22:11,100
me something.
They're telling me, this is for
408
00:22:11,100 --> 00:22:14,300
my subscribers only, right?
And so we respect that and
409
00:22:14,300 --> 00:22:19,200
typically, initially, we were
just essentially scraping all
410
00:22:19,200 --> 00:22:21,600
the links that somebody posted
on Twitter accounts.
411
00:22:21,600 --> 00:22:24,500
We followed that was the first
version and then we noticed a
412
00:22:24,508 --> 00:22:28,400
lot of those links once I click
on them, as human, I see a pay
413
00:22:28,400 --> 00:22:30,900
wall.
So we started adding a blacklist
414
00:22:30,900 --> 00:22:33,400
to the to love.
Void this domain, avoid this
415
00:22:33,400 --> 00:22:36,700
domain avoid this domain and we
try to maintain it.
416
00:22:36,700 --> 00:22:39,000
So that we avoid every domain
with a paywall.
417
00:22:39,800 --> 00:22:43,700
What I should mention is that
there is so much content out
418
00:22:43,700 --> 00:22:47,800
there that if you're avoiding
every website with a pay wall,
419
00:22:48,200 --> 00:22:51,700
you're not missing any news.
Maybe you're only reading five
420
00:22:51,700 --> 00:22:55,000
versions of this news article
instead of 20, right?
421
00:22:55,000 --> 00:22:58,500
But you're not missing out on
any news of any event out there
422
00:22:58,900 --> 00:23:02,000
because so many outlets right
about there.
423
00:23:02,200 --> 00:23:06,100
I think of that, right?
And so, probably the closest
424
00:23:06,100 --> 00:23:09,800
that we come to having a
difficulty as business news
425
00:23:10,000 --> 00:23:13,200
because just so many of them are
gate, right?
426
00:23:13,400 --> 00:23:16,400
But even there CNBC is not
gated, right?
427
00:23:17,100 --> 00:23:22,400
The mainstream networks, ABC,
NBC CBS PBS, they're not gated
428
00:23:22,400 --> 00:23:24,500
them.
They have a business section,
429
00:23:24,500 --> 00:23:26,700
right?
And so there is enough versions
430
00:23:26,700 --> 00:23:30,900
of even business articles to
just get all the news you want
431
00:23:30,900 --> 00:23:33,500
without ever going.
Into a gate that resource.
432
00:23:34,000 --> 00:23:36,100
Yeah, I think same applies to
take as well.
433
00:23:36,100 --> 00:23:40,300
So but one it's like a comment.
I want to add here.
434
00:23:41,600 --> 00:23:44,600
I understand, you know, media
companies they need to make
435
00:23:44,600 --> 00:23:49,300
money, right.
But putting this strong
436
00:23:49,300 --> 00:23:55,700
headlines on a website with ads
and being gated is something I
437
00:23:55,700 --> 00:23:59,300
cannot understand.
Honestly, speaking like II get
438
00:23:59,500 --> 00:24:01,800
too much, you know like
sometimes.
439
00:24:02,900 --> 00:24:04,900
Why?
Okay, I you want to make it get
440
00:24:04,900 --> 00:24:07,600
it, I understand.
But why you are pushing to me
441
00:24:07,600 --> 00:24:10,600
ads and still I cannot access
the content.
442
00:24:10,600 --> 00:24:13,600
So I hope you can do something
for this.
443
00:24:14,800 --> 00:24:19,600
Just like do two things before
we finish works with you today.
444
00:24:19,800 --> 00:24:26,400
Now again like you're using I
and you know when we talk about
445
00:24:26,400 --> 00:24:28,900
either the ethical aspect of it,
right?
446
00:24:28,900 --> 00:24:34,800
So in general, how do you think
We can keep, you know what, we
447
00:24:34,800 --> 00:24:40,200
call it the ethical
responsibility on in any product
448
00:24:40,200 --> 00:24:43,800
that leverage a i specially in
something that it's touching,
449
00:24:43,800 --> 00:24:48,300
you know, trying to filter the
content to the, to the end user.
450
00:24:49,100 --> 00:24:52,500
So, in our case, I think it's
actually, the answer is pretty
451
00:24:52,500 --> 00:24:57,000
simple, because we are using
only supervised machine
452
00:24:57,000 --> 00:25:00,200
learning, right?
If we just open the model and we
453
00:25:00,200 --> 00:25:04,300
open the data set, Everybody can
tell us if we're being
454
00:25:04,300 --> 00:25:06,700
unethical, right?
It's pretty obvious.
455
00:25:06,700 --> 00:25:10,000
What our model is learning to do
based on these two things.
456
00:25:10,600 --> 00:25:15,700
And so, our answer to this is
transparency, we have to show it
457
00:25:15,700 --> 00:25:17,900
to the world and they can keep
us honest.
458
00:25:17,900 --> 00:25:22,300
Like, we want to be unable to do
anything bad, just because if we
459
00:25:22,300 --> 00:25:24,500
do anything bad, then somebody
will catch it.
460
00:25:24,600 --> 00:25:26,600
It's the same.
I think in cybersecurity you
461
00:25:26,600 --> 00:25:28,500
have the same thing.
The safest tools are the open
462
00:25:28,500 --> 00:25:30,600
source ones.
Because then, if there's any
463
00:25:30,600 --> 00:25:32,400
backdoor that somebody built,
Attend to it.
464
00:25:34,700 --> 00:25:37,300
Then it's open source.
Somebody will notice that back
465
00:25:37,300 --> 00:25:40,000
door, right?
That's why there was this big
466
00:25:40,000 --> 00:25:43,800
hack of LastPass recently, and
all my friends are in last past.
467
00:25:43,800 --> 00:25:46,900
I'm using B Warden, right?
Because I don't trust last past.
468
00:25:46,900 --> 00:25:50,500
The source is closed.
So I'm only using an open source
469
00:25:50,500 --> 00:25:52,100
tool when it comes to something
like this.
470
00:25:52,400 --> 00:25:54,000
So that's kind of our approach
to it.
471
00:25:54,000 --> 00:25:57,500
Now granted, I should put a
caveat, we made our code Source
472
00:25:57,500 --> 00:25:59,900
available, not open source, so
people can look at it.
473
00:25:59,900 --> 00:26:03,100
They cannot copy it and reuse
That's kind of a business
474
00:26:03,100 --> 00:26:04,900
decision to be somewhere in the
middle between those.
475
00:26:05,200 --> 00:26:09,500
Now there's a bigger problem
when you get to Giant
476
00:26:09,500 --> 00:26:12,400
unsupervised language models
that just train themselves on
477
00:26:12,400 --> 00:26:17,500
the entire internet, right?
What GPT for is, I don't think
478
00:26:17,900 --> 00:26:21,400
that we can police their ethics.
They have the police themselves
479
00:26:21,400 --> 00:26:24,100
because we have no way of
knowing what they did, in some
480
00:26:24,100 --> 00:26:27,000
sense, by the time the end
product comes to us.
481
00:26:27,300 --> 00:26:30,400
All we can do is kind of poke it
a little bit from the sides, but
482
00:26:30,400 --> 00:26:34,400
we don't actually You know,
what's underneath and if it was
483
00:26:34,400 --> 00:26:37,800
mostly trained on Neo Nazi
propaganda, we might not know,
484
00:26:38,200 --> 00:26:40,000
right?
So it's up to them.
485
00:26:40,000 --> 00:26:44,600
And here, I hope that we can
develop some relatively standard
486
00:26:44,600 --> 00:26:49,400
models of Ethics, but I'm also a
little bit skeptical because I
487
00:26:49,400 --> 00:26:52,400
don't think we have those for
humans, right?
488
00:26:52,400 --> 00:26:55,900
The ethical models that we have
for human behavior are pretty
489
00:26:55,900 --> 00:26:59,100
bad.
If you think about it, in fact,
490
00:26:59,700 --> 00:27:01,500
they're kind of different in
each sphere.
491
00:27:01,900 --> 00:27:03,200
Right?
So if you look at the legal
492
00:27:03,200 --> 00:27:06,800
sphere, what we have is purely
deontic ethics, right?
493
00:27:06,900 --> 00:27:10,000
Somebody wrote the law.
Even if that law is outdated and
494
00:27:10,000 --> 00:27:13,500
let's say the law in the US on
eavesdropping was written. 1973
495
00:27:14,000 --> 00:27:16,600
we still follow that law as
written, even though it was
496
00:27:16,600 --> 00:27:21,600
written 1973, right?
So it's It's inefficient but
497
00:27:21,600 --> 00:27:24,300
that's what we have right.
Should we have that for AI
498
00:27:24,300 --> 00:27:27,300
obviously, that model with
failed because the I just moves
499
00:27:27,300 --> 00:27:30,300
too quickly, right?
Then once you talk to the
500
00:27:30,300 --> 00:27:34,100
entrepreneurs themselves, a lot
of them follow some version of
501
00:27:34,100 --> 00:27:38,000
utilitarian ethics, right?
Basically, if the result is
502
00:27:38,000 --> 00:27:41,000
good, then the action was good
and that's pretty stupid because
503
00:27:41,000 --> 00:27:43,000
when you make the action, you
cannot predict the results.
504
00:27:43,000 --> 00:27:48,600
So how is that an ethical
standard for make for choosing
505
00:27:48,600 --> 00:27:51,800
good actions, right?
So, I don't know.
506
00:27:52,200 --> 00:27:56,200
I hope that we can come to some
version of Greek virtue ethics
507
00:27:56,200 --> 00:27:58,000
when it comes to this sort of
thing.
508
00:27:58,400 --> 00:28:01,400
But I'm not holding my breath,
because most humans don't follow
509
00:28:01,400 --> 00:28:03,600
that either.
So, why would a i models follow?
510
00:28:03,800 --> 00:28:06,100
Yeah, it would be very
interesting to watch this space,
511
00:28:06,100 --> 00:28:10,300
but yeah, I like the analogy
that you did because sometimes,
512
00:28:10,300 --> 00:28:14,600
you know, I used to know to, to
go and do consultancies in
513
00:28:14,600 --> 00:28:17,300
different fields with customers
and then, you know, we touched
514
00:28:17,300 --> 00:28:20,800
on the compliance, for example,
and you find out on Policies.
515
00:28:20,800 --> 00:28:24,200
And you find out that someone
wrote that document, when he
516
00:28:24,200 --> 00:28:28,100
eats 25 his back and okay, I
found it like this, I don't
517
00:28:28,100 --> 00:28:30,200
know, you know.
So I knew but with a is
518
00:28:30,200 --> 00:28:33,800
dangerous because as you said
it's changing on daily basis
519
00:28:33,800 --> 00:28:40,000
almost now they're talking about
GPT 5 and people are asking
520
00:28:40,000 --> 00:28:40,700
them.
No, no, no.
521
00:28:40,700 --> 00:28:43,600
Don't release it yet.
We still didn't digest that one.
522
00:28:44,600 --> 00:28:47,500
But anyway, we'll see.
Finally, you know, question for
523
00:28:47,500 --> 00:28:50,500
you, Alex?
What do you think the A future
524
00:28:50,500 --> 00:28:54,100
of information.
Ecosystem overall gonna be in
525
00:28:54,100 --> 00:28:57,200
the coming years and content
creation, and all the space.
526
00:28:57,900 --> 00:29:00,700
So, I think you have to consider
that a, I will make.
527
00:29:01,700 --> 00:29:04,100
First of all, it will make
content creation, at least the
528
00:29:04,100 --> 00:29:06,500
low quality content creation,
much easier.
529
00:29:06,800 --> 00:29:09,600
So, you will see an even bigger
explosion that you've seen.
530
00:29:09,600 --> 00:29:13,600
Then you've seen in the past 10
years and bad content being
531
00:29:13,600 --> 00:29:16,500
generated.
We will be drowning in jump at
532
00:29:16,500 --> 00:29:19,200
the same time.
One of the other interesting
533
00:29:19,200 --> 00:29:24,400
effects that Ai and specifically
the GPT style generative.
534
00:29:24,400 --> 00:29:29,900
A I will have is that I think
search with search results on
535
00:29:29,900 --> 00:29:33,600
the page will probably go away
as a tool because you will now
536
00:29:33,600 --> 00:29:36,300
have search engines that can
just answer your question.
537
00:29:36,700 --> 00:29:40,600
Give you one answer that was
generated instead of here, are
538
00:29:40,600 --> 00:29:45,000
55,000 answers to choose from.
And that means that most of
539
00:29:45,000 --> 00:29:47,500
these websites, that generate,
the junk won't get much less
540
00:29:47,500 --> 00:29:50,500
traffic from search engines.
So combine these Two things
541
00:29:50,500 --> 00:29:54,300
together, junk is, cheaper to
generate, but harder to
542
00:29:54,300 --> 00:29:57,700
monetize, because less money is
coming in from search engines,
543
00:29:58,200 --> 00:30:00,100
right?
And it's an entire race to the
544
00:30:00,100 --> 00:30:03,000
bottom, where everybody will try
to create more and more junk to
545
00:30:03,008 --> 00:30:06,300
fight for the last scraps of
traffic that are available.
546
00:30:06,800 --> 00:30:10,100
Basically we have to create
filters otherwise we will draw.
547
00:30:10,500 --> 00:30:13,100
So I don't know if the filter is
that we are making are the right
548
00:30:13,100 --> 00:30:16,600
ones, maybe somebody else will
make better ones but whatever it
549
00:30:16,600 --> 00:30:20,000
is, we're going to need filters
because the junk is coming.
550
00:30:20,400 --> 00:30:24,300
If it's going to be like a giant
Avalanche of elephants blowing
551
00:30:24,300 --> 00:30:27,100
bubbles, the is an example that
we just touched.
552
00:30:27,100 --> 00:30:32,500
Yeah, so that's kind of my dire
side of the production but I'll
553
00:30:32,500 --> 00:30:35,200
give you a hopeful note as well.
If you look at the closest
554
00:30:35,200 --> 00:30:38,200
historic, parallel to what we're
seeing now, right?
555
00:30:38,200 --> 00:30:43,300
The printing, press was invented
in the 1430s, we saw 200 years
556
00:30:43,300 --> 00:30:47,200
of inquisitions, witch hunts 52
religious wars, all over Europe.
557
00:30:47,300 --> 00:30:50,000
People were drowning and new
junk being published.
558
00:30:50,200 --> 00:30:53,400
And that they just couldn't
digest or process right?
559
00:30:53,600 --> 00:30:58,500
But then what came out of it is
the scientific method and 1635 I
560
00:30:58,508 --> 00:31:03,000
think peer review in 1665 right?
And then the Enlightenment and
561
00:31:03,000 --> 00:31:05,900
then Universal literacy all over
Europe.
562
00:31:06,000 --> 00:31:08,500
Right.
And So eventually the outcome is
563
00:31:08,500 --> 00:31:10,500
good.
We just need to learn how to get
564
00:31:10,500 --> 00:31:12,100
there without blowing ourselves
up.
565
00:31:13,100 --> 00:31:15,600
Yeah, cool.
And just as a final note on
566
00:31:15,600 --> 00:31:20,800
this, I liked you you know
positivity on That's because
567
00:31:20,800 --> 00:31:23,900
this is what I'm telling people
who usually are showing like
568
00:31:23,900 --> 00:31:25,200
little bit.
They are scared.
569
00:31:25,500 --> 00:31:27,200
I'm telling them.
No.
570
00:31:27,300 --> 00:31:29,900
Okay.
It might be looking scary now
571
00:31:29,900 --> 00:31:34,400
but if you think on the long run
I believe yeah as you said it
572
00:31:34,400 --> 00:31:37,300
might be a moment of
Enlightenment for Humanity.
573
00:31:37,500 --> 00:31:41,200
We need to be careful of course.
Well elect thank you very much
574
00:31:41,200 --> 00:31:46,600
for your time today.
It was I think a very like Rich
575
00:31:46,600 --> 00:31:51,100
conversation with you on this
topic Thank you, everyone for
576
00:31:51,100 --> 00:31:55,100
tuning in with us today.
Thank you for watching or if you
577
00:31:55,100 --> 00:31:58,100
are listening, as usual.
If you have any comment, if you
578
00:31:58,100 --> 00:32:01,400
have any question about this
episode or future, you know,
579
00:32:01,400 --> 00:32:04,600
topics you want to bring in,
please reach out to me by email
580
00:32:05,000 --> 00:32:08,300
LinkedIn, or Twitter and until
we meet in the next episode.
581
00:32:08,300 --> 00:32:09,300
Thank you very much.
00:00:06,500 --> 00:00:09,400
Hello, and welcome to a new
episode of the CTO show with
2
00:00:09,400 --> 00:00:10,800
Mohammed.
My name is Mohammed.
3
00:00:11,200 --> 00:00:13,900
And as, you know, in each
episode, I share some
4
00:00:13,900 --> 00:00:17,800
information about the latest
trends inside strategies, and
5
00:00:17,800 --> 00:00:22,000
topic, like cybersecurity
digital transformation, emerging
6
00:00:22,000 --> 00:00:25,000
Tech, startups and
Entrepreneurship and sometime.
7
00:00:25,000 --> 00:00:28,300
I have thought leaders, and
innovators and entrepreneurs to
8
00:00:28,400 --> 00:00:33,400
let them share with us, you
know, their expertise in the
9
00:00:33,400 --> 00:00:35,700
industry.
They are in and how they are
10
00:00:35,700 --> 00:00:37,900
making.
Impact in the world of tech and
11
00:00:37,900 --> 00:00:40,000
business.
And today I'm very pleased to
12
00:00:40,000 --> 00:00:44,200
have with me on the show.
Alex, Alex is joining me from
13
00:00:44,200 --> 00:00:49,100
the US Alex.
He is the an entrepreneur for
14
00:00:49,100 --> 00:00:51,300
long time.
I will leave him to introduce
15
00:00:51,500 --> 00:00:55,400
himself and what he does.
Alright, so thank you so much
16
00:00:55,400 --> 00:00:57,200
for having me.
It's great to be on the show.
17
00:00:57,500 --> 00:01:00,600
My name is Alex Fink.
I'm the founder and CEO of the
18
00:01:00,600 --> 00:01:03,000
other web or other web ink,
right?
19
00:01:03,700 --> 00:01:07,800
We are a company that is
dedicated to Trying to fix the
20
00:01:07,800 --> 00:01:11,900
world's information ecosystem or
in other words, if we make it a
21
00:01:11,900 --> 00:01:14,400
little more practical, we want
to help people consume higher
22
00:01:14,400 --> 00:01:17,200
quality information.
Now, the way that works is first
23
00:01:17,200 --> 00:01:20,800
we've developed a set of filters
that evaluate information and
24
00:01:20,800 --> 00:01:22,700
figure out what is junk and
throw it out.
25
00:01:22,700 --> 00:01:26,100
And then we created an
information platform that
26
00:01:26,100 --> 00:01:31,300
Aggregates content from all over
the web news commentary podcasts
27
00:01:31,300 --> 00:01:34,900
research, studies, everything we
can get our Crawlers on and then
28
00:01:34,900 --> 00:01:37,400
we filter the junk out and we
Create a feed for people that
29
00:01:37,400 --> 00:01:40,700
they can customize so that they
decide what they consume instead
30
00:01:40,700 --> 00:01:42,600
of social media deciding for
them.
31
00:01:43,700 --> 00:01:46,900
Great great.
What was your opinion motive
32
00:01:46,900 --> 00:01:50,700
Alex to start this?
What how it all started?
33
00:01:51,200 --> 00:01:55,100
So it's been brewing in my mind
for a very long time, right?
34
00:01:55,100 --> 00:01:57,300
And I've been an information
junkie for a long time.
35
00:01:57,300 --> 00:02:01,800
So I've been watching this
ecosystem deteriorate basically
36
00:02:01,800 --> 00:02:04,800
throughout my life and almost
every single outlet.
37
00:02:04,800 --> 00:02:08,100
If you look at it today if it's
This the 20 years ago it was
38
00:02:08,100 --> 00:02:12,000
better 20 years ago, they are
all becoming worse and so I was
39
00:02:12,000 --> 00:02:14,600
trying to figure out why for a
long time and trying to figure
40
00:02:14,600 --> 00:02:18,100
out what incentives, drive them
towards all becoming clickbait
41
00:02:18,100 --> 00:02:21,500
centrally.
Now on a personal level, I've
42
00:02:21,500 --> 00:02:24,600
been building perception systems
for 15 years, so cameras,
43
00:02:24,600 --> 00:02:25,800
computer vision, that sort of
thing.
44
00:02:26,100 --> 00:02:29,500
And so, sometime last year, I
had this crisis of conscience of
45
00:02:30,200 --> 00:02:33,800
does the world need more cameras
and I decided it doesn't.
46
00:02:33,800 --> 00:02:36,000
And I need to do something more
useful with my life.
47
00:02:36,200 --> 00:02:39,500
To avoid having this disconnect
where I built cameras to make
48
00:02:39,500 --> 00:02:41,700
money.
But I keep thinking that the
49
00:02:41,708 --> 00:02:44,000
biggest problem is that people
are consuming junk.
50
00:02:44,200 --> 00:02:48,200
So, I decided to connect those
two and actually start devoting
51
00:02:48,200 --> 00:02:52,100
my time towards fixing this big
problem that I see in the world
52
00:02:52,100 --> 00:02:54,700
that everything around us is
becoming junk and that's what
53
00:02:54,700 --> 00:02:58,200
people can seal.
That's, I mean, it's very
54
00:02:59,500 --> 00:03:02,800
inspirational.
I would say, especially, you
55
00:03:02,800 --> 00:03:05,900
know, like everyone has, you
said they try to make money, but
56
00:03:05,900 --> 00:03:08,200
for you, It's like you go out on
a mission.
57
00:03:08,500 --> 00:03:12,600
Now, if I can, you know, ask you
this.
58
00:03:12,700 --> 00:03:15,700
So it's all about Miss
information, right?
59
00:03:15,700 --> 00:03:18,300
And all these things.
So, how do you think artificial
60
00:03:18,300 --> 00:03:22,300
intelligence can play a role in
fighting against?
61
00:03:22,300 --> 00:03:26,700
You know, this misinformation
propaganda and how you at other
62
00:03:26,700 --> 00:03:29,800
web you are leveraging this in
to filter the content.
63
00:03:29,900 --> 00:03:32,700
So first of all, let me say that
it can play a role in both
64
00:03:32,700 --> 00:03:35,300
directions, right?
And especially now that you're
65
00:03:35,300 --> 00:03:38,300
seeing these And of big advances
in generative AI.
66
00:03:38,700 --> 00:03:42,800
They can play a role in making
things worse, because if you
67
00:03:42,800 --> 00:03:45,500
think about what it took 10
years ago to write the clickbait
68
00:03:45,500 --> 00:03:49,400
article, you have to a be test
various headlines.
69
00:03:49,400 --> 00:03:53,300
You have to actually compose
reasonably coherent text.
70
00:03:54,400 --> 00:03:57,100
You don't have to do any of that
stuff anymore, judge if he can
71
00:03:57,100 --> 00:03:59,500
do that for you.
And so right now that same
72
00:03:59,500 --> 00:04:02,300
person generating Jung called a
can generate 10 times, the
73
00:04:02,308 --> 00:04:05,300
amount of junk that he used to
generate a year ago.
74
00:04:05,700 --> 00:04:09,500
So If left unchecked to its own
devices, this entire ecosystem
75
00:04:09,500 --> 00:04:13,300
will get worse because of a.
I now, the way we view AI is
76
00:04:13,300 --> 00:04:17,200
thinking, okay?
What is needed to clean?
77
00:04:17,200 --> 00:04:21,000
Things up, probably more
editors, and we need those
78
00:04:21,000 --> 00:04:24,400
editors to be more consistent.
We need them to never have a bad
79
00:04:24,400 --> 00:04:26,200
day.
We need them to have great
80
00:04:26,200 --> 00:04:29,200
throughput.
So essentially, we are training
81
00:04:29,200 --> 00:04:32,500
models that behave as newspaper
editors, right?
82
00:04:32,700 --> 00:04:36,000
So we take some particular
trait, like, checking the
83
00:04:36,200 --> 00:04:38,500
Headline, to make sure it
matches the body of the article.
84
00:04:38,700 --> 00:04:42,400
And we trained about a model
that does just that, right?
85
00:04:42,400 --> 00:04:45,300
And it does it almost as good as
a human editor, but it never has
86
00:04:45,300 --> 00:04:50,300
a bad day and never and it can
be completely transparent.
87
00:04:50,300 --> 00:04:54,500
So we also open the source of
all our models made them Source
88
00:04:54,500 --> 00:04:57,800
available because we want people
to be able to actually look at
89
00:04:57,800 --> 00:04:59,600
them and see.
There is nothing hidden in
90
00:04:59,600 --> 00:05:01,200
there.
We're not just randomly
91
00:05:01,200 --> 00:05:03,800
penalizing everybody who's right
wing or everybody who's left
92
00:05:03,800 --> 00:05:06,000
wing, we're just trying to train
it on.
93
00:05:06,300 --> 00:05:10,000
Academic data sets or data sets
that we scrambled together to
94
00:05:10,000 --> 00:05:12,500
make them as balanced and clear
as possible.
95
00:05:12,500 --> 00:05:18,000
So that's where a I can do some
good but ultimately it's a
96
00:05:18,000 --> 00:05:20,100
kitchen knife.
You can make a salad or you can
97
00:05:20,100 --> 00:05:25,100
stab someone it's up to you.
Yeah, so if if I understood
98
00:05:25,100 --> 00:05:30,400
right, so should we say that
your platform would get better
99
00:05:30,400 --> 00:05:33,200
with time because the you know,
it's a matter of fact, the
100
00:05:33,200 --> 00:05:35,900
better you train a model, the
battery.
101
00:05:36,100 --> 00:05:38,600
Becomes later.
So is this like what we should
102
00:05:38,600 --> 00:05:42,600
expect from from other web.
Absolutely, it gets better and
103
00:05:42,600 --> 00:05:45,600
time because we gather more data
and we feed that back and
104
00:05:45,600 --> 00:05:48,100
retrain the model.
But also because we get feedback
105
00:05:48,100 --> 00:05:51,000
from our users.
So the initial set of models we
106
00:05:51,000 --> 00:05:56,100
did our entire feedback was we
showed some examples of the
107
00:05:56,100 --> 00:05:58,900
results to 5 friends that I
have.
108
00:05:58,900 --> 00:06:01,600
And they told me whether it kind
of conforms to their Intuition
109
00:06:01,600 --> 00:06:05,000
or not.
So that's a very low bar for
110
00:06:05,100 --> 00:06:07,900
like, the last I was just
testing your model against
111
00:06:07,900 --> 00:06:10,100
reality, right?
We certainly didn't have the
112
00:06:10,100 --> 00:06:13,600
budget that GPT forehead now
where they had thousands of
113
00:06:13,600 --> 00:06:16,700
people working for nine months
if I understood correctly.
114
00:06:16,700 --> 00:06:20,100
Right?
Tragic, tragic jet GPT for was
115
00:06:20,100 --> 00:06:23,300
done in summer.
They released it now right?
116
00:06:23,300 --> 00:06:26,700
So this entire time they've been
doing human feedback essential
117
00:06:27,200 --> 00:06:30,400
so now we have natural human
feedback, we have a quarter
118
00:06:30,400 --> 00:06:33,600
million users that actually read
the news and they flag stuff
119
00:06:33,600 --> 00:06:36,000
that seems off right?
That's it.
120
00:06:36,100 --> 00:06:38,300
Big piece of feedback for us,
right?
121
00:06:38,300 --> 00:06:43,100
We can also contact them and ask
them and now we also rolled out
122
00:06:43,500 --> 00:06:47,700
a version of the app that
essentially forces, the users to
123
00:06:48,300 --> 00:06:50,700
interact with content to certain
way, right?
124
00:06:50,700 --> 00:06:53,100
They have to swipe an item to
get to the next one.
125
00:06:53,300 --> 00:06:56,400
And the direction of their swipe
is feedback for us about how
126
00:06:56,400 --> 00:06:59,300
good this item was.
And so over time, we will learn
127
00:06:59,300 --> 00:07:03,100
to incorporate that real human
feedback as well and to making
128
00:07:03,100 --> 00:07:05,600
our models better.
And I'm talking here about the
129
00:07:05,600 --> 00:07:08,400
filtering Girls, I'm talking
about the Sorting of the feed
130
00:07:08,800 --> 00:07:11,600
and I'm also talking about
things like summarizes because
131
00:07:11,600 --> 00:07:15,500
we show users a summary of the
article that we create instead
132
00:07:15,500 --> 00:07:17,600
of the original excerpt from the
article.
133
00:07:18,400 --> 00:07:22,500
Yeah.
Like maybe this is should be
134
00:07:22,500 --> 00:07:27,300
also related but like you
started this so people can get
135
00:07:27,500 --> 00:07:33,100
as you said like more clean and
remove the junk from from their
136
00:07:33,100 --> 00:07:36,000
feet but it will like what do
you think?
137
00:07:36,100 --> 00:07:39,500
Think, you know, people should
really care about especially in
138
00:07:39,500 --> 00:07:43,400
this, you know, we are living
now in a, in a world where you
139
00:07:43,408 --> 00:07:45,600
have a lot of things coming in,
something comes from the
140
00:07:45,600 --> 00:07:48,800
mainstream media.
Something comes from digital
141
00:07:49,000 --> 00:07:51,200
media.
Something comes from even, you
142
00:07:51,200 --> 00:07:54,800
know, companies where they work.
So what do you think, you know,
143
00:07:54,800 --> 00:08:00,800
they can do to navigate better
this complex landscape of online
144
00:08:00,800 --> 00:08:03,600
information?
So, I think the best analogy to
145
00:08:03,600 --> 00:08:07,200
keep in mind is the way we Food,
right?
146
00:08:07,900 --> 00:08:09,600
Information is food for the
brain.
147
00:08:09,600 --> 00:08:13,500
So the way that we look at
improving your normal diet is
148
00:08:13,500 --> 00:08:16,000
probably the way that we should
look at improving our digital
149
00:08:16,000 --> 00:08:18,100
day.
So when it comes to normal food,
150
00:08:18,100 --> 00:08:20,300
if you go to the doctor, doctor
tells you you're unhealthy, you
151
00:08:20,300 --> 00:08:22,500
need to eat better food, you
come back home?
152
00:08:22,500 --> 00:08:25,200
What's your next step?
The first step is probably going
153
00:08:25,200 --> 00:08:28,600
to be to cut out the things that
are obviously Just Junk.
154
00:08:28,600 --> 00:08:30,600
They're terrible, write the
empty calories.
155
00:08:30,900 --> 00:08:33,299
I don't put sugar in what your
drink.
156
00:08:33,500 --> 00:08:35,900
Don't eat cookies all day,
remove the obvious.
157
00:08:36,100 --> 00:08:40,000
We bad stuff, that's step one.
So we try to help with that.
158
00:08:40,100 --> 00:08:43,799
But obviously users still need
to pay some attention to other
159
00:08:43,799 --> 00:08:45,700
sources of empty calories that
they have.
160
00:08:45,800 --> 00:08:48,400
Now, once you do that, your job
is not over.
161
00:08:48,700 --> 00:08:51,600
You still have to figure out how
to balance the things that you
162
00:08:51,600 --> 00:08:56,200
do it in a reasonable way.
And I've seen some pretty crazy
163
00:08:56,200 --> 00:08:59,400
studies about what happens when
you overdose on one type of
164
00:08:59,400 --> 00:09:02,900
content.
There's a study from 2013 where
165
00:09:02,900 --> 00:09:07,800
researchers compared people who
were at the Marathon bombing and
166
00:09:07,800 --> 00:09:11,100
actually witnessed the bombing
to people who watched six hours
167
00:09:11,100 --> 00:09:15,300
of news coverage about the
bombing the group that watch the
168
00:09:15,300 --> 00:09:19,100
news coverage had higher chances
of PTSD than the group that what
169
00:09:19,100 --> 00:09:23,400
that was in the event.
So news is obviously good.
170
00:09:23,400 --> 00:09:25,100
You want to know what's going on
in the world?
171
00:09:25,400 --> 00:09:29,500
But 15 minutes is good 6 hours
on a single bombing is probably
172
00:09:29,500 --> 00:09:32,800
really bad, right?
And so you have to learn to
173
00:09:32,800 --> 00:09:35,800
balance what you have and
there's many different sources
174
00:09:35,800 --> 00:09:36,500
of information.
Information.
175
00:09:36,500 --> 00:09:39,700
We could consume could be Dance
videos on Tick.
176
00:09:39,700 --> 00:09:44,000
Tock could be serious.
News coverage could be books
177
00:09:44,000 --> 00:09:47,200
written by somebody. 2,000 years
ago about philosophy, right?
178
00:09:47,200 --> 00:09:50,500
You want to balance all of these
out in some reasonable way?
179
00:09:51,500 --> 00:09:53,300
What is it?
That depends on you?
180
00:09:53,300 --> 00:09:55,000
There's probably different
preferences.
181
00:09:55,000 --> 00:09:58,000
Just like with diet, see of the
keto people, and the vegan
182
00:09:58,000 --> 00:10:01,400
people in the Mediterranean, all
of these right there, all better
183
00:10:01,400 --> 00:10:03,800
than just mindlessly consuming.
What's in front of you?
184
00:10:04,000 --> 00:10:07,600
So I think it's the same thing
with You should just sit down
185
00:10:07,800 --> 00:10:10,400
maybe track a little bit.
What you're consuming right now.
186
00:10:10,600 --> 00:10:12,400
Yeah.
And try to determine what you
187
00:10:12,400 --> 00:10:14,500
want to do.
I want to mention one more thing
188
00:10:14,500 --> 00:10:20,200
with regards to how we handle it
and the other web and that the
189
00:10:20,200 --> 00:10:23,500
vast majority of platforms out
there, try to maximize your time
190
00:10:23,500 --> 00:10:25,900
on site because then they can
show you more ads.
191
00:10:25,900 --> 00:10:28,700
That's their incentive.
But if you think about your
192
00:10:28,700 --> 00:10:33,500
incentive as a human, you don't
want to maximize any single food
193
00:10:33,500 --> 00:10:35,900
group.
Why would you want to spend?
194
00:10:36,000 --> 00:10:38,800
And too much time on any
particular site, you want to get
195
00:10:38,800 --> 00:10:42,700
as much value from this site in
as little time as possible,
196
00:10:42,700 --> 00:10:44,600
right?
You don't want to spend as much
197
00:10:44,600 --> 00:10:48,400
time as possible on Technical
and so, but the average person
198
00:10:48,400 --> 00:10:51,700
spends an hour and a half, a day
of data, which is pretty bad if
199
00:10:51,700 --> 00:10:54,300
you think about it.
So our approach to it, as we
200
00:10:54,300 --> 00:10:56,500
want to help you to get as much
value as you can.
201
00:10:56,500 --> 00:11:00,000
And then stop, there is no
reason for you to just continue
202
00:11:00,000 --> 00:11:01,400
consuming more and more and
more.
203
00:11:01,900 --> 00:11:05,400
We think that if you tell us you
want to read 50 summaries of
204
00:11:05,400 --> 00:11:09,400
articles, We're going to stop
you at 50 because you don't want
205
00:11:09,400 --> 00:11:11,600
to read more 50 is your quarter,
right?
206
00:11:12,000 --> 00:11:14,300
It's counterintuitive.
It goes proudly against
207
00:11:14,300 --> 00:11:17,700
Financial incentives.
It's why we had to register as a
208
00:11:17,700 --> 00:11:20,900
public benefit Corporation
because I think as a c chord, we
209
00:11:20,900 --> 00:11:23,000
would not be able to do this
very long, right.
210
00:11:23,000 --> 00:11:27,900
Our investors would force us to
maximize time on site like
211
00:11:27,900 --> 00:11:31,300
everybody else does.
So how do you monetize just out
212
00:11:31,300 --> 00:11:33,500
of curiosity?
We haven't started yet.
213
00:11:33,900 --> 00:11:37,700
If we have some idea, Ideas on
what we want to experiment with
214
00:11:37,900 --> 00:11:39,600
but we want to start those
experiments.
215
00:11:39,600 --> 00:11:42,400
When there is enough users that
we can actually see what value
216
00:11:42,400 --> 00:11:45,400
they get out of the platform and
how we can monetize without
217
00:11:45,400 --> 00:11:48,600
interfering with that.
One of the interesting
218
00:11:49,000 --> 00:11:55,400
hypotheses we have is that we
think when users consume content
219
00:11:55,400 --> 00:11:59,600
and kind of a passive way like a
feed, eventually they notice
220
00:11:59,600 --> 00:12:00,800
something that Sparks their
interest.
221
00:12:00,800 --> 00:12:03,500
Like, you see a phrase in an
article that you want to know
222
00:12:03,500 --> 00:12:08,000
more about if this happens in a
And then you typically just ask
223
00:12:08,000 --> 00:12:11,200
what's that, right?
But if this happens online, we
224
00:12:11,200 --> 00:12:13,700
have this weird pattern where
people open the browser window
225
00:12:13,700 --> 00:12:16,700
type something into the address
bar, click enter select the
226
00:12:16,700 --> 00:12:19,500
result and then what they see is
usually unrelated to what they
227
00:12:19,500 --> 00:12:21,700
searched because the entire
complex was lost and all these
228
00:12:21,700 --> 00:12:25,200
transitions, right?
So we implemented a what's that
229
00:12:25,200 --> 00:12:27,600
into the platform?
And we see that the average
230
00:12:27,600 --> 00:12:30,900
newsreader on our platform uses
that two or three times per
231
00:12:30,900 --> 00:12:35,300
session and so and when they do,
we show them something that
232
00:12:35,300 --> 00:12:39,100
looks a lot like Search results.
And we know pretty well how to
233
00:12:39,100 --> 00:12:42,800
put ads on search results in a
way that doesn't interfere with
234
00:12:42,800 --> 00:12:46,300
the users experience.
So our hope, and our hypothesis
235
00:12:46,300 --> 00:12:48,700
right now is ads on search
results are enough.
236
00:12:49,100 --> 00:12:53,000
And we never need to put ads on
the feed because that's annoying
237
00:12:53,000 --> 00:12:56,600
and we don't want to do that.
But ultimately, we have to test
238
00:12:56,600 --> 00:12:59,800
and see how users react and make
sure users are happy.
239
00:13:00,100 --> 00:13:01,700
Yeah, that's really interesting
model.
240
00:13:01,700 --> 00:13:05,300
I would say now.
My understanding.
241
00:13:05,300 --> 00:13:10,200
And what we discussed is you
know, like having this garbage
242
00:13:10,200 --> 00:13:13,700
cleaning I would say of read
content.
243
00:13:13,700 --> 00:13:19,500
Right now, something I brought
many times on, on the, on the
244
00:13:19,500 --> 00:13:23,100
show is about the fake where I
talked myself, you know, about,
245
00:13:23,200 --> 00:13:26,500
you know, the fake and what's
happening and even II tried to
246
00:13:26,500 --> 00:13:31,600
do a few want to call it a proof
of concept where I used multiple
247
00:13:31,700 --> 00:13:35,500
AI tools and, you know, I Myself
a video, of course it was like
248
00:13:35,500 --> 00:13:39,700
still a better version of that
software, but we've seen
249
00:13:39,700 --> 00:13:43,900
recently with the generative AI
really thinks that are scary.
250
00:13:45,100 --> 00:13:48,500
How do you think this can be
treated by mean from the
251
00:13:48,500 --> 00:13:50,600
mission?
You are in now because you're
252
00:13:50,600 --> 00:13:53,700
trying to remove right?
Like faking use you're trying to
253
00:13:53,700 --> 00:13:55,600
to clean something, which
nonsense.
254
00:13:56,000 --> 00:13:58,200
How do you think you can tackle
this?
255
00:13:59,100 --> 00:14:02,600
So it's an interesting problem
and there will be a sort of an
256
00:14:02,600 --> 00:14:05,900
arms race between the fakers and
the guys who try to spot the
257
00:14:05,900 --> 00:14:08,800
fakers, right?
The way that I would view this,
258
00:14:08,800 --> 00:14:11,700
as first of all, if you look at
all the different types of bad
259
00:14:11,700 --> 00:14:18,300
content out there content that
is faked on purpose with effort
260
00:14:18,300 --> 00:14:22,400
from the faker is actually the
minority most of the bad content
261
00:14:22,400 --> 00:14:26,300
out there is not fake news.
It's just not news at all,
262
00:14:26,800 --> 00:14:27,700
right?
It's CNN.
263
00:14:27,700 --> 00:14:30,600
Publishing an article with a
title like stop what you're
264
00:14:30,600 --> 00:14:33,300
doing and watch this elephant
play with bubbles, that's an
265
00:14:33,300 --> 00:14:39,000
actual CNN article, right?
So, I think that to actually
266
00:14:39,000 --> 00:14:43,600
focus on Thinking, critically
about content and trying to
267
00:14:43,600 --> 00:14:47,500
notice the fakes we first have
to get rid of the obviously
268
00:14:47,500 --> 00:14:49,900
about stuff.
The stuff that is bad, just
269
00:14:49,900 --> 00:14:53,000
based on form without even
looking at the substance, and
270
00:14:53,000 --> 00:14:55,300
that is the majority.
So once you clean all of that
271
00:14:55,300 --> 00:15:00,600
out, you will be left with the
more stubborn cases of Russian
272
00:15:00,600 --> 00:15:06,000
propaganda or of the Democrats,
trying to create actual fake
273
00:15:06,000 --> 00:15:08,600
stories about Republicans or
Republicans trying to create
274
00:15:08,600 --> 00:15:10,100
actual fake stories about
Democrats.
275
00:15:10,100 --> 00:15:13,900
Those will be Harder to spot and
I don't think you can ever clean
276
00:15:13,900 --> 00:15:17,400
it out completely.
So even if you look at science,
277
00:15:17,700 --> 00:15:19,400
right, you look at scientific
peer-reviewed.
278
00:15:19,400 --> 00:15:24,000
Publishing the last figure I saw
is that in almost every decade,
279
00:15:24,000 --> 00:15:28,500
this figure is kind of fixed the
number of completely fake
280
00:15:29,600 --> 00:15:34,100
experiments published or papers
published was about 14%, right?
281
00:15:34,100 --> 00:15:35,800
So, fourteen percent of
everything.
282
00:15:35,800 --> 00:15:39,000
You see that this published in
peer-reviewed journals is
283
00:15:39,000 --> 00:15:41,000
usually a paper that describes
an experiment.
284
00:15:41,000 --> 00:15:43,900
That They're didn't happen or
the result of the experiment is
285
00:15:43,900 --> 00:15:47,800
not what the paper says it.
Right, it's hard to spot.
286
00:15:48,500 --> 00:15:51,400
You can't ever, if you're a peer
reviewer, you cannot just drive
287
00:15:51,400 --> 00:15:53,900
to somebody's lab and start
counting the mice, right?
288
00:15:54,300 --> 00:15:56,600
You have to believe them, when
they tell you, I started with
289
00:15:56,600 --> 00:15:59,000
1713 survive.
Therefore, this is the death
290
00:15:59,000 --> 00:16:02,100
rate, right?
So, typically peer reviewers,
291
00:16:02,100 --> 00:16:05,100
focus on foreign they don't
focus on going to the lab and
292
00:16:05,100 --> 00:16:07,700
double-checking the number of
miles they just see, okay,
293
00:16:07,700 --> 00:16:11,700
structured correctly, described
correctly headline, You subtract
294
00:16:11,700 --> 00:16:13,500
matches body?
Correct references.
295
00:16:13,600 --> 00:16:14,800
All right.
Clear to publish.
296
00:16:17,400 --> 00:16:19,100
Detecting, whether something was
fake.
297
00:16:19,100 --> 00:16:22,200
All along, might take years, you
will not find it in real time.
298
00:16:22,700 --> 00:16:26,000
So let's start from the easy
problems, then the intermediate
299
00:16:26,000 --> 00:16:29,300
problems and maybe there will be
some number of intractable
300
00:16:29,300 --> 00:16:32,200
problems that we cannot fix.
Now, when you're talking about
301
00:16:32,200 --> 00:16:36,900
deep deep Faith videos, that's
an interesting conundrum in that
302
00:16:37,900 --> 00:16:40,900
You will see those people will
create those videos, then a
303
00:16:40,900 --> 00:16:43,900
journalist will write about it
and then we will take it up in
304
00:16:43,900 --> 00:16:48,600
our aggregator.
So to some extent, the best we
305
00:16:48,600 --> 00:16:50,700
can do is cross-reference
between the journals.
306
00:16:50,700 --> 00:16:54,200
Because again, we cannot be the
guys who verified the video
307
00:16:54,200 --> 00:16:57,700
itself.
We are writing about the person
308
00:16:57,700 --> 00:17:01,400
who wrote about the video in
some sense and so the best we
309
00:17:01,400 --> 00:17:03,300
can do is come up with some
heuristics on.
310
00:17:03,300 --> 00:17:08,599
How do you cross reference Maybe
if left-wing Outlets publish
311
00:17:08,599 --> 00:17:11,099
something, right?
Wing Outlets say that it's fake,
312
00:17:11,900 --> 00:17:15,300
then it's only half true until
both sides, agree on it, right?
313
00:17:16,500 --> 00:17:18,900
That's actually a heuristic.
I use for myself.
314
00:17:19,000 --> 00:17:22,400
Generally speaking things that
are obviously true typically
315
00:17:22,400 --> 00:17:25,700
things that people who disagree
on everything still agree on,
316
00:17:26,300 --> 00:17:28,300
right?
Otherwise, if one side says it's
317
00:17:28,300 --> 00:17:30,000
true in the other one says it's
not.
318
00:17:30,000 --> 00:17:32,200
It's probably only half true it
best.
319
00:17:32,800 --> 00:17:36,400
So I don't know, I don't think
it's entirely solvable.
320
00:17:36,400 --> 00:17:39,600
We will continue to develop
better tools, the fakers will
321
00:17:39,600 --> 00:17:42,600
continue to develop better tools
as well, so it will be an arms
322
00:17:42,600 --> 00:17:44,600
race.
Yeah, it's similar to, you know,
323
00:17:44,600 --> 00:17:47,500
because I cover cyber security
as well at the same case as
324
00:17:47,500 --> 00:17:49,500
right?
So always you have the
325
00:17:50,400 --> 00:17:53,900
researchers trying to, you know,
close the gap and then you see
326
00:17:53,900 --> 00:17:56,900
the other side, the bad guys.
They are also using, you know,
327
00:17:56,900 --> 00:18:00,200
latest Technologies and so it's
like an air is that it's
328
00:18:00,200 --> 00:18:03,300
never-ending.
But one thing which is also
329
00:18:03,300 --> 00:18:07,200
like, Very interesting in what
you're trying to do.
330
00:18:07,400 --> 00:18:12,500
So usually you know you try to
keep people and you said this,
331
00:18:12,500 --> 00:18:15,900
you try to keep people on your
platform as much as possible and
332
00:18:15,900 --> 00:18:18,400
now you're doing the complete
opposite.
333
00:18:18,500 --> 00:18:22,200
I mean in a sense that you're
telling people enough right?
334
00:18:23,400 --> 00:18:27,000
But from user Behavior
perspective, do you think like
335
00:18:27,000 --> 00:18:32,200
this will be something that the
end user will be fine with it
336
00:18:32,200 --> 00:18:35,400
like or not?
You've done some, some, I mean,
337
00:18:35,400 --> 00:18:39,700
testing on that.
So testing is ongoing so far.
338
00:18:39,700 --> 00:18:41,900
The number of users is growing
pretty fast, right.
339
00:18:41,900 --> 00:18:45,400
But we'll see at what point we
need to cross the chasm, right?
340
00:18:45,400 --> 00:18:48,200
We were going to probably hit
the ceiling at some point with
341
00:18:48,200 --> 00:18:53,100
early adopters and then have to
change what we provide to fit
342
00:18:53,300 --> 00:18:54,800
more normal users.
Right?
343
00:18:55,300 --> 00:18:57,500
I don't know when that
transition point is coming, but
344
00:18:57,500 --> 00:19:01,000
it's coming at some point.
So we're still looking at that
345
00:19:01,000 --> 00:19:02,400
feedback.
It's not that we're telling
346
00:19:02,400 --> 00:19:04,700
users enough, right?
I think what we actually
347
00:19:04,700 --> 00:19:07,100
implemented is a
congratulations.
348
00:19:07,400 --> 00:19:10,100
You're now well-informed you
reach their quota for the day,
349
00:19:10,300 --> 00:19:11,500
right?
If they want to continue, they
350
00:19:11,500 --> 00:19:13,200
can continue.
But at least they got this.
351
00:19:13,200 --> 00:19:16,100
Reminder, this was your quarter,
you're done.
352
00:19:16,700 --> 00:19:19,100
It's really want to keep going,
right, right.
353
00:19:19,200 --> 00:19:21,300
But it's also phrased as a
reward, which gives them
354
00:19:21,300 --> 00:19:24,400
positive emotion, instead of
negative emotion, right?
355
00:19:25,100 --> 00:19:26,600
But at least that's the
hypothesis.
356
00:19:26,600 --> 00:19:30,400
Was we still have to test all of
these nudges that we've built
357
00:19:30,400 --> 00:19:33,200
into our system?
So I don't know.
358
00:19:33,200 --> 00:19:35,500
I'm more Read about the
different angle tool.
359
00:19:35,500 --> 00:19:40,800
This, my real concern is that
what every other tool does?
360
00:19:40,800 --> 00:19:43,500
It's not just that they're
trying to maximize time on site,
361
00:19:43,900 --> 00:19:46,800
but they're trying to maximize
your time on site by using the
362
00:19:46,800 --> 00:19:50,800
tools of addiction by variable
rewards to everything that you
363
00:19:50,800 --> 00:19:53,200
see by triggering strong
emotions.
364
00:19:53,400 --> 00:19:57,400
So that, even if that emotion is
- you crave strong, emotion,
365
00:19:57,400 --> 00:20:00,600
next time you see a headline,
right, and people have gotten so
366
00:20:00,600 --> 00:20:03,200
accustomed to these oversized
emotions.
367
00:20:03,200 --> 00:20:07,200
Whenever they Read the news that
now when we give them a clean
368
00:20:07,200 --> 00:20:11,000
version of the news that doesn't
have any of that hate or
369
00:20:11,000 --> 00:20:13,500
agitation or anxiety or that
kind of stuff.
370
00:20:14,000 --> 00:20:16,300
They might be missing something,
right?
371
00:20:16,300 --> 00:20:19,400
They might have been so
accustomed to a know the
372
00:20:19,400 --> 00:20:22,300
sugar-laden version of a cookie
that now that we give them a
373
00:20:22,308 --> 00:20:24,100
real organic cookie.
They don't know what to do with
374
00:20:24,100 --> 00:20:26,400
it anymore, or at least a lot of
people want.
375
00:20:26,500 --> 00:20:30,400
And so, that's my big concern
and that's where most of our
376
00:20:30,700 --> 00:20:32,900
growth hacking experiments are
focusing.
377
00:20:33,100 --> 00:20:37,300
Its what can we substitute?
The bad stuff with once we
378
00:20:37,300 --> 00:20:41,700
remove it because it seems like
just removing might not work.
379
00:20:41,700 --> 00:20:45,800
It's again, just imagine that
some company decided to remove
380
00:20:45,900 --> 00:20:49,200
all the sugar out of older
cookies, people learn going to
381
00:20:49,200 --> 00:20:51,600
buy those cookies, right?
You have to put at least some
382
00:20:51,600 --> 00:20:54,300
erythritol back.
Something to make it resemble,
383
00:20:54,300 --> 00:20:56,200
what people are used to eat,
right?
384
00:20:56,400 --> 00:20:58,800
So that's kind of our big
challenge.
385
00:20:58,800 --> 00:21:03,400
I think the challenge of
reminding people that Their
386
00:21:04,000 --> 00:21:06,200
quota for the day was 50 and
they're about 50.
387
00:21:06,200 --> 00:21:09,000
Now, I don't think people are
going to revolt against that,
388
00:21:09,000 --> 00:21:11,700
that sounds like just the pure
benefit to them, right?
389
00:21:12,100 --> 00:21:17,800
But removing the team sport Vibe
of just hate the other guy.
390
00:21:18,800 --> 00:21:21,700
People might revolt against that
because they're enjoying the
391
00:21:21,700 --> 00:21:24,900
team sport or at least they
think they do because they're so
392
00:21:24,900 --> 00:21:26,200
accustomed to it.
Right.
393
00:21:26,300 --> 00:21:28,700
Yeah.
One thing also that came to my
394
00:21:28,700 --> 00:21:32,100
mind, how you can deal with, I
mean, do you have anything?
395
00:21:32,200 --> 00:21:35,800
Thing that can, you know,
overcome gated content.
396
00:21:36,800 --> 00:21:40,700
So right now we try to avoid it
because we respect the fact that
397
00:21:40,700 --> 00:21:44,600
content is Gator, right?
The reality is most of the Gated
398
00:21:44,600 --> 00:21:47,800
content is actually gated in a
way that only stops humans and
399
00:21:47,800 --> 00:21:50,600
does not stop crawlers at all
because they still want to be
400
00:21:50,600 --> 00:21:52,500
indexed by every search engine
of the right.
401
00:21:52,700 --> 00:21:57,300
And so if I just point my
crawler to a link from Financial
402
00:21:57,300 --> 00:21:59,700
Times, it's going to scrape the
full article from Financial
403
00:21:59,700 --> 00:22:01,900
Times.
Do I want to publish even
404
00:22:02,200 --> 00:22:04,100
Headline or the summary of that
article.
405
00:22:04,100 --> 00:22:06,200
No.
Because it seems to me like if
406
00:22:06,200 --> 00:22:09,500
they put the gate in front of
Financial Times, they're telling
407
00:22:09,500 --> 00:22:11,100
me something.
They're telling me, this is for
408
00:22:11,100 --> 00:22:14,300
my subscribers only, right?
And so we respect that and
409
00:22:14,300 --> 00:22:19,200
typically, initially, we were
just essentially scraping all
410
00:22:19,200 --> 00:22:21,600
the links that somebody posted
on Twitter accounts.
411
00:22:21,600 --> 00:22:24,500
We followed that was the first
version and then we noticed a
412
00:22:24,508 --> 00:22:28,400
lot of those links once I click
on them, as human, I see a pay
413
00:22:28,400 --> 00:22:30,900
wall.
So we started adding a blacklist
414
00:22:30,900 --> 00:22:33,400
to the to love.
Void this domain, avoid this
415
00:22:33,400 --> 00:22:36,700
domain avoid this domain and we
try to maintain it.
416
00:22:36,700 --> 00:22:39,000
So that we avoid every domain
with a paywall.
417
00:22:39,800 --> 00:22:43,700
What I should mention is that
there is so much content out
418
00:22:43,700 --> 00:22:47,800
there that if you're avoiding
every website with a pay wall,
419
00:22:48,200 --> 00:22:51,700
you're not missing any news.
Maybe you're only reading five
420
00:22:51,700 --> 00:22:55,000
versions of this news article
instead of 20, right?
421
00:22:55,000 --> 00:22:58,500
But you're not missing out on
any news of any event out there
422
00:22:58,900 --> 00:23:02,000
because so many outlets right
about there.
423
00:23:02,200 --> 00:23:06,100
I think of that, right?
And so, probably the closest
424
00:23:06,100 --> 00:23:09,800
that we come to having a
difficulty as business news
425
00:23:10,000 --> 00:23:13,200
because just so many of them are
gate, right?
426
00:23:13,400 --> 00:23:16,400
But even there CNBC is not
gated, right?
427
00:23:17,100 --> 00:23:22,400
The mainstream networks, ABC,
NBC CBS PBS, they're not gated
428
00:23:22,400 --> 00:23:24,500
them.
They have a business section,
429
00:23:24,500 --> 00:23:26,700
right?
And so there is enough versions
430
00:23:26,700 --> 00:23:30,900
of even business articles to
just get all the news you want
431
00:23:30,900 --> 00:23:33,500
without ever going.
Into a gate that resource.
432
00:23:34,000 --> 00:23:36,100
Yeah, I think same applies to
take as well.
433
00:23:36,100 --> 00:23:40,300
So but one it's like a comment.
I want to add here.
434
00:23:41,600 --> 00:23:44,600
I understand, you know, media
companies they need to make
435
00:23:44,600 --> 00:23:49,300
money, right.
But putting this strong
436
00:23:49,300 --> 00:23:55,700
headlines on a website with ads
and being gated is something I
437
00:23:55,700 --> 00:23:59,300
cannot understand.
Honestly, speaking like II get
438
00:23:59,500 --> 00:24:01,800
too much, you know like
sometimes.
439
00:24:02,900 --> 00:24:04,900
Why?
Okay, I you want to make it get
440
00:24:04,900 --> 00:24:07,600
it, I understand.
But why you are pushing to me
441
00:24:07,600 --> 00:24:10,600
ads and still I cannot access
the content.
442
00:24:10,600 --> 00:24:13,600
So I hope you can do something
for this.
443
00:24:14,800 --> 00:24:19,600
Just like do two things before
we finish works with you today.
444
00:24:19,800 --> 00:24:26,400
Now again like you're using I
and you know when we talk about
445
00:24:26,400 --> 00:24:28,900
either the ethical aspect of it,
right?
446
00:24:28,900 --> 00:24:34,800
So in general, how do you think
We can keep, you know what, we
447
00:24:34,800 --> 00:24:40,200
call it the ethical
responsibility on in any product
448
00:24:40,200 --> 00:24:43,800
that leverage a i specially in
something that it's touching,
449
00:24:43,800 --> 00:24:48,300
you know, trying to filter the
content to the, to the end user.
450
00:24:49,100 --> 00:24:52,500
So, in our case, I think it's
actually, the answer is pretty
451
00:24:52,500 --> 00:24:57,000
simple, because we are using
only supervised machine
452
00:24:57,000 --> 00:25:00,200
learning, right?
If we just open the model and we
453
00:25:00,200 --> 00:25:04,300
open the data set, Everybody can
tell us if we're being
454
00:25:04,300 --> 00:25:06,700
unethical, right?
It's pretty obvious.
455
00:25:06,700 --> 00:25:10,000
What our model is learning to do
based on these two things.
456
00:25:10,600 --> 00:25:15,700
And so, our answer to this is
transparency, we have to show it
457
00:25:15,700 --> 00:25:17,900
to the world and they can keep
us honest.
458
00:25:17,900 --> 00:25:22,300
Like, we want to be unable to do
anything bad, just because if we
459
00:25:22,300 --> 00:25:24,500
do anything bad, then somebody
will catch it.
460
00:25:24,600 --> 00:25:26,600
It's the same.
I think in cybersecurity you
461
00:25:26,600 --> 00:25:28,500
have the same thing.
The safest tools are the open
462
00:25:28,500 --> 00:25:30,600
source ones.
Because then, if there's any
463
00:25:30,600 --> 00:25:32,400
backdoor that somebody built,
Attend to it.
464
00:25:34,700 --> 00:25:37,300
Then it's open source.
Somebody will notice that back
465
00:25:37,300 --> 00:25:40,000
door, right?
That's why there was this big
466
00:25:40,000 --> 00:25:43,800
hack of LastPass recently, and
all my friends are in last past.
467
00:25:43,800 --> 00:25:46,900
I'm using B Warden, right?
Because I don't trust last past.
468
00:25:46,900 --> 00:25:50,500
The source is closed.
So I'm only using an open source
469
00:25:50,500 --> 00:25:52,100
tool when it comes to something
like this.
470
00:25:52,400 --> 00:25:54,000
So that's kind of our approach
to it.
471
00:25:54,000 --> 00:25:57,500
Now granted, I should put a
caveat, we made our code Source
472
00:25:57,500 --> 00:25:59,900
available, not open source, so
people can look at it.
473
00:25:59,900 --> 00:26:03,100
They cannot copy it and reuse
That's kind of a business
474
00:26:03,100 --> 00:26:04,900
decision to be somewhere in the
middle between those.
475
00:26:05,200 --> 00:26:09,500
Now there's a bigger problem
when you get to Giant
476
00:26:09,500 --> 00:26:12,400
unsupervised language models
that just train themselves on
477
00:26:12,400 --> 00:26:17,500
the entire internet, right?
What GPT for is, I don't think
478
00:26:17,900 --> 00:26:21,400
that we can police their ethics.
They have the police themselves
479
00:26:21,400 --> 00:26:24,100
because we have no way of
knowing what they did, in some
480
00:26:24,100 --> 00:26:27,000
sense, by the time the end
product comes to us.
481
00:26:27,300 --> 00:26:30,400
All we can do is kind of poke it
a little bit from the sides, but
482
00:26:30,400 --> 00:26:34,400
we don't actually You know,
what's underneath and if it was
483
00:26:34,400 --> 00:26:37,800
mostly trained on Neo Nazi
propaganda, we might not know,
484
00:26:38,200 --> 00:26:40,000
right?
So it's up to them.
485
00:26:40,000 --> 00:26:44,600
And here, I hope that we can
develop some relatively standard
486
00:26:44,600 --> 00:26:49,400
models of Ethics, but I'm also a
little bit skeptical because I
487
00:26:49,400 --> 00:26:52,400
don't think we have those for
humans, right?
488
00:26:52,400 --> 00:26:55,900
The ethical models that we have
for human behavior are pretty
489
00:26:55,900 --> 00:26:59,100
bad.
If you think about it, in fact,
490
00:26:59,700 --> 00:27:01,500
they're kind of different in
each sphere.
491
00:27:01,900 --> 00:27:03,200
Right?
So if you look at the legal
492
00:27:03,200 --> 00:27:06,800
sphere, what we have is purely
deontic ethics, right?
493
00:27:06,900 --> 00:27:10,000
Somebody wrote the law.
Even if that law is outdated and
494
00:27:10,000 --> 00:27:13,500
let's say the law in the US on
eavesdropping was written. 1973
495
00:27:14,000 --> 00:27:16,600
we still follow that law as
written, even though it was
496
00:27:16,600 --> 00:27:21,600
written 1973, right?
So it's It's inefficient but
497
00:27:21,600 --> 00:27:24,300
that's what we have right.
Should we have that for AI
498
00:27:24,300 --> 00:27:27,300
obviously, that model with
failed because the I just moves
499
00:27:27,300 --> 00:27:30,300
too quickly, right?
Then once you talk to the
500
00:27:30,300 --> 00:27:34,100
entrepreneurs themselves, a lot
of them follow some version of
501
00:27:34,100 --> 00:27:38,000
utilitarian ethics, right?
Basically, if the result is
502
00:27:38,000 --> 00:27:41,000
good, then the action was good
and that's pretty stupid because
503
00:27:41,000 --> 00:27:43,000
when you make the action, you
cannot predict the results.
504
00:27:43,000 --> 00:27:48,600
So how is that an ethical
standard for make for choosing
505
00:27:48,600 --> 00:27:51,800
good actions, right?
So, I don't know.
506
00:27:52,200 --> 00:27:56,200
I hope that we can come to some
version of Greek virtue ethics
507
00:27:56,200 --> 00:27:58,000
when it comes to this sort of
thing.
508
00:27:58,400 --> 00:28:01,400
But I'm not holding my breath,
because most humans don't follow
509
00:28:01,400 --> 00:28:03,600
that either.
So, why would a i models follow?
510
00:28:03,800 --> 00:28:06,100
Yeah, it would be very
interesting to watch this space,
511
00:28:06,100 --> 00:28:10,300
but yeah, I like the analogy
that you did because sometimes,
512
00:28:10,300 --> 00:28:14,600
you know, I used to know to, to
go and do consultancies in
513
00:28:14,600 --> 00:28:17,300
different fields with customers
and then, you know, we touched
514
00:28:17,300 --> 00:28:20,800
on the compliance, for example,
and you find out on Policies.
515
00:28:20,800 --> 00:28:24,200
And you find out that someone
wrote that document, when he
516
00:28:24,200 --> 00:28:28,100
eats 25 his back and okay, I
found it like this, I don't
517
00:28:28,100 --> 00:28:30,200
know, you know.
So I knew but with a is
518
00:28:30,200 --> 00:28:33,800
dangerous because as you said
it's changing on daily basis
519
00:28:33,800 --> 00:28:40,000
almost now they're talking about
GPT 5 and people are asking
520
00:28:40,000 --> 00:28:40,700
them.
No, no, no.
521
00:28:40,700 --> 00:28:43,600
Don't release it yet.
We still didn't digest that one.
522
00:28:44,600 --> 00:28:47,500
But anyway, we'll see.
Finally, you know, question for
523
00:28:47,500 --> 00:28:50,500
you, Alex?
What do you think the A future
524
00:28:50,500 --> 00:28:54,100
of information.
Ecosystem overall gonna be in
525
00:28:54,100 --> 00:28:57,200
the coming years and content
creation, and all the space.
526
00:28:57,900 --> 00:29:00,700
So, I think you have to consider
that a, I will make.
527
00:29:01,700 --> 00:29:04,100
First of all, it will make
content creation, at least the
528
00:29:04,100 --> 00:29:06,500
low quality content creation,
much easier.
529
00:29:06,800 --> 00:29:09,600
So, you will see an even bigger
explosion that you've seen.
530
00:29:09,600 --> 00:29:13,600
Then you've seen in the past 10
years and bad content being
531
00:29:13,600 --> 00:29:16,500
generated.
We will be drowning in jump at
532
00:29:16,500 --> 00:29:19,200
the same time.
One of the other interesting
533
00:29:19,200 --> 00:29:24,400
effects that Ai and specifically
the GPT style generative.
534
00:29:24,400 --> 00:29:29,900
A I will have is that I think
search with search results on
535
00:29:29,900 --> 00:29:33,600
the page will probably go away
as a tool because you will now
536
00:29:33,600 --> 00:29:36,300
have search engines that can
just answer your question.
537
00:29:36,700 --> 00:29:40,600
Give you one answer that was
generated instead of here, are
538
00:29:40,600 --> 00:29:45,000
55,000 answers to choose from.
And that means that most of
539
00:29:45,000 --> 00:29:47,500
these websites, that generate,
the junk won't get much less
540
00:29:47,500 --> 00:29:50,500
traffic from search engines.
So combine these Two things
541
00:29:50,500 --> 00:29:54,300
together, junk is, cheaper to
generate, but harder to
542
00:29:54,300 --> 00:29:57,700
monetize, because less money is
coming in from search engines,
543
00:29:58,200 --> 00:30:00,100
right?
And it's an entire race to the
544
00:30:00,100 --> 00:30:03,000
bottom, where everybody will try
to create more and more junk to
545
00:30:03,008 --> 00:30:06,300
fight for the last scraps of
traffic that are available.
546
00:30:06,800 --> 00:30:10,100
Basically we have to create
filters otherwise we will draw.
547
00:30:10,500 --> 00:30:13,100
So I don't know if the filter is
that we are making are the right
548
00:30:13,100 --> 00:30:16,600
ones, maybe somebody else will
make better ones but whatever it
549
00:30:16,600 --> 00:30:20,000
is, we're going to need filters
because the junk is coming.
550
00:30:20,400 --> 00:30:24,300
If it's going to be like a giant
Avalanche of elephants blowing
551
00:30:24,300 --> 00:30:27,100
bubbles, the is an example that
we just touched.
552
00:30:27,100 --> 00:30:32,500
Yeah, so that's kind of my dire
side of the production but I'll
553
00:30:32,500 --> 00:30:35,200
give you a hopeful note as well.
If you look at the closest
554
00:30:35,200 --> 00:30:38,200
historic, parallel to what we're
seeing now, right?
555
00:30:38,200 --> 00:30:43,300
The printing, press was invented
in the 1430s, we saw 200 years
556
00:30:43,300 --> 00:30:47,200
of inquisitions, witch hunts 52
religious wars, all over Europe.
557
00:30:47,300 --> 00:30:50,000
People were drowning and new
junk being published.
558
00:30:50,200 --> 00:30:53,400
And that they just couldn't
digest or process right?
559
00:30:53,600 --> 00:30:58,500
But then what came out of it is
the scientific method and 1635 I
560
00:30:58,508 --> 00:31:03,000
think peer review in 1665 right?
And then the Enlightenment and
561
00:31:03,000 --> 00:31:05,900
then Universal literacy all over
Europe.
562
00:31:06,000 --> 00:31:08,500
Right.
And So eventually the outcome is
563
00:31:08,500 --> 00:31:10,500
good.
We just need to learn how to get
564
00:31:10,500 --> 00:31:12,100
there without blowing ourselves
up.
565
00:31:13,100 --> 00:31:15,600
Yeah, cool.
And just as a final note on
566
00:31:15,600 --> 00:31:20,800
this, I liked you you know
positivity on That's because
567
00:31:20,800 --> 00:31:23,900
this is what I'm telling people
who usually are showing like
568
00:31:23,900 --> 00:31:25,200
little bit.
They are scared.
569
00:31:25,500 --> 00:31:27,200
I'm telling them.
No.
570
00:31:27,300 --> 00:31:29,900
Okay.
It might be looking scary now
571
00:31:29,900 --> 00:31:34,400
but if you think on the long run
I believe yeah as you said it
572
00:31:34,400 --> 00:31:37,300
might be a moment of
Enlightenment for Humanity.
573
00:31:37,500 --> 00:31:41,200
We need to be careful of course.
Well elect thank you very much
574
00:31:41,200 --> 00:31:46,600
for your time today.
It was I think a very like Rich
575
00:31:46,600 --> 00:31:51,100
conversation with you on this
topic Thank you, everyone for
576
00:31:51,100 --> 00:31:55,100
tuning in with us today.
Thank you for watching or if you
577
00:31:55,100 --> 00:31:58,100
are listening, as usual.
If you have any comment, if you
578
00:31:58,100 --> 00:32:01,400
have any question about this
episode or future, you know,
579
00:32:01,400 --> 00:32:04,600
topics you want to bring in,
please reach out to me by email
580
00:32:05,000 --> 00:32:08,300
LinkedIn, or Twitter and until
we meet in the next episode.
581
00:32:08,300 --> 00:32:09,300
Thank you very much.
