Del Harvey: The strangeness of scale at Twitter

104,892 views ・ 2014-03-27

TED


Please double-click on the English subtitles below to play the video.

00:12
My job at Twitter
0
12984
1291
00:14
is to ensure user trust,
1
14275
1978
00:16
protect user rights and keep users safe,
2
16253
2837
00:19
both from each other
3
19090
1260
00:20
and, at times, from themselves.
4
20350
3899
00:24
Let's talk about what scale looks like at Twitter.
5
24249
4275
00:28
Back in January 2009,
6
28524
2870
00:31
we saw more than two million new tweets each day
7
31394
3331
00:34
on the platform.
8
34725
1764
00:36
January 2014, more than 500 million.
9
36489
5908
00:42
We were seeing two million tweets
10
42397
2492
00:44
in less than six minutes.
11
44889
2176
00:47
That's a 24,900-percent increase.
12
47065
6984
00:54
Now, the vast majority of activity on Twitter
13
54049
3253
00:57
puts no one in harm's way.
14
57302
1503
00:58
There's no risk involved.
15
58805
1935
01:00
My job is to root out and prevent activity that might.
16
60740
5753
01:06
Sounds straightforward, right?
17
66493
1973
01:08
You might even think it'd be easy,
18
68466
1152
01:09
given that I just said the vast majority
19
69618
2170
01:11
of activity on Twitter puts no one in harm's way.
20
71788
3810
01:15
Why spend so much time
21
75598
2169
01:17
searching for potential calamities
22
77767
2743
01:20
in innocuous activities?
23
80510
2900
01:23
Given the scale that Twitter is at,
24
83410
2940
01:26
a one-in-a-million chance happens
25
86350
2357
01:28
500 times a day.
26
88707
4876
01:33
It's the same for other companies
27
93583
1445
01:35
dealing at this sort of scale.
28
95028
1471
01:36
For us, edge cases,
29
96499
1708
01:38
those rare situations that are unlikely to occur,
30
98207
3625
01:41
are more like norms.
31
101832
2622
01:44
Say 99.999 percent of tweets
32
104454
3942
01:48
pose no risk to anyone.
33
108396
1888
01:50
There's no threat involved.
34
110284
1066
01:51
Maybe people are documenting travel landmarks
35
111350
2954
01:54
like Australia's Heart Reef,
36
114304
1963
01:56
or tweeting about a concert they're attending,
37
116267
2921
01:59
or sharing pictures of cute baby animals.
38
119188
4747
02:03
After you take out that 99.999 percent,
39
123935
4509
02:08
that tiny percentage of tweets remaining
40
128444
3529
02:11
works out to roughly
41
131973
2389
02:14
150,000 per month.
42
134362
3475
02:17
The sheer scale of what we're dealing with
43
137837
2456
02:20
makes for a challenge.
44
140293
2312
02:22
You know what else makes my role
45
142605
1178
02:23
particularly challenging?
46
143783
3107
02:26
People do weird things.
47
146890
5123
02:32
(Laughter)
48
152013
1829
02:33
And I have to figure out what they're doing,
49
153842
2391
02:36
why, and whether or not there's risk involved,
50
156233
2249
02:38
often without much in terms of context
51
158482
2168
02:40
or background.
52
160650
1847
02:42
I'm going to show you some examples
53
162497
2077
02:44
that I've run into during my time at Twitter --
54
164574
2005
02:46
these are all real examples —
55
166579
1620
02:48
of situations that at first seemed cut and dried,
56
168199
2653
02:50
but the truth of the matter was something
57
170852
1643
02:52
altogether different.
58
172495
1550
02:54
The details have been changed
59
174045
1977
02:56
to protect the innocent
60
176022
1257
02:57
and sometimes the guilty.
61
177279
3233
03:00
We'll start off easy.
62
180512
3005
03:03
["Yo bitch"]
63
183517
1793
03:05
If you saw a Tweet that only said this,
64
185310
3228
03:08
you might think to yourself,
65
188538
1694
03:10
"That looks like abuse."
66
190232
1653
03:11
After all, why would you want to receive the message,
67
191885
3107
03:14
"Yo, bitch."
68
194992
2218
03:17
Now, I try to stay relatively hip
69
197210
4663
03:21
to the latest trends and memes,
70
201873
2512
03:24
so I knew that "yo, bitch"
71
204385
2704
03:27
was also often a common greeting between friends,
72
207089
3154
03:30
as well as being a popular "Breaking Bad" reference.
73
210243
4262
03:34
I will admit that I did not expect
74
214505
2487
03:36
to encounter a fourth use case.
75
216992
2841
03:39
It turns out it is also used on Twitter
76
219833
3104
03:42
when people are role-playing as dogs.
77
222937
3062
03:45
(Laughter)
78
225999
5279
03:51
And in fact, in that case,
79
231278
1666
03:52
it's not only not abusive,
80
232944
1609
03:54
it's technically just an accurate greeting.
81
234553
3139
03:57
(Laughter)
82
237692
2889
04:00
So okay, determining whether or not
83
240581
2071
04:02
something is abusive without context,
84
242652
1848
04:04
definitely hard.
85
244500
1592
04:06
Let's look at spam.
86
246092
2717
04:08
Here's an example of an account engaged
87
248809
1960
04:10
in classic spammer behavior,
88
250769
1668
04:12
sending the exact same message
89
252437
1559
04:13
to thousands of people.
90
253996
1804
04:15
While this is a mockup I put together using my account,
91
255800
2793
04:18
we see accounts doing this all the time.
92
258593
3001
04:21
Seems pretty straightforward.
93
261594
1979
04:23
We should just automatically suspend accounts
94
263573
2053
04:25
engaging in this kind of behavior.
95
265626
3307
04:28
Turns out there's some exceptions to that rule.
96
268933
3210
04:32
Turns out that that message could also be a notification
97
272143
2883
04:35
you signed up for that the International Space Station is passing overhead
98
275026
3889
04:38
because you wanted to go outside
99
278915
1846
04:40
and see if you could see it.
100
280761
1948
04:42
You're not going to get that chance
101
282709
1225
04:43
if we mistakenly suspend the account
102
283934
1847
04:45
thinking it's spam.
103
285781
2266
04:48
Okay. Let's make the stakes higher.
104
288047
3526
04:51
Back to my account,
105
291573
1916
04:53
again exhibiting classic behavior.
106
293489
3505
04:56
This time it's sending the same message and link.
107
296994
2643
04:59
This is often indicative of something called phishing,
108
299637
2774
05:02
somebody trying to steal another person's account information
109
302411
3178
05:05
by directing them to another website.
110
305589
2203
05:07
That's pretty clearly not a good thing.
111
307792
4194
05:11
We want to, and do, suspend accounts
112
311986
1930
05:13
engaging in that kind of behavior.
113
313916
2624
05:16
So why are the stakes higher for this?
114
316540
3247
05:19
Well, this could also be a bystander at a rally
115
319787
2999
05:22
who managed to record a video
116
322786
1910
05:24
of a police officer beating a non-violent protester
117
324696
3270
05:27
who's trying to let the world know what's happening.
118
327966
2975
05:30
We don't want to gamble
119
330941
1643
05:32
on potentially silencing that crucial speech
120
332584
2517
05:35
by classifying it as spam and suspending it.
121
335101
2929
05:38
That means we evaluate hundreds of parameters
122
338030
2879
05:40
when looking at account behaviors,
123
340909
1688
05:42
and even then, we can still get it wrong
124
342597
2016
05:44
and have to reevaluate.
125
344613
2236
05:46
Now, given the sorts of challenges I'm up against,
126
346849
3708
05:50
it's crucial that I not only predict
127
350557
2696
05:53
but also design protections for the unexpected.
128
353253
3784
05:57
And that's not just an issue for me,
129
357037
2342
05:59
or for Twitter, it's an issue for you.
130
359379
2087
06:01
It's an issue for anybody who's building or creating
131
361466
2406
06:03
something that you think is going to be amazing
132
363872
1925
06:05
and will let people do awesome things.
133
365797
2789
06:08
So what do I do?
134
368586
2866
06:11
I pause and I think,
135
371452
3318
06:14
how could all of this
136
374770
2095
06:16
go horribly wrong?
137
376865
3793
06:20
I visualize catastrophe.
138
380658
4453
06:25
And that's hard. There's a sort of
139
385111
2463
06:27
inherent cognitive dissonance in doing that,
140
387574
2848
06:30
like when you're writing your wedding vows
141
390422
1812
06:32
at the same time as your prenuptial agreement.
142
392234
2646
06:34
(Laughter)
143
394880
1696
06:36
But you still have to do it,
144
396576
2373
06:38
particularly if you're marrying 500 million tweets per day.
145
398949
4446
06:43
What do I mean by "visualize catastrophe?"
146
403395
3097
06:46
I try to think of how something as
147
406492
2762
06:49
benign and innocuous as a picture of a cat
148
409254
3228
06:52
could lead to death,
149
412482
1104
06:53
and what to do to prevent that.
150
413586
2326
06:55
Which happens to be my next example.
151
415912
2383
06:58
This is my cat, Eli.
152
418295
3110
07:01
We wanted to give users the ability
153
421405
1981
07:03
to add photos to their tweets.
154
423386
2073
07:05
A picture is worth a thousand words.
155
425459
1597
07:07
You only get 140 characters.
156
427056
2009
07:09
You add a photo to your tweet,
157
429065
1200
07:10
look at how much more content you've got now.
158
430265
3038
07:13
There's all sorts of great things you can do
159
433303
1677
07:14
by adding a photo to a tweet.
160
434980
2007
07:16
My job isn't to think of those.
161
436987
2280
07:19
It's to think of what could go wrong.
162
439267
2747
07:22
How could this picture
163
442014
1892
07:23
lead to my death?
164
443906
3539
07:27
Well, here's one possibility.
165
447445
3160
07:30
There's more in that picture than just a cat.
166
450605
3086
07:33
There's geodata.
167
453691
2092
07:35
When you take a picture with your smartphone
168
455783
2212
07:37
or digital camera,
169
457995
1299
07:39
there's a lot of additional information
170
459294
1654
07:40
saved along in that image.
171
460948
1616
07:42
In fact, this image also contains
172
462564
1932
07:44
the equivalent of this,
173
464496
1805
07:46
more specifically, this.
174
466301
3079
07:49
Sure, it's not likely that someone's going to try
175
469380
1956
07:51
to track me down and do me harm
176
471336
2285
07:53
based upon image data associated
177
473621
1784
07:55
with a picture I took of my cat,
178
475405
1948
07:57
but I start by assuming the worst will happen.
179
477353
3651
08:01
That's why, when we launched photos on Twitter,
180
481004
2338
08:03
we made the decision to strip that geodata out.
181
483342
3821
08:07
(Applause)
182
487163
5847
08:13
If I start by assuming the worst
183
493010
2613
08:15
and work backwards,
184
495623
947
08:16
I can make sure that the protections we build
185
496570
2553
08:19
work for both expected
186
499123
1768
08:20
and unexpected use cases.
187
500891
2078
08:22
Given that I spend my days and nights
188
502969
2945
08:25
imagining the worst that could happen,
189
505914
2541
08:28
it wouldn't be surprising if my worldview was gloomy.
190
508455
4257
08:32
(Laughter)
191
512712
1783
08:34
It's not.
192
514495
1417
08:35
The vast majority of interactions I see --
193
515912
3876
08:39
and I see a lot, believe me -- are positive,
194
519788
3901
08:43
people reaching out to help
195
523689
1924
08:45
or to connect or share information with each other.
196
525613
3448
08:49
It's just that for those of us dealing with scale,
197
529061
3323
08:52
for those of us tasked with keeping people safe,
198
532384
3800
08:56
we have to assume the worst will happen,
199
536184
2546
08:58
because for us, a one-in-a-million chance
200
538730
4227
09:02
is pretty good odds.
201
542957
2749
09:05
Thank you.
202
545706
1864
09:07
(Applause)
203
547570
4000
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7