WEBVTT
Kind: captions
Language: en

1
00:00:00.000 --> 00:00:02.160
As I record this, it's February 2021.

2
00:00:02.160 --> 00:00:05.360
We're nearly two months into
Britain's third lockdown right now,

3
00:00:05.360 --> 00:00:06.160
and I've got a problem.

4
00:00:06.800 --> 00:00:08.240
I'm out of ideas.

5
00:00:08.240 --> 00:00:11.040
I've now recorded all the computer science
and language ideas I've got,

6
00:00:11.040 --> 00:00:12.000
at least, all the good ones,

7
00:00:12.000 --> 00:00:14.320
and to be honest, I'm kinda
burned out on those anyway.

8
00:00:14.320 --> 00:00:15.680
I can travel to film stuff,

9
00:00:15.680 --> 00:00:18.800
but it has to be outdoors,
and within England.

10
00:00:18.800 --> 00:00:21.760
Apparently I filmed most of those stories
years ago when I was starting out.

11
00:00:21.760 --> 00:00:23.520
I tried hiring a couple of freelance researchers,

12
00:00:23.520 --> 00:00:24.640
they were wonderful,
they did their best,

13
00:00:24.640 --> 00:00:28.480
they found some stories that would be great
for other people or for BBC documentaries.

14
00:00:28.480 --> 00:00:32.800
But there seems to be some aspect to a video
that's "right for this channel"

15
00:00:32.800 --> 00:00:33.920
that I can't put into words.

16
00:00:33.920 --> 00:00:36.720
And if I can't do that,
I can't expect anyone else to.

17
00:00:37.360 --> 00:00:39.520
But: I just got access to GPT-3,

18
00:00:39.520 --> 00:00:43.120
the state-of-the-art artificial intelligence
text generator made by OpenAI.

19
00:00:43.120 --> 00:00:44.960
I'm not going to go into the details
of how it works here,

20
00:00:44.960 --> 00:00:47.280
other people have done that
far better than I can, but in short:

21
00:00:47.280 --> 00:00:50.320
it has read and stored a
huge amount of the web.

22
00:00:50.320 --> 00:00:51.440
And if you give it some text,

23
00:00:51.440 --> 00:00:53.760
it will try to continue that text
in the same style.

24
00:00:53.760 --> 00:00:55.920
But it's not just repeating
what it's already seen:

25
00:00:55.920 --> 00:00:59.280
it's coming up with brand new,
never-before-written sentences.

26
00:00:59.280 --> 00:01:03.120
It is incredibly clever
in specific but very limited ways.

27
00:01:03.120 --> 00:01:07.120
And the challenge of "come up with
new video ideas in this style"

28
00:01:07.120 --> 00:01:09.040
is exactly what it's meant to be good at.

29
00:01:09.040 --> 00:01:10.880
So I gave it a list of
some of my previous videos:

30
00:01:10.880 --> 00:01:14.560
ones that I'd filmed solo in England,
and I asked it: come up with more.

31
00:01:14.560 --> 00:01:17.200
Maybe it could find the connection
that I can't describe,

32
00:01:17.200 --> 00:01:19.280
pull from all the billions
of things in its memory,

33
00:01:19.280 --> 00:01:22.240
and find the stories that I've missed.
[loud car driving by]

34
00:01:22.240 --> 00:01:25.120
It took a little bit of tweaking to
get the prompt and the parameters right.

35
00:01:25.120 --> 00:01:28.720
At one point it generated part of a
fake interview with me, which was weird.

36
00:01:28.720 --> 00:01:32.240
But pretty quickly it was churning out
plausible-sounding video titles.

37
00:01:32.240 --> 00:01:34.800
In the end, I generated more than
a thousand of them.

38
00:01:34.800 --> 00:01:37.440
I knew most of them wouldn't
be any good, but I thought,

39
00:01:37.440 --> 00:01:39.840
maybe it can manage
at least one in a thousand?

40
00:01:40.480 --> 00:01:42.320
And... it sorta did.

41
00:01:42.960 --> 00:01:45.840
Now, OpenAI's terms of service don't
let me give you the full list.

42
00:01:45.840 --> 00:01:48.000
I have to curate them,
and show you a sample.

43
00:01:48.000 --> 00:01:50.080
Those are the terms and
conditions I agreed to.

44
00:01:50.080 --> 00:01:54.480
But that's not actually much of a problem,
because most of the titles were just... dull.

45
00:01:54.480 --> 00:01:56.320
Which is fair enough, most of the video ideas 

46
00:01:56.320 --> 00:01:58.880
I come up with also get 
discarded almost immediately

47
00:01:58.880 --> 00:02:00.480
because they're just not good enough.

48
00:02:00.480 --> 00:02:04.960
The AI suggested titles like
"The World's First Bridge Over A Motorway".

49
00:02:04.960 --> 00:02:07.440
"An Industrial Warehouse
In The Middle Of Nowhere."

50
00:02:07.440 --> 00:02:09.520
and "The Beach Where You Can Hear the Sea."

51
00:02:09.520 --> 00:02:11.680
[ocean waves]

52
00:02:11.680 --> 00:02:12.960
I... can hear the sea.

53
00:02:13.760 --> 00:02:16.320
Most of the ideas were plausible,
they're valid sentences,

54
00:02:16.320 --> 00:02:17.840
they're just not interesting.

55
00:02:17.840 --> 00:02:20.480
There were a few suggestions
that were just nonsense.

56
00:02:20.480 --> 00:02:22.800
"The British Road That Is Also A Boat."

57
00:02:22.800 --> 00:02:25.120
"The Cliff That Refuses To Be A Cliff."

58
00:02:25.120 --> 00:02:27.440
"The Green Death And
The Industrial Revolution."

59
00:02:27.440 --> 00:02:29.600
I'm actually fairly sure that
last one's a Doctor Who episode.

60
00:02:30.160 --> 00:02:32.880
GPT-3 has a setting called "temperature",

61
00:02:32.880 --> 00:02:34.960
which is basically how
predictable it should be.

62
00:02:34.960 --> 00:02:37.040
If I turned the temperature down too low,

63
00:02:37.040 --> 00:02:40.320
it just repeated existing video titles
over and over back to me.

64
00:02:40.320 --> 00:02:44.400
Too high, and I got ideas like
"Jeremy Clarkson's Lottery of Death"

65
00:02:44.400 --> 00:02:47.520
which I swear was a genuine suggestion
it gave me and which I assume

66
00:02:47.520 --> 00:02:49.280
will be coming to Amazon Prime next year.

67
00:02:49.280 --> 00:02:54.000
But when I got the temperature right,
it did actually suggest some real places!

68
00:02:54.000 --> 00:02:56.160
"The Abandoned Homes of Wigan Pier".

69
00:02:56.160 --> 00:02:59.360
Well, Wigan Pier's a derelict industrial site
in the West Midlands that's being redeveloped,

70
00:02:59.360 --> 00:03:03.760
it's not a story I could tell, and it's not
one I know about, but someone else could do that.

71
00:03:03.760 --> 00:03:06.960
"The Super-Sized Megalithic Stones of Avebury".

72
00:03:06.960 --> 00:03:08.560
I made a video about Avebury years ago

73
00:03:08.560 --> 00:03:10.000
when I was just starting out,
filming on my phone,

74
00:03:10.000 --> 00:03:12.480
Avebury is like Stonehenge but better.

75
00:03:12.480 --> 00:03:13.920
A+ suggestion...

76
00:03:13.920 --> 00:03:16.960
if I hadn't already done it.
It also suggested stories about

77
00:03:16.960 --> 00:03:20.800
the RRS Discovery, an early 20th-century
steamship that explored the Antarctic,

78
00:03:20.800 --> 00:03:24.800
about the Aberfan disaster, where more than
a hundred people died in a mining landslip,

79
00:03:24.800 --> 00:03:27.520
about the New River, an aqueduct
that runs through London,

80
00:03:27.520 --> 00:03:31.040
and about the Post Office Railway,
where I've actually visited before,

81
00:03:31.040 --> 00:03:32.240
just not to film.

82
00:03:32.240 --> 00:03:34.800
None of those were quite right
to make a video about, sadly,

83
00:03:34.800 --> 00:03:37.360
I couldn't find the narrative hook
that means they work for this channel,

84
00:03:37.360 --> 00:03:39.120
but they're close.

85
00:03:39.120 --> 00:03:42.080
It also suggested "The Failed Bomb Plot Pigeon".

86
00:03:42.080 --> 00:03:45.600
That was an article that was used
on Citation Needed a few years ago!

87
00:03:45.600 --> 00:03:50.000
It even suggested a video idea that I've
already filmed but that I haven't released yet.

88
00:03:50.000 --> 00:03:51.440
I'll keep that title secret for now.

89
00:03:51.440 --> 00:03:54.800
But... yeah. It worked. I wasn't expecting that.

90
00:03:54.800 --> 00:03:58.720
Videos talking about testing AI are usually
comedy about how bad the results are,

91
00:03:58.720 --> 00:03:59.840
they're not supposed to work.

92
00:04:00.880 --> 00:04:03.280
There were some suggestions outside the UK, too.

93
00:04:03.280 --> 00:04:05.920
"The Airport Which Lands Planes On Water."

94
00:04:05.920 --> 00:04:09.360
I googled it, and that would be
Vancouver Harbour's seaplane terminal.

95
00:04:09.360 --> 00:04:12.160
I've been there. I've seen it before,
I'd never thought to film it.

96
00:04:12.160 --> 00:04:16.080
Also, there's "How The CIA Plotted To
Kill Castro With Explosive Seashells."

97
00:04:16.080 --> 00:04:18.720
Which actually happened, that's a real thing,
but I googled those words

98
00:04:18.720 --> 00:04:22.160
and they have never been put in
that exact order anywhere online.

99
00:04:22.160 --> 00:04:24.640
It's not a video I can film,
but it's a good idea for someone.

100
00:04:24.640 --> 00:04:27.280
But the most interesting category of suggestion

101
00:04:27.280 --> 00:04:30.560
was the one where the AI came up with videos
that I would love to make...

102
00:04:31.360 --> 00:04:33.120
but which are completely fictional.

103
00:04:33.120 --> 00:04:36.160
Because of course it would:
it doesn't understand any of the words,

104
00:04:36.160 --> 00:04:39.680
it's not checking if they're true,
it's just looking for patterns.

105
00:04:39.680 --> 00:04:43.520
So with the help of an animator,
here are some videos that I can't make

106
00:04:44.080 --> 00:04:45.120
because they're not real.

107
00:04:45.760 --> 00:04:48.880
The Village That Was Saved
By Synchronised Traffic Lights.

108
00:04:48.880 --> 00:04:51.040
The White Cube At The End Of The World.

109
00:04:51.040 --> 00:04:54.080
The Sanity-Defying Crossroads
on the Isle of Skye.

110
00:04:54.080 --> 00:04:58.320
How A Secret Gadget In The Beam Of The
Olympic Stadium Helped Deliver Britain's Medals.

111
00:04:58.320 --> 00:05:01.440
The Long Forgotten History of
the British Moon Landings.

112
00:05:01.440 --> 00:05:04.640
The Island With A Gravity Defying
Vertical Mountain.

113
00:05:04.640 --> 00:05:07.360
The Strange Light That
Floats Over Oxfordshire.

114
00:05:07.360 --> 00:05:10.080
The Lost Village of Lofthouse
That's Now Invisible.

115
00:05:10.080 --> 00:05:13.280
The Dream of a Russian Utopia
in East Yorkshire.

116
00:05:13.280 --> 00:05:14.960
I want to film all of those!

117
00:05:14.960 --> 00:05:19.200
There were hundreds of perfect video ideas
for my channel. But... they're not real.

118
00:05:19.200 --> 00:05:21.040
They're all in some weird parallel universe,

119
00:05:21.040 --> 00:05:24.320
and I don't have a word, I don't think there's
a word in the English language for

120
00:05:24.320 --> 00:05:27.760
nostalgia for a thing that never existed,
that never could have existed.

121
00:05:28.560 --> 00:05:29.840
But what if those were videos?

122
00:05:29.840 --> 00:05:32.800
What if I could click on any of those
topics and get a video about it?

123
00:05:32.800 --> 00:05:34.880
So as an experiment,
I took one of those subjects,

124
00:05:34.880 --> 00:05:37.120
the one about the Russian utopia
in East Yorkshire,

125
00:05:37.120 --> 00:05:39.440
and I told the AI to write the script.

126
00:05:39.440 --> 00:05:40.720
I just gave it one instruction.

127
00:05:41.440 --> 00:05:43.120
And it did it. Really well.

128
00:05:43.120 --> 00:05:44.800
So I filmed it,
or at least the start of it.

129
00:05:44.800 --> 00:05:49.440
Which is why I'm standing here, in another
locked-down windswept bit of England,

130
00:05:49.440 --> 00:05:52.240
because the AI gave me exactly
what I asked for...

131
00:05:53.120 --> 00:05:54.560
even though it wasn't true.

132
00:05:54.560 --> 00:05:57.280
In the late 1800s,
a wealthy Russian noblewoman

133
00:05:57.280 --> 00:06:00.240
by the name Baroness Helena von Hahn
had a dream.

134
00:06:00.240 --> 00:06:03.280
A dream of a perfect world in which
the relentless march of capitalism

135
00:06:03.280 --> 00:06:05.600
would be slowed enough to allow people
to enjoy their lives

136
00:06:05.600 --> 00:06:07.520
and not be forced to work constantly.

137
00:06:07.520 --> 00:06:09.520
She found her chance when she
learned that Thomas Kirby,

138
00:06:09.520 --> 00:06:12.720
an English businessman and entrepreneur
had just bought up large tracts of land

139
00:06:12.720 --> 00:06:15.680
in an area called Holderness,
in West Yorkshire.

140
00:06:15.680 --> 00:06:18.640
Through her friendship with his sister,
she convinced him to sell it to her.

141
00:06:18.640 --> 00:06:20.960
The site was dubbed "Kirby's Dream".

142
00:06:20.960 --> 00:06:24.240
The baroness then set about building
schools, hospitals and parks on the site

143
00:06:24.240 --> 00:06:26.000
as well as planting thousands of trees.

144
00:06:26.800 --> 00:06:29.440
Unfortunately, it was too good to last.

145
00:06:29.440 --> 00:06:32.000
I didn't run that a hundred times
and pick the best script.

146
00:06:32.000 --> 00:06:34.160
That was the first draft it gave me.

147
00:06:34.160 --> 00:06:35.120
None of that's real.

148
00:06:35.120 --> 00:06:38.000
There was a Russian noblewoman with that name,
but she wasn't a Baroness.

149
00:06:38.000 --> 00:06:40.480
There are no records of an enterpreneur
called Thomas Kirby,

150
00:06:40.480 --> 00:06:42.160
the land purchase never happened.

151
00:06:42.160 --> 00:06:47.120
This is Holderness! And it's just farmland,
there never was any Russian utopia.

152
00:06:47.120 --> 00:06:50.320
I could have asked for a script
for any of the thousand video titles.

153
00:06:50.320 --> 00:06:52.160
Or a thousand different scripts
for any of them,

154
00:06:52.160 --> 00:06:54.800
a thousand different failed
Russian utopias in Yorkshire.

155
00:06:55.440 --> 00:06:57.200
There's a reason that the company, OpenAI,

156
00:06:57.200 --> 00:06:59.760
is putting limits and ethical guidelines
on its text generator,

157
00:06:59.760 --> 00:07:03.040
because they're the only things
stopping people from asking for

158
00:07:03.040 --> 00:07:06.640
human-sounding political arguments,
or fake angry customer reviews,

159
00:07:06.640 --> 00:07:08.480
or... well, anything.

160
00:07:09.120 --> 00:07:11.840
Three years ago, text generation
like that was impossible.

161
00:07:11.840 --> 00:07:14.320
New technologies do tend to
follow a sigmoid curve,

162
00:07:14.320 --> 00:07:15.280
nothing happens for ages,

163
00:07:15.280 --> 00:07:18.320
then there's a massive revolution and
it feels like the world's changing,

164
00:07:18.880 --> 00:07:20.960
then it reaches a new fairly steady state.

165
00:07:20.960 --> 00:07:22.880
That happened with the web,
with smartphones,

166
00:07:22.880 --> 00:07:24.960
and I suspect it's happening with AI.

167
00:07:24.960 --> 00:07:28.080
The catch is, we don't know
where we are on that curve yet.

168
00:07:28.080 --> 00:07:29.680
You never do until it flattens out.

169
00:07:29.680 --> 00:07:31.760
It's possible that this
is as far as we go,

170
00:07:31.760 --> 00:07:34.640
that the best we'll get is that
an AI can generate a fake script,

171
00:07:34.640 --> 00:07:35.600
and it's a novelty.

172
00:07:36.320 --> 00:07:38.480
But if someone can take that and
add the requirement that

173
00:07:38.480 --> 00:07:40.160
what it writes has to be true,

174
00:07:40.160 --> 00:07:42.640
that it has to crossreference
with history books and encyclopedias,

175
00:07:42.640 --> 00:07:46.480
like IBM's Watson did to win Jeopardy
ten years ago...

176
00:07:47.200 --> 00:07:49.920
well, if that happens,
my job will get a lot easier.

177
00:07:50.560 --> 00:07:54.800
Until then, I'm going to ask it for another
thousand ideas and see what it comes up with.

178
00:07:57.040 --> 00:07:59.760
I thought I should check,
and, yeah. It's fine.

179
00:08:00.720 --> 00:08:11.840
It's still there.