1
00:00:05,720 --> 00:00:07,720
Speaker 1: Hey, welcome to Stuff to Blow Your Mind. My name

2
00:00:07,760 --> 00:00:10,760
Speaker 1: is Robert Lamb and I'm Joe McCormick, and it's Saturday.

3
00:00:10,800 --> 00:00:12,640
Speaker 1: Time to go into the vault. This time we are

4
00:00:13,119 --> 00:00:15,040
Speaker 1: uh oh. This is going to be part one of

5
00:00:15,160 --> 00:00:17,640
Speaker 1: the pair of episodes we did last year called flate

6
00:00:17,760 --> 00:00:23,040
Speaker 1: Us x Mockina. This one originally published March nineteen, and

7
00:00:23,079 --> 00:00:26,320
Speaker 1: it's about why it's so funny when machines fail, Yes,

8
00:00:26,480 --> 00:00:29,920
Speaker 1: fart out of the machine. What is that? Uh, you

9
00:00:29,960 --> 00:00:32,440
Speaker 1: know you're maybe wondering what that means. It's basically like

10
00:00:32,800 --> 00:00:36,440
Speaker 1: what happens when when AI, despite all of its abilities,

11
00:00:36,520 --> 00:00:39,600
Speaker 1: creates something that is just a flop? And why is

12
00:00:39,640 --> 00:00:41,640
Speaker 1: it so amusing? And what can we learn from that?

13
00:00:41,720 --> 00:00:44,720
Speaker 1: What does it reveal about artificial intelligence in general? Well,

14
00:00:44,760 --> 00:00:50,320
Speaker 1: if you want to find out, we'll listen to this episode.

15
00:00:50,840 --> 00:00:53,680
Speaker 1: Welcome to Stuff to Blow your Mind from how Stop

16
00:00:53,720 --> 00:01:02,000
Speaker 1: Works dot com. Hey you, welcome to Stuff to Blow

17
00:01:02,000 --> 00:01:04,880
Speaker 1: your Mind. My name is Robert Lamb and I'm Joe McCormick,

18
00:01:04,959 --> 00:01:06,760
Speaker 1: and today we're gonna be talking about one of my

19
00:01:06,840 --> 00:01:11,839
Speaker 1: favorite comedy subjects. What's funny about the way that machines fail?

20
00:01:12,440 --> 00:01:14,000
Speaker 1: And here's just heads up. This is gonna be a

21
00:01:14,040 --> 00:01:16,680
Speaker 1: two part episode because we ended up going super long.

22
00:01:16,800 --> 00:01:19,360
Speaker 1: Like the machines, we can't stop. So I want to

23
00:01:19,400 --> 00:01:23,720
Speaker 1: start with a particular example, probably my favorite funny thing

24
00:01:23,840 --> 00:01:28,840
Speaker 1: on the Internet these days, the hilarious almost successes of

25
00:01:29,080 --> 00:01:35,000
Speaker 1: artificial intelligence trying to generate examples of human language almost

26
00:01:35,040 --> 00:01:38,360
Speaker 1: but not quite human. Yes, I don't know why this

27
00:01:38,480 --> 00:01:41,360
Speaker 1: does it for me, but aside from Highlander to the Quickening,

28
00:01:41,440 --> 00:01:45,759
Speaker 1: pretty much nothing makes me laugh harder than language generated

29
00:01:45,760 --> 00:01:49,120
Speaker 1: by artificial neural networks and machine learning. And we'll explain

30
00:01:49,160 --> 00:01:51,600
Speaker 1: a little bit more about exactly how this works in

31
00:01:51,640 --> 00:01:54,040
Speaker 1: a minute, but first I just thought we should look

32
00:01:54,080 --> 00:01:56,440
Speaker 1: at a few examples of what this is like. If

33
00:01:56,440 --> 00:01:59,480
Speaker 1: you're on the Internet, you've probably encountered this at some point,

34
00:02:00,280 --> 00:02:03,560
Speaker 1: because it's become popular in the past few years, especially

35
00:02:03,640 --> 00:02:06,160
Speaker 1: for its comedic value. If you're not on the internet, boy,

36
00:02:06,160 --> 00:02:08,840
Speaker 1: are you in for a treat um by far I

37
00:02:08,840 --> 00:02:11,240
Speaker 1: would say the best of this stuff I've come across

38
00:02:11,440 --> 00:02:15,000
Speaker 1: is traceable back to the blog of a person named

39
00:02:15,120 --> 00:02:18,080
Speaker 1: Janelle Shane, who I think her day job as she

40
00:02:18,120 --> 00:02:21,800
Speaker 1: works in industrial researching optics, but as a hobby she

41
00:02:21,919 --> 00:02:26,239
Speaker 1: trains neural networks with text based data, uh, text based

42
00:02:26,320 --> 00:02:31,120
Speaker 1: data sets to spit out these amazing simulations of types

43
00:02:31,160 --> 00:02:34,160
Speaker 1: of human language, and so they'll they'll be in categories

44
00:02:34,200 --> 00:02:38,160
Speaker 1: like it. She gets an AI program to write recipes

45
00:02:38,200 --> 00:02:41,120
Speaker 1: for foods or to come up with the names of

46
00:02:41,320 --> 00:02:43,959
Speaker 1: paint colors or something. And the way of course you

47
00:02:43,960 --> 00:02:45,920
Speaker 1: would do this is and we'll explain more of the

48
00:02:45,919 --> 00:02:48,680
Speaker 1: details in a bit, but you'd train it on existing

49
00:02:48,840 --> 00:02:51,880
Speaker 1: names of paint colors, or you train it on existing

50
00:02:51,919 --> 00:02:54,280
Speaker 1: recipes of food. Right. So the end result here is

51
00:02:54,280 --> 00:02:57,919
Speaker 1: you essentially have a machine trying to human but not

52
00:02:58,080 --> 00:03:01,720
Speaker 1: quite pulling it off, yes, but but doing so in

53
00:03:01,880 --> 00:03:05,360
Speaker 1: such hilarious fashion. Yeah, so you should absolutely look up

54
00:03:05,440 --> 00:03:08,880
Speaker 1: Janell Shane's blog. It's called AI weirdness dot com. It's

55
00:03:08,919 --> 00:03:11,400
Speaker 1: reliably such a source of joy. But I want to

56
00:03:11,440 --> 00:03:13,359
Speaker 1: start in fact, I will say that if you were

57
00:03:13,480 --> 00:03:15,400
Speaker 1: if you're scratching your head and you're thinking, oh, didn't

58
00:03:15,440 --> 00:03:19,400
Speaker 1: I see something hilarious uh in this vein recently, there's

59
00:03:19,520 --> 00:03:22,280
Speaker 1: there's a there's a high probability that it originated from

60
00:03:22,400 --> 00:03:26,120
Speaker 1: AI weirdness dot com. Yes, that that blog is just awesome,

61
00:03:26,360 --> 00:03:28,000
Speaker 1: But I wanted to look at a few examples of

62
00:03:28,040 --> 00:03:30,280
Speaker 1: what this is like. So one thing is consider the work,

63
00:03:30,560 --> 00:03:33,560
Speaker 1: uh that that Jenell Shane did training in neural network

64
00:03:33,639 --> 00:03:36,080
Speaker 1: to come up with names for D and D spells.

65
00:03:36,840 --> 00:03:39,400
Speaker 1: So you take the Dungeons and Dragons manual and you'd

66
00:03:39,440 --> 00:03:42,240
Speaker 1: feed in all the actual names of spells to the

67
00:03:42,240 --> 00:03:44,280
Speaker 1: sceneral network, so it gets a sense of what these

68
00:03:44,280 --> 00:03:46,640
Speaker 1: things are like, and then it tries to come up

69
00:03:46,680 --> 00:03:49,360
Speaker 1: with similar types of names on its own. Now, to

70
00:03:49,360 --> 00:03:51,600
Speaker 1: give everybody just a quick idea first of what actual

71
00:03:51,720 --> 00:03:56,160
Speaker 1: dungeons and dragon spells are named. You have everything from

72
00:03:56,320 --> 00:04:01,800
Speaker 1: Magic Missile or Crown of Madness to Evard's Black Tentacles

73
00:04:01,960 --> 00:04:04,200
Speaker 1: or anytime there's a wizard name in there, you know

74
00:04:04,240 --> 00:04:07,160
Speaker 1: you're in for some good stuff. Um, well, there's you know,

75
00:04:07,160 --> 00:04:10,680
Speaker 1: stuff like glyph of Warding or what's the what's the

76
00:04:10,720 --> 00:04:13,320
Speaker 1: one I'm trying to think of. Oh, stuff like Leo

77
00:04:13,360 --> 00:04:19,000
Speaker 1: Moon's Secret Chest or Leoman's Tiny Hut. That's another great one. Well,

78
00:04:19,040 --> 00:04:21,120
Speaker 1: so here are the ones that it came up with.

79
00:04:21,440 --> 00:04:25,760
Speaker 1: How about selections like Mr of Light, Confusing, Storm of

80
00:04:25,800 --> 00:04:32,400
Speaker 1: the Giffling, Song of Goom, Song of the Darn, Ward

81
00:04:32,520 --> 00:04:36,760
Speaker 1: of Snay to the pooda beast, primal rear. You've got

82
00:04:36,839 --> 00:04:39,720
Speaker 1: to watch out for primal rear. Someone storm bear, Now,

83
00:04:39,720 --> 00:04:42,760
Speaker 1: that one sounds legitimate. That's one of the beauties of

84
00:04:42,800 --> 00:04:47,080
Speaker 1: these exercises is is when there's one that either almost

85
00:04:47,160 --> 00:04:50,120
Speaker 1: works or actually does work. Because I think some in

86
00:04:50,279 --> 00:04:52,840
Speaker 1: storm Bear, I can I can easily imagine. I can

87
00:04:52,880 --> 00:04:57,480
Speaker 1: describe the storm bear blasting out of the portal and

88
00:04:57,760 --> 00:05:00,360
Speaker 1: rushing into combat on behalf of your you know, your

89
00:05:00,440 --> 00:05:04,240
Speaker 1: your storm mage. Yeah, it's almost kind of effortlessly evocative.

90
00:05:05,080 --> 00:05:07,720
Speaker 1: A divine boom. How about that? That one sounds pretty

91
00:05:07,760 --> 00:05:11,119
Speaker 1: good too, Soul of the bill. Now, now we're flying

92
00:05:11,160 --> 00:05:14,640
Speaker 1: back down the hill. Farc mate, how about a charm

93
00:05:14,720 --> 00:05:20,839
Speaker 1: of the CODs, death of the Sun. Okay that that's

94
00:05:21,080 --> 00:05:23,280
Speaker 1: that would have to be a high level spell. But okay, okay,

95
00:05:23,360 --> 00:05:27,200
Speaker 1: three more greater flick. Okay, that's that's just a can

96
00:05:27,240 --> 00:05:30,200
Speaker 1: trip here, that's just a flick, A magical flick of

97
00:05:30,240 --> 00:05:34,599
Speaker 1: the ear curse clam. I like it. Dav Ing fire

98
00:05:35,640 --> 00:05:38,800
Speaker 1: is confusing. Now, while these phrases I think are mostly

99
00:05:39,000 --> 00:05:42,080
Speaker 1: funny on their own, I think they're probably even funnier

100
00:05:42,120 --> 00:05:44,479
Speaker 1: if you're an actual D and D player, because you

101
00:05:44,560 --> 00:05:47,600
Speaker 1: not only get the pleasure of the nonsense words and

102
00:05:47,640 --> 00:05:50,839
Speaker 1: the you know, the syllables that seem out of place

103
00:05:50,880 --> 00:05:54,200
Speaker 1: in their context, like Dave does not really seem to

104
00:05:54,200 --> 00:05:56,679
Speaker 1: go very well with some kind of magical fire spell.

105
00:05:57,440 --> 00:05:59,440
Speaker 1: But if you actually play D and D you probably

106
00:05:59,480 --> 00:06:02,880
Speaker 1: also get some humor from just like seeing the little

107
00:06:02,920 --> 00:06:06,560
Speaker 1: resonances that these spells have with actual spells that you

108
00:06:06,600 --> 00:06:10,320
Speaker 1: would recognize. I now want to run a gaming session

109
00:06:10,360 --> 00:06:13,800
Speaker 1: where there's some sort of warp effect in place where

110
00:06:13,880 --> 00:06:18,600
Speaker 1: suddenly all the magic users are forced to use spells

111
00:06:18,600 --> 00:06:21,120
Speaker 1: from this list and they don't know what they're going

112
00:06:21,160 --> 00:06:24,400
Speaker 1: to exactly do until they cast them. Even better, what

113
00:06:24,480 --> 00:06:27,120
Speaker 1: if they had to What if your character suddenly had

114
00:06:27,160 --> 00:06:30,279
Speaker 1: amnesia and then had to act as if they had

115
00:06:30,400 --> 00:06:33,680
Speaker 1: bios that were also generated by an artificial neural network.

116
00:06:33,800 --> 00:06:35,919
Speaker 1: That's right, because AI weirdness dot Com also has a

117
00:06:35,960 --> 00:06:39,679
Speaker 1: wonderful piece titled D and D Character Bios. Now making

118
00:06:39,720 --> 00:06:42,479
Speaker 1: slightly more sense. In fact, I would say this post

119
00:06:42,640 --> 00:06:44,720
Speaker 1: from I think this was last week or something we

120
00:06:44,720 --> 00:06:46,840
Speaker 1: were looking at. I think this inspired me to want

121
00:06:46,839 --> 00:06:48,960
Speaker 1: to do this episode. Well, we should read a couple

122
00:06:49,000 --> 00:06:53,520
Speaker 1: of these. Okay, I'll read the first one here quote.

123
00:06:54,000 --> 00:06:58,479
Speaker 1: Frick found his old family's fortune and his curiosity, and

124
00:06:58,560 --> 00:07:00,640
Speaker 1: he went to a small city to see if he

125
00:07:00,680 --> 00:07:04,440
Speaker 1: could find a work in the goldfish. He heard stories

126
00:07:04,520 --> 00:07:07,719
Speaker 1: of a goldfish, a goldfish, a sea monster, and a

127
00:07:07,760 --> 00:07:10,520
Speaker 1: silver fish, a sea monster and a ship that was

128
00:07:10,560 --> 00:07:14,440
Speaker 1: a ship of exploration. The ship was full of fish

129
00:07:14,480 --> 00:07:18,520
Speaker 1: and evil some treasure, but it was not to be

130
00:07:19,120 --> 00:07:22,320
Speaker 1: when Frick found the ship. He rushed back and found

131
00:07:22,360 --> 00:07:26,080
Speaker 1: the ship full of treasure and full of fish. He

132
00:07:26,160 --> 00:07:29,760
Speaker 1: wanted to be a pirate and fight it. Now, I

133
00:07:29,840 --> 00:07:32,160
Speaker 1: like that because there's a there's a lot of silliness

134
00:07:32,360 --> 00:07:35,560
Speaker 1: in there, but it does have the basic shape of

135
00:07:35,560 --> 00:07:38,000
Speaker 1: a bio and and and and and it's weirdness, and

136
00:07:38,040 --> 00:07:42,800
Speaker 1: it's the stuff that is more nonsensical actually feels suitably

137
00:07:43,040 --> 00:07:46,800
Speaker 1: magical and fantastic. You know that there's this this fish

138
00:07:46,840 --> 00:07:49,000
Speaker 1: that's not a fish, that's also a ship that's full

139
00:07:49,040 --> 00:07:51,600
Speaker 1: of treasure and fish, and he's going to become a

140
00:07:51,600 --> 00:07:54,600
Speaker 1: pirate and fight it. And I think you could run

141
00:07:54,640 --> 00:07:57,080
Speaker 1: with that. This is one of the fascinating things about

142
00:07:57,120 --> 00:08:00,840
Speaker 1: neural net generated text is that it often has the

143
00:08:00,880 --> 00:08:04,840
Speaker 1: format of what you're going for, correct it just doesn't

144
00:08:04,880 --> 00:08:09,440
Speaker 1: have the sense of it correct like it will esthetically

145
00:08:09,600 --> 00:08:12,679
Speaker 1: and in shape be very much like what you're looking for,

146
00:08:13,000 --> 00:08:16,840
Speaker 1: but keywords do not make any sense at all. Then again,

147
00:08:16,880 --> 00:08:19,320
Speaker 1: another one that I thought was kind of interesting in

148
00:08:19,400 --> 00:08:22,400
Speaker 1: what it showed about common themes in in D and

149
00:08:22,480 --> 00:08:25,280
Speaker 1: D character bios, or I guess maybe fantasy more generally,

150
00:08:25,800 --> 00:08:29,240
Speaker 1: was one bio that included the lines um the Orc

151
00:08:29,320 --> 00:08:32,040
Speaker 1: Warlock was captured and killed by a group of Orcs.

152
00:08:32,040 --> 00:08:34,800
Speaker 1: He was imprisoned and forced to work for a giant tome.

153
00:08:35,200 --> 00:08:37,520
Speaker 1: He was imprisoned and imprisoned for a while until he

154
00:08:37,600 --> 00:08:40,320
Speaker 1: was rescued by a group of adventurers. He was imprisoned

155
00:08:40,360 --> 00:08:42,440
Speaker 1: and imprisoned for a while until he was rescued by

156
00:08:42,440 --> 00:08:44,440
Speaker 1: a group of adventurers who were looking for a group

157
00:08:44,480 --> 00:08:47,640
Speaker 1: of adventurers to help eradicate the orch is ferocity. See

158
00:08:47,679 --> 00:08:49,880
Speaker 1: now that's wonderful. I actually really like that, And it's

159
00:08:49,880 --> 00:08:52,920
Speaker 1: a little you know, nonsensical. But uh, first of all,

160
00:08:53,040 --> 00:08:56,280
Speaker 1: you know, being an orc is hard. It's it's it's

161
00:08:56,280 --> 00:09:00,560
Speaker 1: a warlike h world for the orc. But but then

162
00:09:00,840 --> 00:09:03,080
Speaker 1: the idea that he's working for a giant tone the

163
00:09:03,160 --> 00:09:06,320
Speaker 1: idea that there's like a magical book employing him. And

164
00:09:06,400 --> 00:09:10,640
Speaker 1: the repetitive imprisonment. Yeah, it's a hard knock life. You're

165
00:09:10,679 --> 00:09:13,959
Speaker 1: always getting imprisoned and then escaping and encountering a band

166
00:09:14,000 --> 00:09:17,199
Speaker 1: of adventurers. Yeah, and working for weird magic books. So

167
00:09:17,520 --> 00:09:19,640
Speaker 1: that one, that one works. I think, of course. Another

168
00:09:19,679 --> 00:09:21,800
Speaker 1: great one on this page is an injury that just

169
00:09:21,840 --> 00:09:25,920
Speaker 1: seems to devolve into endless like dozens of repetitions of

170
00:09:26,040 --> 00:09:30,360
Speaker 1: variations of the phrase big cat. Yes, little did he know.

171
00:09:30,600 --> 00:09:34,360
Speaker 1: During the monastery's course of time, when the monastery's training

172
00:09:34,440 --> 00:09:37,920
Speaker 1: and growth was complete, his mother told big cat, big cat,

173
00:09:38,000 --> 00:09:40,200
Speaker 1: and big cat to drive their own path and test

174
00:09:40,280 --> 00:09:43,320
Speaker 1: big cat with big cats, army of big cats, army

175
00:09:43,360 --> 00:09:46,800
Speaker 1: of big cats and big cats and big cats, big cats,

176
00:09:46,840 --> 00:09:48,960
Speaker 1: big cat, big cat, and it goes on, yes for

177
00:09:49,040 --> 00:09:53,280
Speaker 1: something for for many many lines. I'm trying to picture

178
00:09:53,360 --> 00:09:56,720
Speaker 1: big cat, though I can't. I don't actually go to

179
00:09:56,760 --> 00:09:59,560
Speaker 1: like tiger or lie in there. I think more of

180
00:09:59,640 --> 00:10:03,400
Speaker 1: a a kind of mental power, great basilisk type of

181
00:10:03,559 --> 00:10:06,200
Speaker 1: fat house cat. I think that would work. I also

182
00:10:06,240 --> 00:10:09,240
Speaker 1: can't help but think of cheshire cats, and of course

183
00:10:09,280 --> 00:10:12,200
Speaker 1: cat Bus from Totoro is being sort of in the

184
00:10:12,280 --> 00:10:16,839
Speaker 1: vein of big Cat, Big Cat, big Cat. Um. How

185
00:10:16,880 --> 00:10:19,439
Speaker 1: about a neural networks trained on a corpus of more

186
00:10:19,480 --> 00:10:23,920
Speaker 1: than forty three thousand questions and answer style jokes. I'll

187
00:10:23,960 --> 00:10:26,440
Speaker 1: just read one of them. Why do you call a

188
00:10:26,480 --> 00:10:33,840
Speaker 1: pastor across the road? He take the chicken? Yeah. Another

189
00:10:34,160 --> 00:10:37,880
Speaker 1: great one from AI weirdness dot Com was a neural

190
00:10:37,920 --> 00:10:41,079
Speaker 1: network designs Halloween costumes. Okay, so you just feeded a

191
00:10:41,080 --> 00:10:43,240
Speaker 1: bunch of Halloween costume names. Yeah. And this one, this

192
00:10:43,280 --> 00:10:45,560
Speaker 1: one was tremendous fun. I remember when it first came.

193
00:10:45,600 --> 00:10:49,080
Speaker 1: I just made me laugh so hard. Highlights include uh,

194
00:10:49,160 --> 00:10:55,520
Speaker 1: Saxy Pumpkins, Um, Disco Monster, Spartan Gandalf. I really like

195
00:10:55,600 --> 00:10:59,480
Speaker 1: that one. Starfleet Shark, Yeah, that's good. Um, let's see

196
00:10:59,559 --> 00:11:03,960
Speaker 1: Martian Devil. That was too believable. Panda clam though, that's

197
00:11:04,000 --> 00:11:09,200
Speaker 1: that's pretty good. Shark cow. That's very interesting given some

198
00:11:09,280 --> 00:11:12,160
Speaker 1: of our past discussions on was it a sharp cow

199
00:11:12,240 --> 00:11:13,800
Speaker 1: that we were discussing, and I believe it was. Oh,

200
00:11:13,920 --> 00:11:17,520
Speaker 1: that was That's Daniel Dennett's thought experiment, exposing what he

201
00:11:17,559 --> 00:11:22,319
Speaker 1: believes to be flaws in Donald David's Donald Davidson's Swampman

202
00:11:22,800 --> 00:11:25,760
Speaker 1: thought experiment it was the cow shark, where he says,

203
00:11:26,360 --> 00:11:29,120
Speaker 1: is it actually meaningful if you posit a cow shark

204
00:11:29,200 --> 00:11:31,480
Speaker 1: to ask whether it's a cow or a shark? A

205
00:11:31,520 --> 00:11:36,360
Speaker 1: little accidental thought experiment here from this list. Uh, you'll

206
00:11:36,400 --> 00:11:39,920
Speaker 1: find less thought experimentation in such entries though, as snape

207
00:11:39,960 --> 00:11:50,760
Speaker 1: scarce snape scarecrow, or or how about lady garbage, Lady garbage.

208
00:11:50,800 --> 00:11:53,040
Speaker 1: Lady garbage is good. So yeah, there are a lot

209
00:11:53,040 --> 00:11:57,760
Speaker 1: of great, great entries on that list. Okay, one last

210
00:11:57,800 --> 00:12:00,160
Speaker 1: thing from General Shane's work, I just must mean in

211
00:12:00,320 --> 00:12:03,319
Speaker 1: some some titles for recipes that she wrote a script

212
00:12:03,400 --> 00:12:07,640
Speaker 1: to come up with. Um. These include chocolate pickle sauce,

213
00:12:07,800 --> 00:12:16,480
Speaker 1: whole chicken cookies, salmon beef style chicken bottom, artichoke, gelatin dogs,

214
00:12:16,520 --> 00:12:21,760
Speaker 1: and crock pot cold water. These are where any of

215
00:12:21,800 --> 00:12:24,920
Speaker 1: these attempted these recipes. Uh well, actually at one point

216
00:12:24,960 --> 00:12:27,600
Speaker 1: she so those were just names at some point. They

217
00:12:27,640 --> 00:12:30,760
Speaker 1: don't have these selected, but she did have it based

218
00:12:30,800 --> 00:12:34,720
Speaker 1: on a corpus of recipes generate new recipes which are

219
00:12:34,760 --> 00:12:38,680
Speaker 1: just nightmares, you know, like huge lists of ingredients that

220
00:12:38,679 --> 00:12:42,160
Speaker 1: would be like, you know, a third cup smore goals,

221
00:12:42,400 --> 00:12:47,280
Speaker 1: you know, a cup of Horseradishuh. Then there was one

222
00:12:47,320 --> 00:12:49,440
Speaker 1: website I said, I don't remember who actually did it

223
00:12:49,520 --> 00:12:52,120
Speaker 1: might have been Super Deluxe. There's somebody made a video

224
00:12:52,559 --> 00:12:55,480
Speaker 1: where they took one of these AI generated recipes and

225
00:12:55,559 --> 00:12:58,679
Speaker 1: just literally made it. Now, they had some trouble because

226
00:12:58,720 --> 00:13:01,160
Speaker 1: some of the ingredients were not real words, they're just

227
00:13:01,480 --> 00:13:04,080
Speaker 1: you know, yeah, like, so they had to I guess

228
00:13:04,080 --> 00:13:07,160
Speaker 1: substitute something or they put in instead of smart goals

229
00:13:07,200 --> 00:13:09,520
Speaker 1: it was coco powder or something. But they ended up

230
00:13:09,559 --> 00:13:13,600
Speaker 1: with these like, uh, pasta shells. I think that had

231
00:13:13,679 --> 00:13:16,520
Speaker 1: like chocolate and stuff all over them. But anyway, so

232
00:13:16,960 --> 00:13:19,400
Speaker 1: that I think that's currently one of the best websites

233
00:13:19,440 --> 00:13:21,760
Speaker 1: on the internet. Go to AI weirdness dot com if

234
00:13:21,800 --> 00:13:25,120
Speaker 1: you want more joy and humor. But um, but I

235
00:13:25,160 --> 00:13:28,960
Speaker 1: was wondering, why is this the funniest thing out there

236
00:13:29,040 --> 00:13:32,320
Speaker 1: for me right now? What is so inherently funny about

237
00:13:32,360 --> 00:13:37,679
Speaker 1: the ways machines create things that are sort of like

238
00:13:37,920 --> 00:13:41,160
Speaker 1: real language, just close enough to be in the zone

239
00:13:41,200 --> 00:13:43,320
Speaker 1: where they are funny, but at the same time far

240
00:13:43,440 --> 00:13:46,320
Speaker 1: enough off that they're they're totally hilarious. And I think

241
00:13:46,360 --> 00:13:49,880
Speaker 1: that there are at least two parts that make this

242
00:13:49,880 --> 00:13:53,120
Speaker 1: this stuff so golden one is that there's something inherently

243
00:13:53,160 --> 00:13:58,040
Speaker 1: funny about machines trying to behave like humans and failing. Specifically,

244
00:13:58,040 --> 00:14:01,280
Speaker 1: it's the ways that they fail, and will definitely explore

245
00:14:01,320 --> 00:14:04,679
Speaker 1: this more throughout the episode. But the way that they

246
00:14:04,679 --> 00:14:10,280
Speaker 1: failed demonstrates a kind of pristine, oblivious quality of stupidity.

247
00:14:10,320 --> 00:14:13,960
Speaker 1: It's like a kind of platonic stupidity isolated from the

248
00:14:14,000 --> 00:14:18,520
Speaker 1: ability to appreciate itself. It's like the funniness of you know,

249
00:14:18,559 --> 00:14:22,360
Speaker 1: watching automatic doors repeatedly trying to close on something that's

250
00:14:22,400 --> 00:14:25,160
Speaker 1: blocking them, Like that that's not itself so funny, but

251
00:14:25,200 --> 00:14:28,280
Speaker 1: you see an inkling of the same thing there that

252
00:14:28,360 --> 00:14:32,000
Speaker 1: comes through in these AI generated texts. But then we

253
00:14:32,080 --> 00:14:35,080
Speaker 1: sort of mentioned this earlier. It's also funny because it

254
00:14:35,120 --> 00:14:39,040
Speaker 1: tends to reveal something interesting about the human culture game

255
00:14:39,160 --> 00:14:41,760
Speaker 1: that it's trying to play like. It brings this sort

256
00:14:41,760 --> 00:14:45,880
Speaker 1: of cold objectivity to phenomena that we don't necessarily always bring,

257
00:14:46,280 --> 00:14:50,360
Speaker 1: and it can identify and awkwardly replicate trends and behaviors

258
00:14:50,360 --> 00:14:52,320
Speaker 1: that we might fail to notice in the same way

259
00:14:52,320 --> 00:14:55,680
Speaker 1: that like you might notice funny things that are actually

260
00:14:55,720 --> 00:14:59,000
Speaker 1: present in the names of D and D spells by

261
00:14:59,040 --> 00:15:02,760
Speaker 1: watching a computer or try and replicate them. Yeah, yeah,

262
00:15:02,800 --> 00:15:05,400
Speaker 1: I think these are these are these are two strong reasons. Okay,

263
00:15:05,400 --> 00:15:07,280
Speaker 1: we're going to take a quick break. We'll be right

264
00:15:07,320 --> 00:15:13,280
Speaker 1: back with more. Alright, we're back now with all things humor,

265
00:15:13,320 --> 00:15:14,880
Speaker 1: And of course we'll get into this in this episode

266
00:15:14,880 --> 00:15:17,840
Speaker 1: as we continue to discuss what humor is and then

267
00:15:17,920 --> 00:15:22,600
Speaker 1: why machines in some cases achieve it, like the absurd

268
00:15:23,480 --> 00:15:29,200
Speaker 1: is is funny, Like the absurdity is hilarious, And it

269
00:15:29,280 --> 00:15:31,440
Speaker 1: seems it seems like we're in a situation where a

270
00:15:31,440 --> 00:15:34,400
Speaker 1: lot of times, some of these, especially these neural net situations,

271
00:15:34,640 --> 00:15:40,280
Speaker 1: where we were accidentally creating absurdity engines, we're creating machines

272
00:15:40,480 --> 00:15:46,200
Speaker 1: that that produce absurdity. And uh, you know, well, you know,

273
00:15:46,200 --> 00:15:48,480
Speaker 1: what can you say, like why is why is something

274
00:15:48,480 --> 00:15:52,280
Speaker 1: that is absurd funny? You know, because we'll get into

275
00:15:52,280 --> 00:15:54,040
Speaker 1: all that in a bit. But but sometimes I feel

276
00:15:54,040 --> 00:15:57,920
Speaker 1: like the answer might be it's funny because it's funny. Well, yeah,

277
00:15:57,920 --> 00:16:01,520
Speaker 1: I mean, there's definitely an ineffable quality of of of humor.

278
00:16:01,640 --> 00:16:04,320
Speaker 1: That is one of the reasons there's so many different

279
00:16:04,360 --> 00:16:06,960
Speaker 1: theories of humor. Then, as I said, we'll explore them

280
00:16:06,960 --> 00:16:09,880
Speaker 1: more as the episode goes on. But um, yeah, it's

281
00:16:09,920 --> 00:16:13,040
Speaker 1: obviously something that's really hard to to narrow down and

282
00:16:13,040 --> 00:16:15,080
Speaker 1: put your finger on. It's that there seemed to be

283
00:16:15,120 --> 00:16:18,880
Speaker 1: all these strange, conflicting, overlapping reasons we find things funny.

284
00:16:19,560 --> 00:16:24,000
Speaker 1: But I do feel like this strain of machine humor

285
00:16:24,080 --> 00:16:27,400
Speaker 1: machine failure humor being one of the funniest types of

286
00:16:27,480 --> 00:16:32,480
Speaker 1: humor UH is bigger than just the AI text. Because

287
00:16:32,520 --> 00:16:35,360
Speaker 1: it got me I started thinking about what are some

288
00:16:35,440 --> 00:16:38,280
Speaker 1: of the funniest scenes in movies I can think of?

289
00:16:38,680 --> 00:16:40,520
Speaker 1: And when I tried to think about that, I can't

290
00:16:40,520 --> 00:16:43,640
Speaker 1: help but think of the dark humor of the hyper

291
00:16:43,760 --> 00:16:48,840
Speaker 1: violent boardroom scene in RoboCop with ED two o nine.

292
00:16:49,080 --> 00:16:51,000
Speaker 1: I would argue one of the I mean, it's not

293
00:16:51,080 --> 00:16:54,680
Speaker 1: for children. This is like a hyper violent, horrible UH scene,

294
00:16:54,680 --> 00:16:57,080
Speaker 1: but it's also in in a in a morbid way,

295
00:16:57,200 --> 00:16:59,640
Speaker 1: one of the funniest scenes I think in film history,

296
00:17:00,640 --> 00:17:02,640
Speaker 1: like and so why is it so funny? I think

297
00:17:02,640 --> 00:17:06,520
Speaker 1: it hinges on parallels between humans and machines and the

298
00:17:06,520 --> 00:17:09,560
Speaker 1: scene and the similarities in the ways they fail. So

299
00:17:10,080 --> 00:17:13,159
Speaker 1: brief recap of the scene is, uh. You know Ronnie

300
00:17:13,200 --> 00:17:17,520
Speaker 1: Cox plays this, you know self serious, bloviating businessman who's

301
00:17:17,520 --> 00:17:20,359
Speaker 1: proudly proclaiming, you know, I've got the technology that's the

302
00:17:20,400 --> 00:17:23,880
Speaker 1: future of policing. And he brings out this robot called

303
00:17:23,960 --> 00:17:26,040
Speaker 1: ED two oh nine. He's got these big guns on it,

304
00:17:26,320 --> 00:17:28,119
Speaker 1: and they're saying that it's going to take over the

305
00:17:28,160 --> 00:17:31,360
Speaker 1: police force and it's this great new technology. And then

306
00:17:31,400 --> 00:17:35,120
Speaker 1: they demonstrate it on a guy and it malfunctions and uh,

307
00:17:35,400 --> 00:17:37,920
Speaker 1: it tells him to drop his weapon in the demonstration,

308
00:17:38,000 --> 00:17:40,040
Speaker 1: and he does, and it doesn't seem to notice he

309
00:17:40,080 --> 00:17:42,680
Speaker 1: has and then it shoots him like a hundred times,

310
00:17:42,720 --> 00:17:45,760
Speaker 1: just a ridiculous amount of times. And but it's something

311
00:17:45,840 --> 00:17:49,919
Speaker 1: about the way that the people in the room fail

312
00:17:50,040 --> 00:17:51,919
Speaker 1: in the same way the machine does. Like he just

313
00:17:52,000 --> 00:17:56,119
Speaker 1: plows through this this horrible violent encounter, and then afterwards

314
00:17:56,119 --> 00:17:58,280
Speaker 1: somebody in the background is like, can we get a

315
00:17:58,280 --> 00:18:02,160
Speaker 1: paramedic after this guy has been shot like a hundred times?

316
00:18:02,720 --> 00:18:05,000
Speaker 1: As if like the machine, they're just sort of like

317
00:18:05,119 --> 00:18:09,680
Speaker 1: carrying on their like route behaviors without understanding what they're

318
00:18:09,680 --> 00:18:13,239
Speaker 1: doing or thinking about them. Yeah, the end to oh

319
00:18:13,320 --> 00:18:17,080
Speaker 1: nine is such a great design in the original RoboCop

320
00:18:17,359 --> 00:18:22,120
Speaker 1: because it it resembles it has animal qualities to it.

321
00:18:22,119 --> 00:18:26,480
Speaker 1: It looks kind of like a bipedal dinosaur. Uh, And

322
00:18:26,840 --> 00:18:30,200
Speaker 1: yet it's also smooth and abstract in so many ways

323
00:18:30,240 --> 00:18:33,480
Speaker 1: that it looks like either a highly designed piece of technology,

324
00:18:33,480 --> 00:18:35,000
Speaker 1: which of course it is. It has you know, it

325
00:18:35,000 --> 00:18:38,640
Speaker 1: looks like it's a nice piece of stereo equipment. But

326
00:18:38,640 --> 00:18:42,240
Speaker 1: but then it also lacks any additional details, like it's

327
00:18:42,240 --> 00:18:46,320
Speaker 1: almost a silhouette of a lie of an animal. Yes, um,

328
00:18:46,960 --> 00:18:49,600
Speaker 1: and the growls and it growls as well. But then

329
00:18:49,680 --> 00:18:52,000
Speaker 1: also later a really funny scene in the movie is

330
00:18:52,000 --> 00:18:56,560
Speaker 1: the discovery that this horrible violent killer robot is can

331
00:18:56,560 --> 00:18:59,439
Speaker 1: be defeated by stairs. Like you can't use stairs. It

332
00:18:59,480 --> 00:19:03,200
Speaker 1: has these one for like all terrain looking um legs

333
00:19:03,200 --> 00:19:06,480
Speaker 1: that it walks on, and that it can't manipulate stairs

334
00:19:06,520 --> 00:19:09,080
Speaker 1: it falls down them and then it's it's like a

335
00:19:09,600 --> 00:19:13,120
Speaker 1: ridiculous upside down duckling. I mean it's so hilarious too,

336
00:19:13,119 --> 00:19:15,520
Speaker 1: because I mean ed to A nine is highly effective

337
00:19:15,600 --> 00:19:19,840
Speaker 1: in other situations if it's just filling a guy with bullets,

338
00:19:20,280 --> 00:19:25,480
Speaker 1: highly effective, um, just battling RoboCop also highly effective for

339
00:19:25,680 --> 00:19:29,000
Speaker 1: this part um and this falls in line I think

340
00:19:29,000 --> 00:19:32,640
Speaker 1: pretty well with our experiences of machines and of AI.

341
00:19:32,720 --> 00:19:35,879
Speaker 1: We can create highly effective specialist in many areas of

342
00:19:35,920 --> 00:19:39,320
Speaker 1: AI and robotics. So you create a machine that just

343
00:19:39,359 --> 00:19:41,879
Speaker 1: puts bullets into this guy. Uh, you know, it does

344
00:19:41,880 --> 00:19:44,600
Speaker 1: a great job. But when it comes to creating a

345
00:19:44,640 --> 00:19:48,560
Speaker 1: general AI or a machine that can navigate the complex

346
00:19:48,720 --> 00:19:51,679
Speaker 1: natural or even or the human created world such as

347
00:19:51,760 --> 00:19:56,000
Speaker 1: the stairs, Uh, there's continual challenge there. Like that's kind

348
00:19:56,000 --> 00:19:58,160
Speaker 1: of That's that's what a lot of of what's going

349
00:19:58,160 --> 00:20:00,800
Speaker 1: on in robotics and AI is about. Yeah, or the

350
00:20:01,000 --> 00:20:03,919
Speaker 1: just recognizing that it's not actually supposed to shoot the

351
00:20:03,960 --> 00:20:07,720
Speaker 1: board executive during the demonstration exactly. Yeah, but then also

352
00:20:07,800 --> 00:20:11,600
Speaker 1: just the idea that it's it's wonderful, it's something and

353
00:20:11,680 --> 00:20:15,080
Speaker 1: terrible at another thing, that imbalance is often where we

354
00:20:15,119 --> 00:20:17,840
Speaker 1: find a lot a lot of hilarity in other um,

355
00:20:18,280 --> 00:20:21,679
Speaker 1: you know, another comedic stories and bits of fiction or

356
00:20:21,680 --> 00:20:25,439
Speaker 1: just situations like one of my favorite cut scenes of

357
00:20:25,440 --> 00:20:28,800
Speaker 1: all times from Conan the Barbarian. So there's a scene

358
00:20:28,800 --> 00:20:31,040
Speaker 1: in it where it coming this is the Arnold Schwarzenegger original.

359
00:20:31,080 --> 00:20:33,640
Speaker 1: There's it's like a blooper outtake. It's a blueper out take.

360
00:20:33,680 --> 00:20:35,560
Speaker 1: You find it, I don't think most of the special DVDs.

361
00:20:35,560 --> 00:20:39,560
Speaker 1: It's available on YouTube as well. But basically Conan has

362
00:20:39,640 --> 00:20:44,280
Speaker 1: just escaped or been released from servitude and he's running

363
00:20:44,320 --> 00:20:48,520
Speaker 1: across the you know, the wasteland. Essentially, wild dogs are

364
00:20:48,600 --> 00:20:51,399
Speaker 1: chasing him in order to to to eat his flesh.

365
00:20:51,560 --> 00:20:54,119
Speaker 1: He manages in the finished film to scramble up some

366
00:20:54,240 --> 00:20:57,280
Speaker 1: rocks and that begins this new story arc him discovering

367
00:20:57,320 --> 00:21:00,720
Speaker 1: this great old sword. But this is like young corporal

368
00:21:00,760 --> 00:21:03,800
Speaker 1: beef Body, giant Arnold Schwartzen. This is the this is

369
00:21:03,840 --> 00:21:06,680
Speaker 1: the Arnold Schwartzenegger that was so ripped he had to

370
00:21:06,720 --> 00:21:09,679
Speaker 1: like lose muscle so that he could actually hold the

371
00:21:09,720 --> 00:21:13,520
Speaker 1: sword correctly. Uh So there's this out take though, where

372
00:21:13,520 --> 00:21:17,480
Speaker 1: he's running in chains. The dogs are chasing him, the

373
00:21:17,520 --> 00:21:20,720
Speaker 1: movie dogs, and then he's scrambling up the stones and

374
00:21:20,760 --> 00:21:24,320
Speaker 1: the dogs catch him and drag him back down and

375
00:21:24,359 --> 00:21:27,800
Speaker 1: he's just screamed on and and cursing the whole time. Yeah,

376
00:21:27,800 --> 00:21:31,359
Speaker 1: it's I watched it after you linked. It very very funny,

377
00:21:31,400 --> 00:21:32,720
Speaker 1: and you see the p a s come in to

378
00:21:32,760 --> 00:21:36,760
Speaker 1: like get the dogs off the dogs and he's like, yeah,

379
00:21:37,160 --> 00:21:40,120
Speaker 1: it's it's funny because the idea of Conan the barbarian

380
00:21:40,280 --> 00:21:43,879
Speaker 1: or even just prime Arnold failing like this, it's just

381
00:21:44,520 --> 00:21:48,359
Speaker 1: start contrast to actual or perceived strength of the character

382
00:21:48,640 --> 00:21:52,080
Speaker 1: and or individual. And it's also funny because he wasn't

383
00:21:52,119 --> 00:21:55,200
Speaker 1: actually mauled to death by movie dogs, right, of course,

384
00:21:55,200 --> 00:21:58,280
Speaker 1: he wasn't actually seriously harmed in this incident, but he

385
00:21:58,560 --> 00:22:02,359
Speaker 1: was apparently convenienced and had his pride wounded. Right. And

386
00:22:02,400 --> 00:22:04,760
Speaker 1: it will come back to that idea, to the degree

387
00:22:04,920 --> 00:22:09,359
Speaker 1: to which you know, the severity of the outcome um

388
00:22:09,720 --> 00:22:12,680
Speaker 1: comes into play and determining if something is funny or not. Yeah,

389
00:22:12,720 --> 00:22:15,080
Speaker 1: I think I can definitely see what you're talking about

390
00:22:15,119 --> 00:22:19,120
Speaker 1: that the failures of technology are especially funny when there

391
00:22:19,119 --> 00:22:23,119
Speaker 1: are other ways that the technology is highly advanced or

392
00:22:23,200 --> 00:22:26,480
Speaker 1: presented as highly advanced. And as long as like nobody

393
00:22:26,520 --> 00:22:29,919
Speaker 1: of course dies, you know, um. But but I do

394
00:22:30,000 --> 00:22:31,800
Speaker 1: wonder too if there's a darker streak and all of

395
00:22:31,800 --> 00:22:35,320
Speaker 1: this too, you know, something that ties into a deeply

396
00:22:35,520 --> 00:22:39,639
Speaker 1: rooted human disdain for the other, especially for others of species.

397
00:22:40,400 --> 00:22:42,720
Speaker 1: There is a quote from C. S. Lewis's The Lion,

398
00:22:42,880 --> 00:22:45,119
Speaker 1: The Which the Wardrobe and the Four Children. That's my

399
00:22:45,200 --> 00:22:48,040
Speaker 1: son suggested alternate title for it. He's like, like, why

400
00:22:48,040 --> 00:22:50,800
Speaker 1: don't they call it the land which Wardrobe and the

401
00:22:50,840 --> 00:22:53,320
Speaker 1: four children like they're in it too, and they're not

402
00:22:53,400 --> 00:22:56,040
Speaker 1: in the title. Seems like a missed opportunity. But anyway,

403
00:22:56,160 --> 00:22:58,280
Speaker 1: there's this wonderful quote from it that I find I

404
00:22:58,320 --> 00:23:03,080
Speaker 1: found kind of creepy on a recent read of the book. Quote,

405
00:23:03,440 --> 00:23:06,560
Speaker 1: but in general, take my advice when you meet anything

406
00:23:06,640 --> 00:23:09,359
Speaker 1: that is going to be human and isn't yet or

407
00:23:09,560 --> 00:23:12,480
Speaker 1: used to be human once and isn't now or ought

408
00:23:12,520 --> 00:23:15,439
Speaker 1: to be human and isn't you keep your eyes on

409
00:23:15,480 --> 00:23:20,040
Speaker 1: it and feel for your hatchet. Well, that is very creepy. Now,

410
00:23:20,080 --> 00:23:21,879
Speaker 1: I think in the context of the book, this is

411
00:23:21,920 --> 00:23:25,800
Speaker 1: going to be referring to like, you know, possessed objects

412
00:23:25,840 --> 00:23:28,720
Speaker 1: and stuff like pep and magical Basically, yeah, basically, the

413
00:23:28,760 --> 00:23:32,199
Speaker 1: message here was talking animals are cool. You know, you

414
00:23:32,240 --> 00:23:34,720
Speaker 1: can hang out with them in Narnia, but there are

415
00:23:34,800 --> 00:23:37,280
Speaker 1: other things in Narnia that are dangerous, and you can

416
00:23:37,320 --> 00:23:40,320
Speaker 1: tell if they're dangerous based on how human they seem

417
00:23:40,480 --> 00:23:42,600
Speaker 1: they're they're trying to be or used to be. Well,

418
00:23:42,600 --> 00:23:45,600
Speaker 1: that's one thing in a in a fantasy context, in

419
00:23:45,600 --> 00:23:48,720
Speaker 1: in a science fiction or even just a real technology context.

420
00:23:48,800 --> 00:23:51,199
Speaker 1: That's that's a different thing entirely, and starts making you

421
00:23:51,240 --> 00:23:56,000
Speaker 1: think about uh, well, you know, fear of advancing technology

422
00:23:56,040 --> 00:24:00,280
Speaker 1: mimicking human behavior, fears of AI. Yeah, yeah, yeah. To

423
00:24:00,320 --> 00:24:03,119
Speaker 1: what extent do we delight in the falls and errors

424
00:24:03,160 --> 00:24:06,400
Speaker 1: of inhuman entities because we don't wish to see them succeed?

425
00:24:06,800 --> 00:24:09,840
Speaker 1: You know? We we celebrate that the telltale signs of

426
00:24:09,880 --> 00:24:12,720
Speaker 1: their otherness because we kind of dread the day when

427
00:24:12,720 --> 00:24:15,240
Speaker 1: there will be no way to tell. And I think

428
00:24:15,240 --> 00:24:17,680
Speaker 1: there's a lot there's a strong argument to be made

429
00:24:17,680 --> 00:24:19,880
Speaker 1: that that day will be here sooner than we think,

430
00:24:19,920 --> 00:24:22,800
Speaker 1: and in many respects it already is. We've talked about,

431
00:24:22,960 --> 00:24:26,760
Speaker 1: for instance, robocalls on the show before UM and and

432
00:24:26,920 --> 00:24:31,679
Speaker 1: just how and also chat bots and to it becomes

433
00:24:31,680 --> 00:24:34,960
Speaker 1: frightening when you look at where we are with the technology. Now, UM,

434
00:24:35,119 --> 00:24:37,600
Speaker 1: certainly you get a robocall, it's not going, hopefully not

435
00:24:37,680 --> 00:24:41,520
Speaker 1: going to UM deceive you long term. But I think

436
00:24:41,560 --> 00:24:43,760
Speaker 1: a lot of us are having that experience where you

437
00:24:43,840 --> 00:24:45,720
Speaker 1: pick one of these up and at first you think

438
00:24:45,720 --> 00:24:47,600
Speaker 1: it is a human you are talking to, and then

439
00:24:47,680 --> 00:24:50,639
Speaker 1: you realize that it is not. Well, I think one

440
00:24:50,680 --> 00:24:53,640
Speaker 1: of the funny things about stuff like chat bots, which

441
00:24:53,680 --> 00:24:56,640
Speaker 1: also deal in language, but can be much much more

442
00:24:56,680 --> 00:25:00,240
Speaker 1: convincing than these, uh, these neural networks that generate you know,

443
00:25:00,320 --> 00:25:04,399
Speaker 1: lists of lists of character bios or something. Is that

444
00:25:04,480 --> 00:25:07,600
Speaker 1: the things that are generally more convincing these days are

445
00:25:07,640 --> 00:25:10,879
Speaker 1: programmed I think, with more explicit rules, they tend to

446
00:25:10,920 --> 00:25:13,800
Speaker 1: have more kind of human meddling and exactly what they're

447
00:25:13,840 --> 00:25:16,800
Speaker 1: going to be doing, and by having less freedom to

448
00:25:16,960 --> 00:25:20,920
Speaker 1: be creative and all that and ultimately having less potential,

449
00:25:21,040 --> 00:25:24,720
Speaker 1: they can actually be more kind of narrowly convincing. The

450
00:25:24,960 --> 00:25:27,000
Speaker 1: I think one of the things that's interesting about the

451
00:25:27,240 --> 00:25:31,600
Speaker 1: neural network generated text is that it's not anywhere close yet.

452
00:25:31,800 --> 00:25:33,680
Speaker 1: You know, you can't really use it yet to make

453
00:25:33,720 --> 00:25:36,480
Speaker 1: things where you go like, yeah, that's definitely real human.

454
00:25:36,520 --> 00:25:39,600
Speaker 1: I mean maybe you can in some again narrow scenarios,

455
00:25:39,640 --> 00:25:42,320
Speaker 1: but we like, for instance, if you have something that

456
00:25:42,320 --> 00:25:45,119
Speaker 1: will tell you what your Wu Tang clan name will be.

457
00:25:45,640 --> 00:25:49,239
Speaker 1: You know, they sort of random generation systems which are

458
00:25:49,240 --> 00:25:51,600
Speaker 1: totally different, different thing. Uh, and we'll get into the

459
00:25:51,640 --> 00:25:54,320
Speaker 1: distinction here, but but you know, systems like that just

460
00:25:54,400 --> 00:26:00,560
Speaker 1: through sheer random um matching of words, those can be effective. Yeah,

461
00:26:00,600 --> 00:26:02,879
Speaker 1: but it doesn't mean that it's you know, it's an

462
00:26:03,000 --> 00:26:05,960
Speaker 1: entirely different kettle of fish in terms of of what's

463
00:26:05,960 --> 00:26:08,720
Speaker 1: going on with neural networks and where they seem to

464
00:26:08,720 --> 00:26:10,480
Speaker 1: be going. Well, maybe we should take a quick break

465
00:26:10,520 --> 00:26:12,719
Speaker 1: and then we come back. We can explain just the

466
00:26:12,720 --> 00:26:16,920
Speaker 1: basics of how this these kind of things actually work.

467
00:26:17,680 --> 00:26:19,800
Speaker 1: All right, we're back. So I'm gonna try to do

468
00:26:19,960 --> 00:26:22,880
Speaker 1: the simple version of how a neural network works, because

469
00:26:22,920 --> 00:26:25,200
Speaker 1: if you get in the weeds, obviously, neural networks become

470
00:26:25,240 --> 00:26:28,720
Speaker 1: extremely complicated. I know, I I spent a lot of

471
00:26:28,760 --> 00:26:31,000
Speaker 1: time deep in a bunch of articles trying to understand

472
00:26:31,000 --> 00:26:34,800
Speaker 1: technical details that I'm not actually gonna end up using here. Um.

473
00:26:35,080 --> 00:26:38,240
Speaker 1: So the simple version is, think of a neural network

474
00:26:38,520 --> 00:26:42,439
Speaker 1: as a machine that transforms values. That's it. You know,

475
00:26:42,720 --> 00:26:45,720
Speaker 1: it has values that come in like number variables, and

476
00:26:45,760 --> 00:26:48,159
Speaker 1: then it puts out values at the end. It's like

477
00:26:48,359 --> 00:26:50,760
Speaker 1: it's kind of like the toaster conveyor belt at quiz

478
00:26:50,760 --> 00:26:53,840
Speaker 1: Nos or one of those sandwich shops. You know, you're untoasted.

479
00:26:53,840 --> 00:26:57,520
Speaker 1: Sandwich comes in, toasted sandwich comes out. If everything goes

480
00:26:57,560 --> 00:27:00,080
Speaker 1: according to plan. Now, if it doesn't go according and

481
00:27:00,240 --> 00:27:02,919
Speaker 1: maybe it splits out something that's on fire or something

482
00:27:02,960 --> 00:27:04,959
Speaker 1: that you know, who knows what goes on in there,

483
00:27:05,240 --> 00:27:07,960
Speaker 1: or nothing comes out because the bagels are building up

484
00:27:08,000 --> 00:27:10,960
Speaker 1: inside of it, right exactly. But a neural networks core

485
00:27:11,040 --> 00:27:14,640
Speaker 1: job is to just accept inputs and produce outputs. Okay.

486
00:27:14,880 --> 00:27:18,600
Speaker 1: An example would be image recognition. Their neural networks designed

487
00:27:18,600 --> 00:27:22,120
Speaker 1: to look at an image, a digital image, and come

488
00:27:22,200 --> 00:27:24,479
Speaker 1: up with a text string that says, this is what

489
00:27:24,600 --> 00:27:26,399
Speaker 1: that is. So look at a picture of a dog

490
00:27:26,440 --> 00:27:29,080
Speaker 1: and say that's a dog. And you've probably seen examples

491
00:27:29,119 --> 00:27:31,800
Speaker 1: of this on the web. So you'd have numerical values

492
00:27:31,840 --> 00:27:33,760
Speaker 1: going in. It might be things like, you know, be

493
00:27:33,800 --> 00:27:36,800
Speaker 1: a field of pixels with numerical values for their color

494
00:27:36,920 --> 00:27:41,359
Speaker 1: and placement, and then it would have an output that's, uh, say,

495
00:27:41,440 --> 00:27:44,440
Speaker 1: a string of text, which would actually be like a ranking,

496
00:27:44,600 --> 00:27:48,160
Speaker 1: like the top ranked string of text that matches with

497
00:27:48,400 --> 00:27:51,600
Speaker 1: those pixels in that configuration. But so the question is

498
00:27:51,640 --> 00:27:53,880
Speaker 1: what's going on inside the machine? How does it turn

499
00:27:54,040 --> 00:27:57,720
Speaker 1: that input into the correct matching output, and then of

500
00:27:57,720 --> 00:28:00,159
Speaker 1: course how does it fail. Because I've also on this

501
00:28:00,200 --> 00:28:06,040
Speaker 1: tremendously amusing um. Tumbler Um recently changed their guidelines about

502
00:28:06,040 --> 00:28:09,040
Speaker 1: what's acceptable content and what's not and stuff to blow

503
00:28:09,080 --> 00:28:13,840
Speaker 1: your mind has has Slash had a tumbler account recently

504
00:28:13,960 --> 00:28:16,800
Speaker 1: just rebranded it as the Transgenesis tumbler account, but it

505
00:28:16,920 --> 00:28:18,280
Speaker 1: had a lot of old stuff to clear your mind

506
00:28:18,320 --> 00:28:20,439
Speaker 1: content on there. And suddenly I got a page of

507
00:28:20,480 --> 00:28:23,760
Speaker 1: all the things that have been flagged for for potentially

508
00:28:23,880 --> 00:28:26,399
Speaker 1: violating the new terms. And it was amusing because some

509
00:28:26,440 --> 00:28:28,919
Speaker 1: of it was like, okay, well that has a classical

510
00:28:28,960 --> 00:28:31,399
Speaker 1: painting on it, it's got like a you know, it's

511
00:28:31,400 --> 00:28:33,800
Speaker 1: something there's something that might look like nudity, and so

512
00:28:33,840 --> 00:28:37,160
Speaker 1: it got flagged makes sense, or the topic is something

513
00:28:37,200 --> 00:28:40,360
Speaker 1: that is a little too sketchy for them, and they're like, okay,

514
00:28:40,400 --> 00:28:43,160
Speaker 1: that's been flagged by the machine. But the most hilarious

515
00:28:43,160 --> 00:28:46,800
Speaker 1: one was a picture of a baby bat on on

516
00:28:46,880 --> 00:28:52,200
Speaker 1: somebody's palm, and that was flagged as as as likely inappropriate.

517
00:28:52,240 --> 00:28:53,840
Speaker 1: And so I was just trying to figure that one out,

518
00:28:53,840 --> 00:28:56,400
Speaker 1: like what is it about a baby bat did it?

519
00:28:56,640 --> 00:28:59,840
Speaker 1: I mean didn't think this was genitalia or like, like

520
00:29:00,040 --> 00:29:02,560
Speaker 1: what because because I know it's somehow clicked off a

521
00:29:02,640 --> 00:29:06,120
Speaker 1: number of boxes and when and then the automated result

522
00:29:06,280 --> 00:29:08,560
Speaker 1: was no baby bats on tumbler. Well, I guess I

523
00:29:08,880 --> 00:29:10,240
Speaker 1: don't know if you know, but I don't know what

524
00:29:10,320 --> 00:29:13,120
Speaker 1: the mechanism for identifying that is. It might be something

525
00:29:13,160 --> 00:29:15,840
Speaker 1: like this, but yeah that I love seeing that kind

526
00:29:15,840 --> 00:29:18,400
Speaker 1: of failure. And notice that we are laughing now like

527
00:29:18,480 --> 00:29:20,880
Speaker 1: it or we were laughing. It is funny that it

528
00:29:20,920 --> 00:29:23,360
Speaker 1: looks at that and says, I don't know about this bad.

529
00:29:23,480 --> 00:29:25,560
Speaker 1: I think this might be porno. Yeah, I mean then

530
00:29:25,600 --> 00:29:28,400
Speaker 1: the stakes are pretty low. I ultimately one picture of

531
00:29:28,400 --> 00:29:30,840
Speaker 1: a baby bat no longer on a tumbler pitch that

532
00:29:30,880 --> 00:29:33,400
Speaker 1: we don't really use. It doesn't really affect me personally,

533
00:29:33,440 --> 00:29:36,240
Speaker 1: but you could see where this could lead to uh

534
00:29:36,960 --> 00:29:41,240
Speaker 1: too far worse problems if given the right scenario. Okay,

535
00:29:41,240 --> 00:29:43,200
Speaker 1: So so back to the neural network. So you've got

536
00:29:43,200 --> 00:29:46,920
Speaker 1: this machine. Inside the machine, values are being transformed. You

537
00:29:46,920 --> 00:29:50,440
Speaker 1: have inputs and outputs and so inside on the inside,

538
00:29:50,440 --> 00:29:54,040
Speaker 1: and neural network consists of layers of sort of stations

539
00:29:54,160 --> 00:29:57,800
Speaker 1: of value transformation that are called referred to as neurons,

540
00:29:58,320 --> 00:30:01,760
Speaker 1: and each neuron essentially exce up a series of numerical

541
00:30:01,840 --> 00:30:05,120
Speaker 1: values as input, and then it just performs some kind

542
00:30:05,120 --> 00:30:08,720
Speaker 1: of mathematical transformation of those values based on what's known

543
00:30:08,760 --> 00:30:12,400
Speaker 1: as the weights of its connections with the sources that

544
00:30:12,480 --> 00:30:16,360
Speaker 1: received the inputs from. So you've got these interlinking sources

545
00:30:16,400 --> 00:30:20,760
Speaker 1: of information, inputs and outputs throughout, and each neuron takes inputs,

546
00:30:21,120 --> 00:30:24,640
Speaker 1: sums them, does some transformation, and produces an output. So

547
00:30:24,720 --> 00:30:26,840
Speaker 1: for each neuron, you've got a bunch of numbers coming

548
00:30:26,840 --> 00:30:29,600
Speaker 1: from different sources. Each one gets treated with a certain

549
00:30:29,640 --> 00:30:32,560
Speaker 1: bias based on where it came from, and then the

550
00:30:32,600 --> 00:30:35,560
Speaker 1: neuron spits out a new value. And these neurons exist

551
00:30:35,600 --> 00:30:38,920
Speaker 1: in layers, so there are these waves of inputs, say

552
00:30:39,480 --> 00:30:43,520
Speaker 1: the pixels and image getting passed and transformed through one

553
00:30:43,640 --> 00:30:47,400
Speaker 1: layer after another of these neurons until finally the network

554
00:30:47,440 --> 00:30:50,840
Speaker 1: produces numbers that constitute its final outputs. In this case,

555
00:30:51,120 --> 00:30:53,440
Speaker 1: this would be something like its top guesses at the

556
00:30:53,520 --> 00:30:56,000
Speaker 1: word string that describes the image you put in at

557
00:30:56,000 --> 00:30:58,760
Speaker 1: the beginning. Now you'll see from this the value of

558
00:30:58,840 --> 00:31:03,080
Speaker 1: neural network depends completely on how well those connections between

559
00:31:03,200 --> 00:31:06,880
Speaker 1: neurons are weighted to produce the correct results. If they

560
00:31:06,920 --> 00:31:09,520
Speaker 1: just have random weights, then the network will just produce

561
00:31:09,680 --> 00:31:11,880
Speaker 1: random output. It won't be any better than making up

562
00:31:11,960 --> 00:31:14,720
Speaker 1: numbers at random. So the network has to be calibrated

563
00:31:14,840 --> 00:31:18,240
Speaker 1: or trained somehow to produce outputs that are correct. And

564
00:31:18,240 --> 00:31:20,560
Speaker 1: there are multiple ways to do this. Uh. It of

565
00:31:20,600 --> 00:31:23,239
Speaker 1: course could be programmed to some degree by hand, right,

566
00:31:23,240 --> 00:31:26,520
Speaker 1: you could have a program or explicitly uh, going in

567
00:31:26,560 --> 00:31:29,200
Speaker 1: and tinkering with waiting rules to try to get the

568
00:31:29,240 --> 00:31:31,720
Speaker 1: outcomes to be better. But can it can also be

569
00:31:31,760 --> 00:31:34,680
Speaker 1: trained through machine learning, which is a process where inputs

570
00:31:34,680 --> 00:31:39,200
Speaker 1: are already associated with correct outputs, Like you've already got

571
00:31:39,240 --> 00:31:42,000
Speaker 1: a text string associated with the image that you put

572
00:31:42,080 --> 00:31:44,240
Speaker 1: in and you say, this is what you should say

573
00:31:44,280 --> 00:31:47,360
Speaker 1: when it comes out, And each time it runs the process,

574
00:31:47,400 --> 00:31:49,760
Speaker 1: it checks to see how far off from the correct

575
00:31:49,880 --> 00:31:52,800
Speaker 1: known output that it was and then tries to change

576
00:31:52,840 --> 00:31:55,719
Speaker 1: the internal waiting to get closer to the correct answer.

577
00:31:55,960 --> 00:31:58,320
Speaker 1: And of course, with automatic machine learning, you can do

578
00:31:58,400 --> 00:32:01,360
Speaker 1: this at scale. You can do a thousands of times.

579
00:32:01,400 --> 00:32:04,520
Speaker 1: You could potentially do it millions of times just training

580
00:32:04,600 --> 00:32:06,520
Speaker 1: over and over, and you might be able to see

581
00:32:06,520 --> 00:32:08,479
Speaker 1: a parallel here with one of the ways that we

582
00:32:08,560 --> 00:32:11,520
Speaker 1: actually learn. Uh. You know, we we learn in multiple ways.

583
00:32:11,560 --> 00:32:15,320
Speaker 1: Sometimes we learn by being taught explicit rules to follow,

584
00:32:15,400 --> 00:32:18,080
Speaker 1: like if we're learning in school what an insect is

585
00:32:18,200 --> 00:32:20,880
Speaker 1: we might learn that an insect is a small animal

586
00:32:20,960 --> 00:32:25,520
Speaker 1: with an exoskeleton that has six legs, or sometimes on

587
00:32:25,560 --> 00:32:28,360
Speaker 1: the other hand, we learned to generalize from particulars. We

588
00:32:28,480 --> 00:32:31,640
Speaker 1: might see lots of animals pictures of lots of animals

589
00:32:31,680 --> 00:32:34,840
Speaker 1: and notice that the ones that are called insects all

590
00:32:34,920 --> 00:32:38,719
Speaker 1: happen to have six legs and exoskeletons, and therefore we

591
00:32:38,800 --> 00:32:42,880
Speaker 1: derive this category called insect from that survey. And in logic,

592
00:32:42,920 --> 00:32:44,959
Speaker 1: of course, this this process where we come up with

593
00:32:45,000 --> 00:32:48,560
Speaker 1: general rules from lots of individual examples, is known as induction.

594
00:32:49,040 --> 00:32:51,640
Speaker 1: So machine learning to train neural networks is kind of

595
00:32:51,680 --> 00:32:56,200
Speaker 1: like allowing computers to learn categories by induction, kind of

596
00:32:56,240 --> 00:32:58,120
Speaker 1: like we do when we just go out and look

597
00:32:58,160 --> 00:33:00,560
Speaker 1: at the world and see what we find. But one

598
00:33:00,600 --> 00:33:03,800
Speaker 1: of the things that really sets humans apart from computers

599
00:33:04,440 --> 00:33:08,680
Speaker 1: is that humans seem to have this amazing, remarkable ability

600
00:33:09,080 --> 00:33:12,200
Speaker 1: to generalize from particulars. We can often get the gist

601
00:33:12,240 --> 00:33:16,040
Speaker 1: of a category from just a tiny handful of examples.

602
00:33:16,080 --> 00:33:19,160
Speaker 1: You know, when you're giving somebody examples of something to

603
00:33:19,400 --> 00:33:21,560
Speaker 1: like give them the gist of what you're talking about,

604
00:33:21,800 --> 00:33:24,680
Speaker 1: you don't usually need to list a million examples, you

605
00:33:24,720 --> 00:33:28,600
Speaker 1: can list two or three maybe, or sometimes even just one. Yeah.

606
00:33:28,680 --> 00:33:31,120
Speaker 1: This gets into the idea of judging a book by

607
00:33:31,160 --> 00:33:34,720
Speaker 1: its cover, right, Yeah, not supposed to, but we often do,

608
00:33:34,800 --> 00:33:37,000
Speaker 1: and sometimes you you can if you pick up on

609
00:33:37,040 --> 00:33:40,080
Speaker 1: particular things on you know, specifically to use the book. Example,

610
00:33:40,600 --> 00:33:45,080
Speaker 1: specific things about the design or the era of the cover. Yeah. Yeah,

611
00:33:45,120 --> 00:33:48,400
Speaker 1: And in fact, sometimes you can you can judge things

612
00:33:48,400 --> 00:33:51,600
Speaker 1: about the contents of a book just by knowing certain

613
00:33:51,640 --> 00:33:54,760
Speaker 1: things about how certain types of books end up with

614
00:33:54,800 --> 00:33:58,520
Speaker 1: certain types of covers. Right. Like you might think I

615
00:33:58,600 --> 00:34:02,120
Speaker 1: tend to like books that have hand drawn illustrations on

616
00:34:02,160 --> 00:34:04,600
Speaker 1: the cover more than I like books that have sort

617
00:34:04,640 --> 00:34:09,200
Speaker 1: of like c g I stock image cover photos. Means

618
00:34:09,200 --> 00:34:12,200
Speaker 1: you probably like books from like, you know, at least

619
00:34:12,200 --> 00:34:15,880
Speaker 1: the nineteen eighties and before. Yeah, because it seems like

620
00:34:15,920 --> 00:34:18,960
Speaker 1: we have far fewer hand drawn, uh you know, covers

621
00:34:18,960 --> 00:34:22,400
Speaker 1: on books these days. Yeah. Why are people putting stock

622
00:34:22,440 --> 00:34:24,680
Speaker 1: photos on the covers of books? I do not get it?

623
00:34:25,000 --> 00:34:27,760
Speaker 1: But you you didn't need to read a million books

624
00:34:27,840 --> 00:34:29,840
Speaker 1: to come to that conclusion. You could probably come to

625
00:34:29,840 --> 00:34:33,799
Speaker 1: that conclusion after reading I don't know three books like you,

626
00:34:34,080 --> 00:34:37,359
Speaker 1: really we get the gist of things really fast, and

627
00:34:37,960 --> 00:34:42,240
Speaker 1: that's in contrast to computers, which really really don't at all.

628
00:34:42,719 --> 00:34:45,120
Speaker 1: This is one of those strange and amazing things about

629
00:34:45,320 --> 00:34:47,879
Speaker 1: neuroscience about the human brain. How do we solve such

630
00:34:47,920 --> 00:34:51,959
Speaker 1: a difficult problem as generalizing from particulars with so few

631
00:34:52,000 --> 00:34:55,080
Speaker 1: examples to draw from. And of course another example of

632
00:34:55,080 --> 00:34:57,800
Speaker 1: the generalizing power of the human brain is in language.

633
00:34:57,840 --> 00:34:59,319
Speaker 1: Like we've been talking about it, like, how is it

634
00:34:59,360 --> 00:35:02,520
Speaker 1: that more us to the time kids learn how to

635
00:35:02,560 --> 00:35:06,560
Speaker 1: speak a language without being taught explicit rules of syntax

636
00:35:06,640 --> 00:35:09,200
Speaker 1: and grammar and the definitions and usages of all the

637
00:35:09,239 --> 00:35:13,640
Speaker 1: common words, and without hearing billions and billions of examples

638
00:35:13,640 --> 00:35:16,640
Speaker 1: of sentences. They just pick it up. Well, there's there

639
00:35:16,680 --> 00:35:19,160
Speaker 1: there are some specific answers to that. If we've talked

640
00:35:19,160 --> 00:35:21,319
Speaker 1: about that on the show before. Yeah, well, I mean

641
00:35:21,400 --> 00:35:23,279
Speaker 1: I think that there there is a good case to

642
00:35:23,320 --> 00:35:26,760
Speaker 1: be made that the human brain is specially geared towards

643
00:35:26,880 --> 00:35:30,239
Speaker 1: language acquisition in childhood. Right, that's sort of like one

644
00:35:30,280 --> 00:35:34,400
Speaker 1: of our species superpowers. And then those windows close, or

645
00:35:34,560 --> 00:35:39,359
Speaker 1: they don't completely close, but they the windows become much

646
00:35:39,440 --> 00:35:43,160
Speaker 1: smaller later on in life. You know, speaking of children,

647
00:35:43,560 --> 00:35:45,800
Speaker 1: they are you know, they are also frequently a font

648
00:35:45,840 --> 00:35:49,000
Speaker 1: of weirdness and beauty, as they too are learning to

649
00:35:49,080 --> 00:35:51,840
Speaker 1: function in the human world, in the in the adult

650
00:35:51,920 --> 00:35:54,799
Speaker 1: human world, and then they say and do things that

651
00:35:55,000 --> 00:35:58,719
Speaker 1: sometimes hit a weird zone that is either hilarious or

652
00:35:58,840 --> 00:36:01,399
Speaker 1: sometimes a little frightening yes, or or even a little

653
00:36:01,400 --> 00:36:04,239
Speaker 1: bit elegant. Yes, I know exactly what you mean. Like,

654
00:36:04,880 --> 00:36:09,239
Speaker 1: kids often do the same funny outputs based on induction

655
00:36:09,880 --> 00:36:13,600
Speaker 1: that these machine learning algorithms do, Like you can see

656
00:36:13,640 --> 00:36:15,640
Speaker 1: it's funny in the same way that they might say

657
00:36:15,640 --> 00:36:18,080
Speaker 1: something that's a little bit off and kind of absurd,

658
00:36:18,440 --> 00:36:22,040
Speaker 1: but you can sort of understand the rules that got

659
00:36:22,080 --> 00:36:25,680
Speaker 1: them there, right. Like one example I always referred to

660
00:36:25,840 --> 00:36:27,840
Speaker 1: is from years and years ago. I went to a

661
00:36:28,200 --> 00:36:30,640
Speaker 1: children's puppet shows before I had a kid. My my

662
00:36:30,680 --> 00:36:32,360
Speaker 1: wife and I went to check this out because it

663
00:36:32,400 --> 00:36:37,360
Speaker 1: was actors from a local improv group, Dad's Garage, Atlanta.

664
00:36:37,680 --> 00:36:41,680
Speaker 1: They put on the show Uncle Grandpa's Hoodli Story Time.

665
00:36:41,719 --> 00:36:43,799
Speaker 1: I think it was okay. So you had these sort

666
00:36:43,800 --> 00:36:46,319
Speaker 1: of seasoned, uh, you know, improv vets, and they were

667
00:36:46,320 --> 00:36:50,360
Speaker 1: doing a puppet show for kids, and they're taking ideas

668
00:36:50,360 --> 00:36:52,239
Speaker 1: from the audience, and they said, like, who should our

669
00:36:52,239 --> 00:36:54,759
Speaker 1: main character be and so they hand the mic to

670
00:36:54,880 --> 00:36:56,920
Speaker 1: some little little girl in the front row and she

671
00:36:56,960 --> 00:37:01,200
Speaker 1: says Batman the Girl, which which is so hilarious and

672
00:37:01,320 --> 00:37:02,960
Speaker 1: I don't think an adult would be able to come

673
00:37:03,040 --> 00:37:05,279
Speaker 1: up with that, but you can sort of tease it

674
00:37:05,320 --> 00:37:07,520
Speaker 1: apart and figure out how she got there and you know,

675
00:37:07,600 --> 00:37:10,040
Speaker 1: with it. But uh, but it's it's just one of

676
00:37:10,160 --> 00:37:13,239
Speaker 1: you know, many examples I'm sure that that anyone with

677
00:37:13,360 --> 00:37:15,480
Speaker 1: children in their life can can turn to where they

678
00:37:15,480 --> 00:37:18,720
Speaker 1: come with something that is just so goofy or weird

679
00:37:18,840 --> 00:37:23,000
Speaker 1: or or sometimes terrified. Well, I think the real funniness

680
00:37:23,000 --> 00:37:25,680
Speaker 1: and pleasure in that is that it's Batman the Girl

681
00:37:25,800 --> 00:37:29,480
Speaker 1: and not bat Girl, Right, Yeah, Like it's almost there,

682
00:37:29,600 --> 00:37:33,880
Speaker 1: and but by not being there, it's it's also like

683
00:37:33,920 --> 00:37:36,560
Speaker 1: it's it's even better, like it's not it's not bad Girl,

684
00:37:36,600 --> 00:37:41,320
Speaker 1: it's Batman the Girl. It's just so nonsensical, uh, and

685
00:37:41,640 --> 00:37:44,359
Speaker 1: beautiful at the same time. You know. Another one of

686
00:37:44,440 --> 00:37:48,640
Speaker 1: the great AI text generation experiments that Janelle Shane did

687
00:37:48,640 --> 00:37:52,319
Speaker 1: on her blog was was generating that you know those

688
00:37:52,440 --> 00:37:56,480
Speaker 1: Valentine's Day candy hearts. Yes, yes, she had a programmed

689
00:37:56,480 --> 00:37:58,920
Speaker 1: sample those and then try to come up with examples

690
00:37:59,160 --> 00:38:02,600
Speaker 1: and ended up saying things I think like like sweat,

691
00:38:02,680 --> 00:38:08,080
Speaker 1: pooh and hole and time hug. Time hug sounds good

692
00:38:08,120 --> 00:38:11,600
Speaker 1: time hug time cop Yeah, but it also sounds like

693
00:38:11,680 --> 00:38:14,840
Speaker 1: something that like it might be a term that aliens

694
00:38:14,920 --> 00:38:17,600
Speaker 1: come up with for human love. It's like they engage

695
00:38:17,680 --> 00:38:20,200
Speaker 1: not in a hug, but in a time hug. It

696
00:38:20,360 --> 00:38:22,400
Speaker 1: is as if they are hugging for the rest of

697
00:38:22,440 --> 00:38:26,360
Speaker 1: their lives, you know, something like that. Another one, all

698
00:38:26,360 --> 00:38:31,120
Speaker 1: hover and then finally bog love. Um. My son, who

699
00:38:31,480 --> 00:38:34,600
Speaker 1: as of this recording is is six almost seven. He

700
00:38:35,000 --> 00:38:37,640
Speaker 1: drew a picture for my wife and I for for

701
00:38:37,719 --> 00:38:43,320
Speaker 1: Valentine's Valentine's Decardi did to school and it depicted dinosaurs

702
00:38:43,440 --> 00:38:45,880
Speaker 1: as these want to do um and they're there are

703
00:38:46,000 --> 00:38:51,080
Speaker 1: some herbivores walking about, there are some carnivores eating the

704
00:38:51,120 --> 00:38:54,520
Speaker 1: flash of fallen dinosaurs. And then, as is typical in

705
00:38:54,560 --> 00:38:57,160
Speaker 1: paleo art, there may be a volcano in the background,

706
00:38:57,200 --> 00:39:00,400
Speaker 1: but then there is a meteor coming in hard and

707
00:39:00,480 --> 00:39:05,560
Speaker 1: fast asteroid. Yeah, and he writes I love you on

708
00:39:05,600 --> 00:39:09,800
Speaker 1: the on the metior, which which is absolutely wonderful because

709
00:39:09,800 --> 00:39:12,600
Speaker 1: it's like at once it is like like this is beautiful,

710
00:39:12,680 --> 00:39:15,120
Speaker 1: like he totally means all of this and it's like

711
00:39:15,280 --> 00:39:18,080
Speaker 1: the most beautiful Valentine I've ever received and yet to

712
00:39:18,120 --> 00:39:21,719
Speaker 1: put it on the the instrument of of the of

713
00:39:21,760 --> 00:39:26,200
Speaker 1: the extinction events. It's just so weird and like it

714
00:39:26,200 --> 00:39:29,680
Speaker 1: it's accidentally brilliant. You know. Yeah, I saw that when

715
00:39:29,719 --> 00:39:31,480
Speaker 1: you put it on the internet. It was the sweetest

716
00:39:31,480 --> 00:39:34,880
Speaker 1: thing I've ever seen it. And it's like, yeah, this

717
00:39:35,000 --> 00:39:37,240
Speaker 1: is that. This is how love works. Love is a

718
00:39:37,280 --> 00:39:40,120
Speaker 1: is a destroyer as well as a creator. Now, you know,

719
00:39:40,200 --> 00:39:41,680
Speaker 1: we do want to stress it with the especially with

720
00:39:41,719 --> 00:39:46,440
Speaker 1: these text based situations. We're not talking about mer random

721
00:39:46,520 --> 00:39:50,160
Speaker 1: mashups of text, such as like the Wu Tang clan

722
00:39:50,440 --> 00:39:54,480
Speaker 1: name generator or um or more literary example would be

723
00:39:54,480 --> 00:39:58,000
Speaker 1: the cut up technique popularized by author William S. Burrows,

724
00:39:58,960 --> 00:40:02,719
Speaker 1: where you just have like a random um mechanism and

725
00:40:02,800 --> 00:40:07,480
Speaker 1: play to sort of splice words or or sentences together

726
00:40:07,600 --> 00:40:12,400
Speaker 1: to get something that may have sort of accidental meaning

727
00:40:12,440 --> 00:40:15,640
Speaker 1: to it. No, the neural net programs are they are

728
00:40:15,680 --> 00:40:20,640
Speaker 1: algorithms attempting as best they can to approximate the quality

729
00:40:20,920 --> 00:40:24,319
Speaker 1: of the the input texts, the texts they're trained on,

730
00:40:24,640 --> 00:40:28,040
Speaker 1: So they're doing their best to make something like this. Uh.

731
00:40:28,080 --> 00:40:29,640
Speaker 1: And then of course, I mean one of the things

732
00:40:29,760 --> 00:40:32,759
Speaker 1: is you might say, well, why don't they just perfectly

733
00:40:32,800 --> 00:40:35,480
Speaker 1: spit back out the text you've trained them on. In fact,

734
00:40:35,960 --> 00:40:38,160
Speaker 1: if you don't tell them not to, they will do that.

735
00:40:38,400 --> 00:40:40,640
Speaker 1: You know, they'll just spit back in exactly what you

736
00:40:40,760 --> 00:40:43,120
Speaker 1: fed in. You have to sort of like change some

737
00:40:43,280 --> 00:40:46,080
Speaker 1: values and tinker with it to prevent what's known as

738
00:40:46,200 --> 00:40:49,920
Speaker 1: overfitting in this world. Uh, to sort of force the

739
00:40:49,960 --> 00:40:52,839
Speaker 1: algorithms to be more creative and try to come up

740
00:40:52,880 --> 00:40:55,640
Speaker 1: with new versions of the kind of thing they've seen

741
00:40:55,920 --> 00:40:58,719
Speaker 1: instead of just completely copying what you fed them. Yeah,

742
00:40:58,719 --> 00:41:01,200
Speaker 1: because we want these machines to to rule the world,

743
00:41:01,239 --> 00:41:05,399
Speaker 1: not just Hollywood, right. Um, But you know, we also

744
00:41:05,440 --> 00:41:07,640
Speaker 1: have to to distinguish that we're also not talking about

745
00:41:07,719 --> 00:41:13,000
Speaker 1: fake i AI generated text, which can certainly be tremendously

746
00:41:13,600 --> 00:41:17,480
Speaker 1: entertaining as well. Yes, that's such a great genre, people

747
00:41:17,640 --> 00:41:22,720
Speaker 1: pretending to be neural networks creating machine learning generated text,

748
00:41:23,480 --> 00:41:27,160
Speaker 1: which is such an amazing reversal of principles that humans

749
00:41:27,160 --> 00:41:32,000
Speaker 1: have intuitively detected. What's funny about machine learning generated text

750
00:41:32,320 --> 00:41:35,640
Speaker 1: and then made fake human designed versions of it to

751
00:41:35,800 --> 00:41:38,640
Speaker 1: exploit that inherent humor right here, don't. I don't know

752
00:41:38,680 --> 00:41:41,480
Speaker 1: how you get deeper on the irony pit than that. Yeah,

753
00:41:41,520 --> 00:41:44,960
Speaker 1: there's a wonderful two thousand eighteen tweet by comedian Keaton

754
00:41:45,040 --> 00:41:48,160
Speaker 1: Patty and this was the tweet quote. I forced a

755
00:41:48,200 --> 00:41:50,800
Speaker 1: bot to watch over a thousand hours of Alive Garden

756
00:41:50,800 --> 00:41:53,200
Speaker 1: commercials and then asked it to write an alive Garden

757
00:41:53,200 --> 00:41:56,919
Speaker 1: commercial of its own. Here is the first page, uh

758
00:41:56,960 --> 00:42:00,000
Speaker 1: and uh. And then he proceeded to include the image

759
00:42:00,120 --> 00:42:04,279
Speaker 1: is of this text that script rather for and all

760
00:42:04,320 --> 00:42:08,239
Speaker 1: of garden commercial and and it's filled with hilarious nonsense

761
00:42:08,320 --> 00:42:13,439
Speaker 1: like I shall eat Italian citizens and unlimited stick and

762
00:42:13,880 --> 00:42:17,000
Speaker 1: playing seeming, you know, to play upon the whole catchphrase

763
00:42:17,040 --> 00:42:20,080
Speaker 1: of what when you're here your home, when you're here

764
00:42:20,120 --> 00:42:23,560
Speaker 1: your family and you hear your family. There's there's one

765
00:42:23,960 --> 00:42:26,480
Speaker 1: from this uh, this fake script that says leave without me,

766
00:42:26,600 --> 00:42:30,040
Speaker 1: I'm home, which I just I love. I remember laughing

767
00:42:30,080 --> 00:42:32,040
Speaker 1: so hard at this when it came out, as well,

768
00:42:32,280 --> 00:42:35,320
Speaker 1: I love what the waitress says lasagna wings with extra italy.

769
00:42:37,320 --> 00:42:41,080
Speaker 1: There's also like really funny stage directions in it. Yeah, well,

770
00:42:41,160 --> 00:42:43,719
Speaker 1: oh yeah, you mentioned the infinite uh where it says

771
00:42:43,800 --> 00:42:48,040
Speaker 1: like we the gluten classic O. We believe the waitress

772
00:42:48,160 --> 00:42:50,240
Speaker 1: that it is from the kitchen. We have no reason

773
00:42:50,280 --> 00:42:53,560
Speaker 1: not to believe. Now, of course, this is just a

774
00:42:53,640 --> 00:42:57,480
Speaker 1: comedian doing this thing, trying to pretend to be inn AI. UM.

775
00:42:57,520 --> 00:42:59,800
Speaker 1: But I was reading an article where they quoted Janelle

776
00:42:59,800 --> 00:43:02,040
Speaker 1: Shae and the author of the AI Weirdness blog, who

777
00:43:02,080 --> 00:43:03,960
Speaker 1: trains all these neural networks to come up with all

778
00:43:03,960 --> 00:43:05,640
Speaker 1: this funny stuff we were talking about at the beginning

779
00:43:05,640 --> 00:43:07,759
Speaker 1: of the episode. And you know, she talks about that

780
00:43:07,840 --> 00:43:12,080
Speaker 1: there are ways to notice uh when something was probably

781
00:43:12,080 --> 00:43:15,239
Speaker 1: written by a human instead of by an actual AI.

782
00:43:15,440 --> 00:43:18,560
Speaker 1: Both can be absurd in similar funny ways. But one

783
00:43:18,560 --> 00:43:21,719
Speaker 1: of the problems with this script UH in passing as

784
00:43:21,760 --> 00:43:24,800
Speaker 1: a real AI generated text is that it's actually too coherent,

785
00:43:25,400 --> 00:43:28,759
Speaker 1: like it's memory. Its memory is too long. It remembers

786
00:43:28,920 --> 00:43:32,640
Speaker 1: characters from many lines earlier and still has them appearing

787
00:43:32,680 --> 00:43:36,359
Speaker 1: and saying things. Actual AI generated texts of a much

788
00:43:36,400 --> 00:43:39,879
Speaker 1: shorter memory. They they're not consistent in that way. They

789
00:43:39,920 --> 00:43:43,120
Speaker 1: don't make Actually, they make even less sense than the

790
00:43:43,160 --> 00:43:47,520
Speaker 1: fake All of Garden commercial. So she's saying, don't reach

791
00:43:47,560 --> 00:43:50,359
Speaker 1: for your hatchet on this one. No, because a real

792
00:43:50,400 --> 00:43:54,520
Speaker 1: neural net generated text only mimics forms, it doesn't mimic meaning.

793
00:43:55,000 --> 00:43:59,520
Speaker 1: And this thing, it's it means too much, it's too clever. Okay,

794
00:43:59,520 --> 00:44:01,759
Speaker 1: so we're gonna go ahead and close out this episode now.

795
00:44:01,800 --> 00:44:03,919
Speaker 1: But again there's going to be a part two where

796
00:44:03,920 --> 00:44:07,040
Speaker 1: we continue this discussion and really get more into the

797
00:44:07,040 --> 00:44:09,640
Speaker 1: meat of the topic. In the meantime, check out Stuff

798
00:44:09,640 --> 00:44:11,359
Speaker 1: to Blow your Mind dot com. That's the mothership. That's

799
00:44:11,360 --> 00:44:13,799
Speaker 1: whe we'll find all the episodes of the show. There's

800
00:44:13,840 --> 00:44:15,799
Speaker 1: also a little button at the top you can click

801
00:44:15,840 --> 00:44:18,040
Speaker 1: on that go to our T shirt store get t

802
00:44:18,200 --> 00:44:21,160
Speaker 1: shirts stickers with a number of cool designs, designs that

803
00:44:21,480 --> 00:44:24,000
Speaker 1: line up with certain show topics like the Great Basilisk

804
00:44:24,160 --> 00:44:27,799
Speaker 1: or various squirrel episodes. Uh, as well as just as

805
00:44:27,840 --> 00:44:32,040
Speaker 1: you like show logo um material as well. And if

806
00:44:32,080 --> 00:44:34,080
Speaker 1: you want to support the show, the best thing that

807
00:44:34,120 --> 00:44:36,560
Speaker 1: you can do is rate and review Stuff to Blow

808
00:44:36,600 --> 00:44:38,120
Speaker 1: your Mind wherever you have the power to do so.

809
00:44:38,160 --> 00:44:41,120
Speaker 1: Wherever you get this podcast, give us some stars, give

810
00:44:41,200 --> 00:44:43,600
Speaker 1: us a nice review. It really helps us out in

811
00:44:43,640 --> 00:44:47,600
Speaker 1: our war against the almighty algorithms. Big thanks as always

812
00:44:47,600 --> 00:44:51,240
Speaker 1: to our excellent audio producers Alex Williams and Tarry Harrison.

813
00:44:51,520 --> 00:44:53,000
Speaker 1: If you would like to get in touch with us

814
00:44:53,000 --> 00:44:55,400
Speaker 1: directly with feedback about this episode or any other, to

815
00:44:55,440 --> 00:44:58,000
Speaker 1: suggest a topic for the future, or just to say hi,

816
00:44:58,239 --> 00:45:01,080
Speaker 1: you can email us at blow Oh the Mind at

817
00:45:01,120 --> 00:45:14,319
Speaker 1: how stuff works dot com for more on this and

818
00:45:14,400 --> 00:45:40,040
Speaker 1: thousands of other topics. Is it how stuff works dot com.