LLMs prove Structuralism was Correct all Along // 258

WEBVTT
1
00:00:07.320 --> 00:00:10.039
But today, okay, some of our patrons. I was saying,
2
00:00:10.279 --> 00:00:15.599
some of them a couple more. Every episode starts saying, okay,
3
00:00:15.640 --> 00:00:20.559
stop talking about AI now I'm bored. But before we stop,
4
00:00:20.719 --> 00:00:23.760
we had to get to this one. This was recommended
5
00:00:23.800 --> 00:00:28.039
to me by another patron, and I'm glad it was
6
00:00:28.079 --> 00:00:33.159
recommended because we are back. Structuralism is back. Hydigerians out there,
7
00:00:33.240 --> 00:00:34.719
you are on suicide watch.
8
00:00:35.840 --> 00:00:36.679
Unnotice.
9
00:00:37.079 --> 00:00:42.520
Language is completely technical and structuralism is confirmed correct. We've
10
00:00:42.560 --> 00:00:44.920
been waiting for this for about one hundred years, a
11
00:00:45.000 --> 00:00:49.000
century after it's invented. Structuralism is the correct view of language,
12
00:00:49.280 --> 00:00:53.799
and Chadjibti proves it. And I'm dropping my vendetta against
13
00:00:53.799 --> 00:00:59.719
anthropic and Sam Altman. Yeah, just for proving structuralism, they
14
00:00:59.799 --> 00:01:03.320
got the human is amount of language. Finally, what's with your.
15
00:01:03.240 --> 00:01:06.079
Beat you guys. I'll never understand your beef against humanism.
16
00:01:06.120 --> 00:01:07.760
I know we've done a couple episodes about it, but
17
00:01:07.799 --> 00:01:08.280
it's just.
18
00:01:10.040 --> 00:01:12.879
I don't. I don't. I can't abide by people being wrong.
19
00:01:12.959 --> 00:01:16.760
But you're you're your Your utterance that it's wrong is
20
00:01:16.879 --> 00:01:17.640
human dependent?
21
00:01:18.599 --> 00:01:20.799
No, it's not. It's linguistically dependent.
22
00:01:21.280 --> 00:01:22.719
Linguistics is human dependent.
23
00:01:23.159 --> 00:01:25.920
No, it's not, as Chet has just proven.
24
00:01:27.200 --> 00:01:29.000
But Chach only works because of humans.
25
00:01:30.159 --> 00:01:32.120
Uh, you don't know that anyway.
26
00:01:32.120 --> 00:01:35.760
I'll let you guys try you guys. You guys can
27
00:01:35.799 --> 00:01:37.760
have your post human you guys can have your little
28
00:01:37.799 --> 00:01:41.239
post humans, post structure or structurreless little jerk off session,
29
00:01:41.239 --> 00:01:42.560
and I'll just be here for the ride.
30
00:01:43.000 --> 00:01:45.120
You can't ask a question and then not pay attention
31
00:01:45.200 --> 00:01:49.120
to the answer. It just demonstrates intellectual incuriosity on your part.
32
00:01:49.439 --> 00:01:51.280
Well, I mean, I generally, I'm trying to I'm trying
33
00:01:51.280 --> 00:01:54.040
to understand. But all right, let's let's as ahead.
34
00:01:53.719 --> 00:01:55.439
And thank you for bringing that up. I'm not a
35
00:01:56.120 --> 00:02:02.079
I've decided that post structuralism isn't real. Structuralism consists of
36
00:02:02.239 --> 00:02:08.439
a few less strict structuralists critiquing a few more strict structuralists.
37
00:02:08.520 --> 00:02:10.639
That's all that post structuralism is. It's made up.
38
00:02:10.680 --> 00:02:13.599
I think you're right about that. I think post structuralism
39
00:02:14.039 --> 00:02:19.800
is a creation of the reception history of French theory
40
00:02:20.120 --> 00:02:24.000
in Anglo in the Anglo sphere, because we got a
41
00:02:24.000 --> 00:02:29.520
lot of those texts translated even before structuralists. Major structuralist
42
00:02:29.560 --> 00:02:35.759
texts were translated into English. We got Drrieda, We got Leotard,
43
00:02:36.439 --> 00:02:40.360
we got people like that. Yeah, and then but we
44
00:02:40.400 --> 00:02:44.840
didn't get things like the art structuralist texts like Grimass
45
00:02:45.240 --> 00:02:48.919
until like later. It was almost like a reverse reception.
46
00:02:49.360 --> 00:02:53.120
We got post structuralism before we got structuralism in the
47
00:02:53.159 --> 00:02:57.520
Anglo world, and then by the time structuralism arrived, it
48
00:02:57.599 --> 00:03:01.599
was like, well, isn't this already been surpass by these
49
00:03:01.639 --> 00:03:04.879
texts which have already claimed that structuralism has no center
50
00:03:05.719 --> 00:03:10.599
and we never But in reality, post structuralism is really
51
00:03:10.680 --> 00:03:14.360
just like a kind of continuation of structuralism. There's no
52
00:03:14.599 --> 00:03:15.520
clean break there.
53
00:03:16.120 --> 00:03:19.520
I'm already on this beat because I think I've done
54
00:03:19.520 --> 00:03:23.120
like thirteen on patron thirteen videos in a row on
55
00:03:23.319 --> 00:03:26.560
just structuralism. So I'm already on this beat. But do
56
00:03:26.560 --> 00:03:29.439
you know they don't there's no French word for post structuralism.
57
00:03:29.479 --> 00:03:31.919
They don't say it. They just call it theory.
58
00:03:32.400 --> 00:03:36.080
Yeah, yeah, because I'm pretty sure it's it's like a
59
00:03:36.120 --> 00:03:44.080
classification effort by American academics. I don't think there. And
60
00:03:44.159 --> 00:03:47.879
structuralism even the term itself was a term of abuse
61
00:03:48.879 --> 00:03:52.280
by like non structuralists, putting it on a certain group
62
00:03:52.360 --> 00:03:57.879
of theorists, linguistic theorists who followed broadly like so sure
63
00:03:58.319 --> 00:04:03.599
and he alms love and people like that. And then
64
00:04:03.840 --> 00:04:07.439
I think it was just adopted as people said, fine,
65
00:04:07.479 --> 00:04:11.319
call us structuralists. We don't fucking care. So it originally
66
00:04:11.400 --> 00:04:12.719
was a term of abuse anyway.
67
00:04:13.759 --> 00:04:18.439
So AI, we've been on AI. We're not doing ever again.
68
00:04:18.480 --> 00:04:21.759
AI with consciousness. We tried it. We put that baby
69
00:04:21.759 --> 00:04:24.360
to bed. That's what the analytics are doing anyway.
70
00:04:25.639 --> 00:04:28.160
And I thought, and some of the phenomenologists too, are
71
00:04:28.199 --> 00:04:28.959
still on that too.
72
00:04:29.040 --> 00:04:33.319
But yes, right right, like Dreyfus, there's a few points
73
00:04:33.360 --> 00:04:36.079
of reference that I wanted to bring back from our
74
00:04:36.120 --> 00:04:40.160
previous episode to bring into here. Oh, by the way,
75
00:04:40.199 --> 00:04:42.800
there's a guy I don't know who he is. I
76
00:04:42.839 --> 00:04:46.439
think he's in New York. His name is Leif Weatherbee,
77
00:04:47.000 --> 00:04:51.079
and he wrote language Machines and it confirmed a lot
78
00:04:51.079 --> 00:04:54.439
of our, at least my suspicions. I feel like Victor's
79
00:04:54.519 --> 00:04:59.519
gonna put up a fuss here, but I'm expecting Eric
80
00:04:59.600 --> 00:05:03.439
to help me beat down Victor here. But we got
81
00:05:03.519 --> 00:05:05.360
with this. What's the guy, the deep mind guy? We
82
00:05:05.399 --> 00:05:09.800
read a deep mind researcher who wrote very astutely to
83
00:05:09.839 --> 00:05:13.160
the surprise of all that simulating gravity on a computer
84
00:05:13.199 --> 00:05:18.000
is not gravity, and simulating photosynthesis on a computer is
85
00:05:18.079 --> 00:05:25.120
not photosynthesis, and thus his argument goes, simulating consciousness on
86
00:05:25.199 --> 00:05:29.439
a computer is also not consciousness. Now there's a big
87
00:05:29.519 --> 00:05:31.879
jump to get to the consciousness, and we're not even
88
00:05:31.879 --> 00:05:35.839
gonna touch that again because it was very disappointing. But
89
00:05:36.360 --> 00:05:39.600
I think anyone can say that simulating a chess game
90
00:05:40.120 --> 00:05:43.879
on a computer is indeed a real chess game. And
91
00:05:43.920 --> 00:05:47.319
now we come to the question today of structuralist linguistics
92
00:05:47.439 --> 00:05:54.240
versus chat GPT. GPT three is the example used by
93
00:05:54.279 --> 00:05:58.279
this book, and the export of that question is, well,
94
00:06:00.079 --> 00:06:02.319
we know that chess on a computer's real chess. We
95
00:06:02.399 --> 00:06:04.879
know that gravity on a computer is not real gravity.
96
00:06:05.120 --> 00:06:07.959
But what about the language question? If language is simulated
97
00:06:08.000 --> 00:06:11.040
by a computer or an l l M, does that
98
00:06:11.120 --> 00:06:16.839
constitute real language or simulated language? The answer may shock you.
99
00:06:17.920 --> 00:06:22.360
Yeah, I guess the what's the humanist answer would be no?
100
00:06:23.480 --> 00:06:24.920
Oh, I mean I think it's yes.
101
00:06:25.639 --> 00:06:26.480
But when when you.
102
00:06:26.480 --> 00:06:29.519
Have to say the humanist answer is no, because only
103
00:06:29.639 --> 00:06:34.319
humans can use language. Everything else is say, is like
104
00:06:34.360 --> 00:06:37.360
a pet, like the parrot, like or like the raven
105
00:06:37.480 --> 00:06:40.079
or like a parrot. Right, It's just everything else is
106
00:06:40.199 --> 00:06:42.519
just like making sounds.
107
00:06:43.319 --> 00:06:45.480
Luckily we have a humanist defender on the podet.
108
00:06:45.519 --> 00:06:48.240
I guess as the I guess as the humanist representative here,
109
00:06:48.759 --> 00:06:52.399
But I still say, but for me, the answer is yes,
110
00:06:52.439 --> 00:06:57.519
if you can, if you can distinguish between language and meaning,
111
00:06:58.879 --> 00:07:01.800
so like, you can't have meaning without human but you
112
00:07:01.839 --> 00:07:03.839
can have language. I mean, you can have patterns, you
113
00:07:03.879 --> 00:07:05.720
can have anything without humans. You can have a bunch
114
00:07:05.720 --> 00:07:09.439
of shit. It's just like pattern stuff that's without humans.
115
00:07:09.519 --> 00:07:12.920
But to have like meaning, which I associate with intentionality,
116
00:07:13.399 --> 00:07:14.160
you need humans.
117
00:07:14.759 --> 00:07:17.000
So one thing a humanist would probably say about this
118
00:07:17.199 --> 00:07:20.360
is that language and meaning are equivalent.
119
00:07:21.079 --> 00:07:23.120
Okay, so I guess I'm not a humanist.
120
00:07:23.160 --> 00:07:24.560
Or like tightly connected.
121
00:07:24.759 --> 00:07:27.079
I'm not saying every humanist thinks that, but I'm thinking
122
00:07:27.079 --> 00:07:33.839
that like meaning derives, meaning is downstream or maybe immediately
123
00:07:33.920 --> 00:07:35.199
upstream of language.
124
00:07:36.120 --> 00:07:38.240
I mean, I guess I'm thinking about meaning a little
125
00:07:38.279 --> 00:07:43.560
bit more expansively, like uhcause, I guess because I guess
126
00:07:43.600 --> 00:07:46.920
I was even thinking about Wickenstein a little bit, you know,
127
00:07:46.959 --> 00:07:49.560
because he he talks about you know, the meaning of
128
00:07:49.600 --> 00:07:51.680
a word is use which I think is somewhat consistent
129
00:07:51.720 --> 00:07:56.120
with like its role in you know, the structure kind
130
00:07:56.160 --> 00:07:59.040
of like it's just like how it's used, but ultimately,
131
00:07:59.040 --> 00:08:01.319
like he also appeals to this thing he calls forms
132
00:08:01.360 --> 00:08:04.360
of life, right, which is like like the the the
133
00:08:04.519 --> 00:08:07.360
ultimately the language. It's a whole, it's a whole pattern
134
00:08:07.439 --> 00:08:10.879
system is like nothing in a sense meaning And I
135
00:08:11.000 --> 00:08:12.519
kind of take him this might be a bit of
136
00:08:12.560 --> 00:08:17.399
a leap to mean like meaning intentionality, Like forms of
137
00:08:17.439 --> 00:08:20.319
life is like the thing that gives the whole structure
138
00:08:20.560 --> 00:08:24.040
like its point. But you could still have like the pattern,
139
00:08:25.240 --> 00:08:26.839
it just wouldn't have the forms of life part.
140
00:08:27.879 --> 00:08:31.480
But do you need the forms of life to turn
141
00:08:31.560 --> 00:08:34.000
into language or what's the connection there? Oh?
142
00:08:34.000 --> 00:08:38.240
Okay, so I mean I think uh oh, like like
143
00:08:38.279 --> 00:08:39.879
do you mean like temporarily.
144
00:08:40.879 --> 00:08:43.240
I mean, you still have on your side the argument
145
00:08:43.440 --> 00:08:48.120
that chatipt did not invent English or whatever language is
146
00:08:48.159 --> 00:08:50.519
it speaks, Yeah, and yet it's capable of doing it.
147
00:08:51.120 --> 00:08:54.039
But you're saying that a specific form of life was
148
00:08:54.120 --> 00:08:57.600
meant to or was required in the first place, to
149
00:08:57.759 --> 00:09:00.000
give rise to English.
150
00:09:00.120 --> 00:09:02.039
Yeah, I think so for that pattern to like to
151
00:09:02.120 --> 00:09:04.720
like emerge, Yeah, I think so. I think there is
152
00:09:04.759 --> 00:09:06.960
a precondition required.
153
00:09:07.399 --> 00:09:09.440
So Hydigerians are not on suicide watch.
154
00:09:09.440 --> 00:09:13.159
You're saying I think, I think not, I mean, I
155
00:09:13.200 --> 00:09:16.519
don't know about Vickenstein. Maybe maybe the real advance is
156
00:09:17.000 --> 00:09:22.039
that meaning is context dependent, so the exact same word
157
00:09:22.200 --> 00:09:25.559
can mean something completely So in a sense, then there's
158
00:09:25.759 --> 00:09:28.679
another way of saying like meaning is not fixed or
159
00:09:29.000 --> 00:09:34.360
there's no fixed meaning to language. It's it's contextually dependent one.
160
00:09:34.519 --> 00:09:38.960
Whether you're saying the same word in one game one
161
00:09:39.039 --> 00:09:42.720
context or another, it can mean something completely different, it
162
00:09:42.720 --> 00:09:45.840
has a different effect. But then it's another thing to
163
00:09:45.919 --> 00:09:53.200
say that language depends on meaning, and for that to
164
00:09:53.279 --> 00:09:55.919
be the case, there has to be intention And so
165
00:09:56.480 --> 00:09:59.559
we're just smuggling consciousness back in because it's not language
166
00:09:59.639 --> 00:10:04.200
unless there's like intentional use. And ultimately the meaning depends
167
00:10:04.240 --> 00:10:08.039
on the intentional use of the of the speaker, which
168
00:10:08.080 --> 00:10:11.759
which then just smuggles consciousness back into it. And that
169
00:10:11.919 --> 00:10:13.440
and we're supposed to be past that.
170
00:10:14.159 --> 00:10:15.159
You could have a pattern.
171
00:10:16.480 --> 00:10:21.279
I can and in structuralism tends to be more in
172
00:10:22.320 --> 00:10:28.480
the formal camp where language should be analyzed separate from
173
00:10:29.159 --> 00:10:33.600
I guess, speakers and receivers, and it should be in
174
00:10:33.840 --> 00:10:37.720
its structure the more. The more classic view would be
175
00:10:39.080 --> 00:10:42.320
let the meaning of language is determined by its intention
176
00:10:42.519 --> 00:10:45.360
by the user, by the author. Then the author is
177
00:10:45.399 --> 00:10:47.679
the one that determines the meaning. And then you have
178
00:10:47.720 --> 00:10:51.720
things like reader response theory that would locate meaning more
179
00:10:51.799 --> 00:10:55.200
on the side of the receiver and the interpreter, and
180
00:10:55.240 --> 00:10:58.559
say that you know the author's intention if it is
181
00:10:58.600 --> 00:11:00.720
a factor, it's just one fact or a book in
182
00:11:00.879 --> 00:11:03.960
among many, and it's not really an important one. So
183
00:11:04.080 --> 00:11:08.799
intention is not really that important, like you, Then you
184
00:11:08.840 --> 00:11:12.279
get to think the intentional fallacy and the and then
185
00:11:12.360 --> 00:11:14.879
I guess the other way is the effective fallacy. But
186
00:11:14.879 --> 00:11:18.240
but yeah, the intentional fallacy where meaning depends on the
187
00:11:18.279 --> 00:11:19.279
intention of the author.
188
00:11:19.960 --> 00:11:22.120
Well, I don't. I didn't mean intentionality in the sense
189
00:11:22.159 --> 00:11:24.960
of like intention of like a speaker. I mean in
190
00:11:25.000 --> 00:11:27.320
the form in the sense in the phenomenological sense of
191
00:11:27.320 --> 00:11:29.120
like intentionality.
192
00:11:28.559 --> 00:11:35.639
Like aboutness. It has to be about about something subjective aboutness. Yeah, like,
193
00:11:35.720 --> 00:11:37.799
I guess you have reference. But in what we're going
194
00:11:37.879 --> 00:11:39.879
to talk about, we're going to go back to so
195
00:11:40.080 --> 00:11:47.720
sersal distinction between signifier and signified, which which are psychological entities.
196
00:11:47.799 --> 00:11:52.320
So the signified is not it's a mental concept. It's
197
00:11:52.360 --> 00:11:55.759
not it's not about anything in the world. It's a
198
00:11:55.799 --> 00:12:00.440
mental concept. So maybe maybe again as structural as kind
199
00:12:00.440 --> 00:12:04.159
of in the middle, it's like the formal side.
200
00:12:05.159 --> 00:12:07.960
Yeah. The punchline of the episode is so sir didn't
201
00:12:07.960 --> 00:12:11.879
even realize how correct he was. He was more correct
202
00:12:11.879 --> 00:12:14.840
than he he It goes further than even he was
203
00:12:14.879 --> 00:12:17.519
willing to go with it with what he had at
204
00:12:17.519 --> 00:12:20.000
his disposal. But I guess the question is to call
205
00:12:20.080 --> 00:12:25.279
back our previous episode. The question is does chat GBT
206
00:12:26.240 --> 00:12:31.399
or Cloud Cloud? Does Cloud understand what the hamburger is?
207
00:12:32.000 --> 00:12:35.679
I was saying Cloud and now you're saying Cloud Cloud final.
208
00:12:36.559 --> 00:12:44.799
I'm making an extra Canadian Oh Cloud, your answer you guys, listeners,
209
00:12:44.840 --> 00:12:48.879
whoever your list Your answer to the question does Cloud
210
00:12:49.159 --> 00:12:53.159
understand what a hamburger is? Is of is of chief
211
00:12:53.279 --> 00:12:58.480
importance because I'm gonna say Cloud does understand what? I
212
00:12:58.519 --> 00:13:01.000
can't do this the whole episode. It's it's too annoying.
213
00:13:01.200 --> 00:13:06.200
I'll just say TAGGYBT, does CHAGYPT understand what a hamburger is?
214
00:13:06.840 --> 00:13:12.080
I'm gonna say yes, And if you say no, then
215
00:13:12.159 --> 00:13:16.480
you're a little bit more on the intentionality or reader
216
00:13:16.799 --> 00:13:21.039
or interpreter side of the interpretership model.
217
00:13:21.039 --> 00:13:22.679
I think that I think the disagreement is going to
218
00:13:22.720 --> 00:13:24.679
come down to what you mean by understand, because I
219
00:13:24.720 --> 00:13:29.279
think there is a sense if you define understand in
220
00:13:29.360 --> 00:13:34.960
the sense of recognize the pattern of the language and
221
00:13:35.039 --> 00:13:38.840
like where something would fit, then sure it understands. But
222
00:13:38.919 --> 00:13:45.159
if understanding is crucially related to a subjective sense of aboutness,
223
00:13:45.279 --> 00:13:48.799
where it has like a meaning in an intentional orientation
224
00:13:48.879 --> 00:13:50.480
towards the world, then the answer would be no.
225
00:13:52.200 --> 00:13:55.440
And someone like systems theories like Nick Las Luhmann would say,
226
00:13:55.480 --> 00:13:59.000
these are two different. If you have a psychic memory
227
00:13:59.039 --> 00:14:02.080
because you've eaten a ham before and you remember what
228
00:14:02.080 --> 00:14:03.840
it tastes like and you remember the ingredients that go
229
00:14:03.919 --> 00:14:08.960
on it, this is one hamburger, and there's no overlap
230
00:14:09.039 --> 00:14:12.279
or connection. There's like a little node where that connects
231
00:14:12.320 --> 00:14:16.600
to the word hamburger and everything that hamburger is connected
232
00:14:16.600 --> 00:14:20.720
to in the signifying matrix that we call language, which
233
00:14:20.720 --> 00:14:25.399
includes lettuce, includes sandwich, includes fries, includes fast food, and
234
00:14:25.679 --> 00:14:28.639
includes all these other terms that are near to it
235
00:14:28.720 --> 00:14:32.000
in proximity. So if you think those two have some
236
00:14:32.159 --> 00:14:39.000
necessary overlap, then you might not be convinced that cloud
237
00:14:39.080 --> 00:14:41.960
or chagbt understands what a hamburger is. But if you're
238
00:14:42.000 --> 00:14:45.120
willing to accept that those things are separate, then you
239
00:14:45.120 --> 00:14:48.279
can say that it does understand one of the hamburgers
240
00:14:48.879 --> 00:14:52.000
and not the other. But we don't have we don't
241
00:14:52.000 --> 00:14:54.639
even really have the other without being able to refer
242
00:14:54.679 --> 00:14:56.440
to it in language, because there's no hamburger in this
243
00:14:56.519 --> 00:14:57.759
conversation except the word.
244
00:14:57.840 --> 00:15:02.600
You know, just quick really quickly. I just want to, like,
245
00:15:03.039 --> 00:15:05.240
just because it's right on this question of under of
246
00:15:05.279 --> 00:15:09.519
like whether LMS understand, I guess, like when it's real.
247
00:15:09.559 --> 00:15:11.960
I guess the moment when it's revealed that the LM
248
00:15:12.039 --> 00:15:15.799
doesn't understand in the subjective sense is whenever it makes
249
00:15:15.840 --> 00:15:19.480
a funny mistake, right, like, because that's happening less and
250
00:15:19.559 --> 00:15:22.360
less now. But there are those cases where LMS will
251
00:15:22.360 --> 00:15:25.600
make these weird mistakes, and I wish I had one
252
00:15:25.639 --> 00:15:28.720
ready to hand, But where it'll like use hamburger in
253
00:15:28.759 --> 00:15:30.639
a context that doesn't really make sense that used to
254
00:15:30.639 --> 00:15:34.080
happen with earlier versions of LMS, then it would be
255
00:15:34.159 --> 00:15:36.000
like kind of clear. You'd be like, oh, okay, this
256
00:15:36.120 --> 00:15:39.320
is not really understanding, right it got the pattern wrong
257
00:15:39.399 --> 00:15:41.559
this time. Anyway, That's all I want to say.
258
00:15:41.639 --> 00:15:44.399
Luhmann's idea of communication, what he splits it up into
259
00:15:44.399 --> 00:15:50.399
three parts, right, information, utterance, and understanding, And all understanding
260
00:15:50.600 --> 00:15:55.639
really is is like the capacity to accept, reject, or
261
00:15:55.759 --> 00:16:05.360
query a previous communication. And I think chatbots can do that, right, Yeah,
262
00:16:05.480 --> 00:16:11.320
because you have Elena Esposito arguing that chatbots are fully
263
00:16:11.480 --> 00:16:18.600
functional communication partners that are completely able to participate in
264
00:16:18.799 --> 00:16:25.480
communication alongside human beings, and therefore they share in the
265
00:16:25.519 --> 00:16:31.559
ability to understand and perpetuate communication in the way that
266
00:16:31.639 --> 00:16:37.000
makes society continue. And that's all it needs, society, all
267
00:16:37.000 --> 00:16:42.480
it needs is to continue communicating. I guess theoretically it
268
00:16:42.480 --> 00:16:46.399
doesn't require humans if humans aren't the only ones that
269
00:16:46.440 --> 00:16:47.279
can communicate.
270
00:16:47.960 --> 00:16:53.000
Yeah, that's another humanist question. In principle, does language require
271
00:16:53.080 --> 00:16:56.720
humans or did just happen in this context that language
272
00:16:56.720 --> 00:17:02.759
and humans are are co interdependent. Language. Language is like
273
00:17:02.799 --> 00:17:09.400
an alien a Zeno virus that we just happened to
274
00:17:09.440 --> 00:17:11.880
pick up at some point, and it has it has
275
00:17:11.920 --> 00:17:16.319
an independent existence of any instantiation in a single in
276
00:17:16.359 --> 00:17:19.200
a single parole. It exists beyond each parole.
277
00:17:19.599 --> 00:17:22.200
Well, I guess language is a medium and we pull
278
00:17:22.240 --> 00:17:27.519
it together into individual utterances that unfold sequentially in time,
279
00:17:28.240 --> 00:17:31.599
but other and that's what you use it for.
280
00:17:32.440 --> 00:17:36.880
Wickenstein, I'm realizing, is a humanist on this in this discourse,
281
00:17:37.960 --> 00:17:42.119
his private language argument, I think could be applied here
282
00:17:43.319 --> 00:17:46.200
to say that, no, you can't have a language without humans,
283
00:17:46.200 --> 00:17:49.799
because the private language argument is more specifically like could
284
00:17:49.839 --> 00:17:53.799
you have a language in just one human alone? And
285
00:17:53.880 --> 00:17:56.720
his whole point is like, that's impossible because the whole
286
00:17:56.759 --> 00:18:01.920
function of language requires verification by another, essentially like it's
287
00:18:02.039 --> 00:18:04.839
for you to remember, oh this this works in this place,
288
00:18:04.960 --> 00:18:09.119
in this meaning function. It requires constant verification, rechecking by
289
00:18:09.160 --> 00:18:12.960
other introlockutors. So you can't have a private language because
290
00:18:12.960 --> 00:18:17.400
you could never be sure Oh does that does that
291
00:18:17.440 --> 00:18:19.839
signify or play this role in this place? Or did
292
00:18:19.839 --> 00:18:22.839
I just make that up? So the rule the game
293
00:18:23.920 --> 00:18:29.000
requires players in a sense, so I think, so I
294
00:18:29.000 --> 00:18:31.079
don't know. Then, Okay, now the question is could you
295
00:18:31.119 --> 00:18:34.480
just have a language with no players? No.
296
00:18:35.079 --> 00:18:38.200
Vickenstein does say if a lion could speak, we wouldn't
297
00:18:38.200 --> 00:18:41.799
be able to understand it, and that I mean, it's
298
00:18:41.799 --> 00:18:45.640
not like we understand computers. The computers doing all the translating.
299
00:18:46.720 --> 00:18:49.759
Yeah, but he says that because because its whole world
300
00:18:49.880 --> 00:18:53.839
is different. Yeah, because because of a human feel.
301
00:18:53.640 --> 00:18:55.839
Like if you had a large enough data set, then
302
00:18:56.680 --> 00:18:58.240
a computer could talk to lions.
303
00:18:58.720 --> 00:19:00.799
Probably don't know about that one.
304
00:19:01.119 --> 00:19:07.440
The implication here structureless linguistics. How is it otherwise? How
305
00:19:07.480 --> 00:19:10.519
is it otherwise that a computer which is a quantitative
306
00:19:10.559 --> 00:19:15.559
machine seems to be able to produce language so convincingly,
307
00:19:15.799 --> 00:19:20.799
like past the Turing test on the daily at this point,
308
00:19:21.960 --> 00:19:26.440
if meaning is this necessary dimension or aspect of it,
309
00:19:27.000 --> 00:19:28.559
and it might be. I mean, this is kind of
310
00:19:28.599 --> 00:19:32.960
what I'm convinced of language is we like using language,
311
00:19:33.039 --> 00:19:36.039
and we enjoy it and we take pleasure in it,
312
00:19:36.400 --> 00:19:39.799
but it's not a necessary part of language for it
313
00:19:39.839 --> 00:19:42.160
to work. And it's why we're going to keep using it,
314
00:19:42.200 --> 00:19:46.160
even though even if the computers get better at better
315
00:19:46.240 --> 00:19:49.680
than us at writing poetry and writing novels, We're still
316
00:19:49.720 --> 00:19:53.119
going to do it because of that sense creation and
317
00:19:53.160 --> 00:19:57.200
we have that we have that dimension when we interface
318
00:19:57.559 --> 00:20:00.400
with language that I don't think there's any evidence Chatchy
319
00:20:00.480 --> 00:20:03.920
Bet does Chatchi bed is not feeling the pleasure of
320
00:20:04.039 --> 00:20:08.359
coherence when it produces a good answer whereas we do.
321
00:20:08.759 --> 00:20:11.440
But I think what this shows is that that part
322
00:20:11.799 --> 00:20:15.240
is not necessary for the back and forth to occur.
323
00:20:15.759 --> 00:20:18.720
And that's something that we've never been able to test before,
324
00:20:19.759 --> 00:20:22.240
and it's something that structuralism would have predicted if you
325
00:20:22.319 --> 00:20:25.079
take it to its conclusions.
326
00:20:25.599 --> 00:20:28.440
Ah, this was a tough This was a tough one.
327
00:20:28.599 --> 00:20:28.759
Yeah.
328
00:20:28.759 --> 00:20:31.640
Should we get to what he actually said? I think
329
00:20:31.640 --> 00:20:35.519
everyone knows the differential stuff of language, right, that's the
330
00:20:35.640 --> 00:20:42.720
basic so Surian discovery or I mean, now it's a discovery.
331
00:20:43.200 --> 00:20:48.000
The hypothesis is that language or words, and the meanings
332
00:20:48.000 --> 00:20:54.119
of terms and the meanings of utterances don't have internal meaning,
333
00:20:54.240 --> 00:20:59.000
they have external differential meaning. So when you're talking about
334
00:20:59.039 --> 00:21:03.319
the value of a Now the the contrary position to
335
00:21:03.440 --> 00:21:06.079
this is that you know what a tree is because
336
00:21:06.079 --> 00:21:08.759
you've seen a lot of trees, So you're copying your
337
00:21:08.799 --> 00:21:11.799
mental impression of a tree when you say tree. That's
338
00:21:11.839 --> 00:21:17.319
sort of a positivistic or a referentialistic understanding of what
339
00:21:17.440 --> 00:21:22.759
language is, or representational representational I think that's the term
340
00:21:22.799 --> 00:21:25.279
he uses. But what we're trying to say here with
341
00:21:25.400 --> 00:21:29.279
a differential view of language is that languages get their
342
00:21:29.359 --> 00:21:32.519
meaning from not just what's around them, but from what
343
00:21:32.640 --> 00:21:36.799
they are not. So it's not it's not that you
344
00:21:36.920 --> 00:21:40.599
refer to the mental image in someone else's head. That
345
00:21:40.640 --> 00:21:44.559
would be a referentialist view, but rather that because of
346
00:21:44.599 --> 00:21:46.599
all the things that it can't be in the context
347
00:21:46.599 --> 00:21:49.079
in which you're using it, then it's got to have
348
00:21:50.160 --> 00:21:53.519
a negative meaning where it doesn't mean any of the
349
00:21:53.519 --> 00:21:56.240
other words that you could have said in that context.
350
00:21:57.160 --> 00:21:59.400
And you need you need this kind of explanation. I mean,
351
00:21:59.400 --> 00:22:01.720
we're not going to cover so sorry at this point,
352
00:22:01.960 --> 00:22:07.279
but you need you need a differentialist view of language
353
00:22:07.319 --> 00:22:10.799
to make sense of certain utterances like that would they
354
00:22:10.839 --> 00:22:15.160
wouldn't make sense in a referentialist view, Like the word
355
00:22:15.240 --> 00:22:19.400
aw aw doesn't refer to anything, but it's completely understandable
356
00:22:19.880 --> 00:22:22.279
to a person in which context you would use, in
357
00:22:22.319 --> 00:22:25.319
which context it's inappropriate to use, or in which context
358
00:22:25.920 --> 00:22:29.559
it's lying to say ow as anyone with two year
359
00:22:29.599 --> 00:22:31.079
olds knows.
360
00:22:31.119 --> 00:22:35.960
Again little a little rusty, but yeah when he divides
361
00:22:36.000 --> 00:22:40.000
it into the signifier and the signified divides a sign right.
362
00:22:41.079 --> 00:22:45.319
And the value value is not just another word for meaning.
363
00:22:47.160 --> 00:22:51.279
Value is something that a term is given based on
364
00:22:51.440 --> 00:22:58.119
its position in a network in relation to a bunch
365
00:22:58.160 --> 00:23:02.400
of other terms that clus are around it for various reasons,
366
00:23:03.279 --> 00:23:07.799
either because they're similar phonetically or they're similar in their meaning.
367
00:23:08.839 --> 00:23:12.240
But value itself doesn't. Isn't just a synonym for meaning.
368
00:23:13.759 --> 00:23:14.720
Value is.
369
00:23:16.599 --> 00:23:21.640
Yeah, Like value could be grouped by similar similarities of meaning.
370
00:23:21.799 --> 00:23:26.279
So you could say, if you're going to construct a sentence,
371
00:23:26.480 --> 00:23:28.559
you might use the word dog, or you might use
372
00:23:28.559 --> 00:23:30.680
the word hound, or you might use the word pooch,
373
00:23:30.799 --> 00:23:34.599
or you might use some other term knine like, doesn't matter.
374
00:23:34.640 --> 00:23:39.079
They're just groups like that, And then you know, then
375
00:23:39.119 --> 00:23:43.839
there's the phonetic similarities between words as well, dog, fog, god,
376
00:23:44.119 --> 00:23:47.599
dog backwards. You could you have those kinds of clusters
377
00:23:47.599 --> 00:23:50.799
as well, and you can get a change in the
378
00:23:50.920 --> 00:23:54.599
value of a term without its meaning or its phonetics
379
00:23:54.640 --> 00:23:58.640
even changing as well. I think I saw an example
380
00:23:58.680 --> 00:24:02.240
of like a hot and cold. If you have a
381
00:24:02.319 --> 00:24:05.000
language that only has hot and cold, and then suddenly
382
00:24:05.039 --> 00:24:09.200
the word freezing is introduced into that language. Cold is
383
00:24:09.200 --> 00:24:11.920
going to change its value, but not its meaning. But
384
00:24:12.039 --> 00:24:16.680
suddenly you're gonna have three terms freezing, cold, and hot. Yeah,
385
00:24:16.720 --> 00:24:20.720
it's the value that the structure itself confers, a value
386
00:24:21.200 --> 00:24:26.279
upon each individual term. That's why it's like, that's the
387
00:24:26.319 --> 00:24:30.880
differential view. I guess there's no, it's not these positive terms.
388
00:24:30.880 --> 00:24:32.480
It's it's a network of differences.
389
00:24:34.119 --> 00:24:37.160
Yeah, so bark, if you're saying, hey, look, bark, what
390
00:24:37.279 --> 00:24:41.079
you're saying by saying hey, look, bark is also saying
391
00:24:41.880 --> 00:24:49.480
not tree, not leaves, not forest. Uh. And also in
392
00:24:49.519 --> 00:24:54.200
the other direction and the other access not meow, not speak,
393
00:24:54.400 --> 00:24:59.000
not mouth, not dog. You're saying bark in whichever context
394
00:24:59.079 --> 00:25:02.480
bark is trying to say nothing nothing else in this
395
00:25:02.599 --> 00:25:05.599
context is what I'm talking about. But this thing, So
396
00:25:05.680 --> 00:25:12.559
it's differentiating, and uh yeah, this is this is the
397
00:25:12.640 --> 00:25:18.279
way llms basically work at root, And the fact that
398
00:25:18.319 --> 00:25:21.079
they can work so well gives some credence to the
399
00:25:21.559 --> 00:25:25.920
differential theory as opposed to the representational theory.
400
00:25:26.559 --> 00:25:29.359
Before when it was just the human model because AI
401
00:25:29.519 --> 00:25:34.960
wasn't really capable of doing this or it wasn't largely recognized.
402
00:25:35.559 --> 00:25:38.400
I mean, yeah, you just had again the human use
403
00:25:38.440 --> 00:25:42.920
of language. But now that lms are fully capable of
404
00:25:44.839 --> 00:25:50.200
using language according to its linguistic value. Because so Sarah
405
00:25:50.359 --> 00:25:53.279
was a linguist first and foremost, and trying to figure
406
00:25:53.319 --> 00:25:57.000
out a way to study language, and it wasn't a
407
00:25:57.039 --> 00:26:00.720
philosopher of language, like he wasn't interested in meaning and
408
00:26:00.759 --> 00:26:03.559
the interaction between words and things in the world. He
409
00:26:03.680 --> 00:26:09.079
was maybe a philosopher of linguistics how to study language
410
00:26:09.720 --> 00:26:14.880
itself as like an isolated system, as an important human
411
00:26:14.920 --> 00:26:19.720
institution for so, sir. But now lllms can use language,
412
00:26:19.759 --> 00:26:22.640
it can be fully fully capable language users.
413
00:26:23.680 --> 00:26:26.640
Yeah, that's that's going to be an upsetting thing for
414
00:26:26.680 --> 00:26:29.440
all the people who say, well, it can't do the
415
00:26:29.559 --> 00:26:32.200
It can't write a poem with feeling, it can't make
416
00:26:32.519 --> 00:26:34.519
it can't make a pretty picture of the way human
417
00:26:34.559 --> 00:26:39.839
would or can it. I have to say I was.
418
00:26:39.799 --> 00:26:41.279
Just gonna say this is the part that I'm actually
419
00:26:41.279 --> 00:26:43.640
pretty sympathetic of about the paper, Like I think that
420
00:26:43.640 --> 00:26:47.680
that's true. I think like it's you know, I think
421
00:26:48.119 --> 00:26:50.359
I think a lot of these like people are going
422
00:26:50.440 --> 00:26:53.680
to be quite shocked and distressed by how good these
423
00:26:53.759 --> 00:26:56.160
language models can probably make art in the future.
424
00:26:56.759 --> 00:26:59.960
There is something with the I mean, I don't know,
425
00:27:00.319 --> 00:27:04.279
you could tell I told I told Cloud to talk
426
00:27:04.519 --> 00:27:07.839
to write a paragraph in the style of Thomas Pinchion,
427
00:27:08.680 --> 00:27:11.920
and may I can I can't separate myself from my bias.
428
00:27:11.960 --> 00:27:14.119
But it was not even close. It was like quite
429
00:27:15.160 --> 00:27:17.359
quite bad. But that I guess it doesn't mean it's
430
00:27:17.359 --> 00:27:17.880
stuck there.
431
00:27:18.440 --> 00:27:19.519
Yeah, it keeps getting better.
432
00:27:19.960 --> 00:27:23.200
Let's say something about how lllms work. And this is
433
00:27:23.279 --> 00:27:27.400
not a non expert way I did. I did some
434
00:27:27.400 --> 00:27:31.240
some research into how they actually spit out the answer
435
00:27:31.319 --> 00:27:34.400
back at you, and this chapter has a bunch contains
436
00:27:34.440 --> 00:27:36.799
a bunch of that information too. A lot of this
437
00:27:36.880 --> 00:27:40.880
chapter is actually about like neural net, a specific neural
438
00:27:40.880 --> 00:27:44.400
net that pre dates uh chat GBT. And it's surprising
439
00:27:44.440 --> 00:27:47.960
they got they got pretty far before the chat GYBT
440
00:27:48.519 --> 00:27:52.039
model came out that everyone got so excited about. And
441
00:27:52.119 --> 00:27:55.039
I do have to say, like absolutely no offense to
442
00:27:55.079 --> 00:27:58.720
mister Weatherby here, but if you this book is written
443
00:27:59.519 --> 00:28:06.319
in like such a dissertation argument style, that is, this
444
00:28:06.359 --> 00:28:08.440
guy said this, and then this other guy said this,
445
00:28:08.519 --> 00:28:10.920
Now allow me to present my intervention in the field.
446
00:28:11.000 --> 00:28:15.200
So it's very very grad student really written, probably because
447
00:28:15.200 --> 00:28:19.279
it probably because it had to be, but it's it's
448
00:28:19.319 --> 00:28:22.279
a little circuitous to get to the point. But there's
449
00:28:22.319 --> 00:28:25.319
a big problem. There's a big problem with humans, with
450
00:28:25.440 --> 00:28:30.599
us being able to imagine how they how the computer
451
00:28:31.160 --> 00:28:33.519
gets to these connections. Because we can only think in
452
00:28:33.599 --> 00:28:37.160
two or three we can only imagine things in two
453
00:28:37.200 --> 00:28:41.880
dimensions are three basically, So if you imagine the term
454
00:28:42.519 --> 00:28:47.319
tree at the center, then pretty nearby in the linguistic field,
455
00:28:47.799 --> 00:28:51.640
field is a rolling Bart's name for it, but pretty
456
00:28:51.680 --> 00:28:55.079
near is like things like oak and bush and forest
457
00:28:55.160 --> 00:28:58.920
and barks and leaves, So they cluster according to context,
458
00:28:59.039 --> 00:29:01.119
like what else are you likely to have in a
459
00:29:01.200 --> 00:29:04.559
sentence next to tree, And there's a bunch of options,
460
00:29:04.839 --> 00:29:07.000
And then a little bit further the way but still close,
461
00:29:07.079 --> 00:29:10.839
you might have things like climbing and lumberjack and squirrel.
462
00:29:11.359 --> 00:29:13.480
And then really far away is something that you would
463
00:29:14.400 --> 00:29:17.119
be surprised to see in the same sentence, which is
464
00:29:18.119 --> 00:29:22.359
institution or or keyboard or wheel or just things that
465
00:29:22.799 --> 00:29:27.680
wouldn't typically be associated with trees in a normal English
466
00:29:27.680 --> 00:29:30.680
sentence or whichever language you're writing it. So we can
467
00:29:30.759 --> 00:29:35.039
imagine that in two or three D space that there's
468
00:29:35.200 --> 00:29:39.960
these clusters, But what an LM does, or even the
469
00:29:40.000 --> 00:29:43.920
neural net that he's talking about, there are hundreds or
470
00:29:43.960 --> 00:29:47.160
even thousands of dimensions, so it's not two D space,
471
00:29:47.359 --> 00:29:53.880
not three D space. It's one thousand dimensional space. And
472
00:29:53.920 --> 00:29:56.319
that's why I mean this is it's a failure of
473
00:29:56.319 --> 00:30:01.720
our imagination. That's kind of disappointing. But so Sir, So
474
00:30:01.720 --> 00:30:07.720
Sir and Jakobsen had like these two axes view of language,
475
00:30:07.839 --> 00:30:12.759
and now we have JGBT working with it one four
476
00:30:13.359 --> 00:30:16.960
dimensional view of language. So that's how you can have
477
00:30:17.079 --> 00:30:23.680
something like cone is in the same area as parabola,
478
00:30:23.960 --> 00:30:26.599
but the same cone is also on a different on
479
00:30:26.680 --> 00:30:29.599
a different dimension next to ice cream, and in a
480
00:30:29.640 --> 00:30:35.279
different dimension is next to what else is cones in
481
00:30:35.319 --> 00:30:38.799
it pyramid or a different sort of shape. So this
482
00:30:38.920 --> 00:30:43.200
is called embedding space. This is not something that so
483
00:30:43.240 --> 00:30:45.759
Sir would have been able to think of, but he
484
00:30:45.799 --> 00:30:52.640
basically conceptualized it without knowing. This is like difference difference
485
00:30:52.640 --> 00:30:57.559
across one thousand dimensions. So the LM, sorry, the LM
486
00:30:57.640 --> 00:31:00.640
gets the data sets, train on the data set, and
487
00:31:00.680 --> 00:31:03.519
it's given the dimension space. But then based on its
488
00:31:03.559 --> 00:31:08.319
training data, it's able to break down every not even
489
00:31:08.319 --> 00:31:11.319
a full word, but every part of a word into
490
00:31:11.359 --> 00:31:14.440
where it fits in this whole thing, and then when
491
00:31:14.440 --> 00:31:16.519
it needs to repiece them together to give a connection
492
00:31:16.640 --> 00:31:19.720
back to you, it pulls again from all of those dimensions,
493
00:31:19.759 --> 00:31:22.400
so we can't picture it. But the point is that
494
00:31:23.839 --> 00:31:28.319
mister what's his name again, weather be weatherb is pointing
495
00:31:28.319 --> 00:31:33.319
to is that this embedding space is differentially constructed and
496
00:31:33.359 --> 00:31:37.279
it's not understandable by a human but in principle we
497
00:31:37.279 --> 00:31:39.440
can kind of see what it means because it's how
498
00:31:39.480 --> 00:31:43.200
so sur said that we understand meaning, Sorry that was long.
499
00:31:44.920 --> 00:31:48.680
No, it's good because my structuralism is a little bit
500
00:31:48.759 --> 00:31:53.880
rusty and his argument is quite dense. There's those those
501
00:31:53.880 --> 00:31:57.920
two acts, those axes you were talking about, the axis
502
00:31:58.000 --> 00:32:04.839
of uh combination and the axis of selection. Like, I
503
00:32:05.119 --> 00:32:10.079
think everyone's basically doing that. AI is doing that when
504
00:32:10.119 --> 00:32:15.119
it's constructing language, I think as are human beings they
505
00:32:15.440 --> 00:32:18.880
because that's what else can you do. You're you're creating
506
00:32:18.920 --> 00:32:23.960
a string a sequence of words, and you're selecting what
507
00:32:24.119 --> 00:32:27.079
goes where, and you're sequencing them, so you're combining them.
508
00:32:27.440 --> 00:32:30.880
That's all you can do. But when it it's.
509
00:32:30.720 --> 00:32:32.480
You got to check out the chapter. It has a
510
00:32:32.480 --> 00:32:33.599
whole section on this.
511
00:32:35.079 --> 00:32:38.480
But when it's mapping the semantic field, yes, it's doing
512
00:32:38.519 --> 00:32:41.880
it like that. It's clustering terms like he gives these
513
00:32:41.920 --> 00:32:49.480
examples like paradigmatically, you have kid, taught and youngster and
514
00:32:49.720 --> 00:32:53.680
onto infinity whatever else clusters around that meaning of like
515
00:32:53.839 --> 00:32:58.319
boy or child. And but then suddenly you have all
516
00:32:58.359 --> 00:33:02.359
these intersections, like kid also has something to do with animals.
517
00:33:02.599 --> 00:33:05.680
Kid is the name of like a goat too, is
518
00:33:05.680 --> 00:33:08.359
it a goat? A young goat is called a kid
519
00:33:08.839 --> 00:33:12.440
as well, So then it's intersected by this other dimension.
520
00:33:13.599 --> 00:33:17.920
And AI can map out this like extremely dense dimensional
521
00:33:18.279 --> 00:33:23.160
semantic field of language that humans can't really do. But
522
00:33:23.200 --> 00:33:27.920
when they're actually generating the text, that's what they're selecting from. Plus,
523
00:33:28.000 --> 00:33:31.920
there's the combining rules, which is like syntax, So you
524
00:33:31.960 --> 00:33:38.160
have the syntagmatic which takes place along the axis of combination,
525
00:33:39.480 --> 00:33:43.680
which is basically just the rules of combination, like what
526
00:33:43.960 --> 00:33:49.000
cone often appears after traffic, so there's a certain statistical
527
00:33:49.039 --> 00:33:53.039
correspondence there because of traffic cone. But then I guess
528
00:33:53.319 --> 00:33:59.039
head might be another combinatory possibility cone head. Right, It's
529
00:33:59.079 --> 00:34:02.359
that you're mapping all the like combinatory possibilities as well,
530
00:34:02.519 --> 00:34:03.559
and AI is just going to.
531
00:34:03.599 --> 00:34:06.160
Be really really good at that. Right.
532
00:34:06.519 --> 00:34:10.159
Apart from the paradigmatic clusterings of terms with similar sounds
533
00:34:10.239 --> 00:34:13.400
or similar meanings. You have the clustering of terms based
534
00:34:13.400 --> 00:34:18.119
on the frequency that they occur together, and you're basically
535
00:34:18.199 --> 00:34:21.639
doing it syntactically at that point as well. So, but
536
00:34:21.679 --> 00:34:25.239
that's that's the basic That's how I'm understanding how how
537
00:34:25.960 --> 00:34:31.599
structuralism tried to explain language. And then AI comes along
538
00:34:31.800 --> 00:34:36.800
and just proves that Structuralism's explanation was like basically on point,
539
00:34:37.960 --> 00:34:40.519
even though AI does it a little differently in his way,
540
00:34:41.320 --> 00:34:43.360
more adept in certain areas.
541
00:34:44.719 --> 00:34:48.000
Yeah. So, uh, this is where I said at the beginning,
542
00:34:48.039 --> 00:34:52.280
it actually went further than structuralism, and that's because so
543
00:34:52.320 --> 00:34:56.360
Sir reduced it, and actually every structuralist did this in
544
00:34:56.440 --> 00:35:00.599
their own terms. The Acobsen I think metaphorm autonomy were
545
00:35:00.639 --> 00:35:03.400
his terms, but they had these two axes of language.
546
00:35:03.480 --> 00:35:07.800
One is basically the word order and then whichever language
547
00:35:07.800 --> 00:35:12.159
you're writing in, like you have the usually have the subject,
548
00:35:12.199 --> 00:35:14.159
then the verb than the noun, and like.
549
00:35:14.159 --> 00:35:18.199
That's that or the principle of contiguity. That's sometimes referred
550
00:35:18.199 --> 00:35:19.119
to as well as.
551
00:35:19.320 --> 00:35:23.320
Yeah, that's another guys, everyone's got two axes. Yeah, yeah,
552
00:35:23.440 --> 00:35:28.239
So the paradigmatic is the vertical and the syntigmatic is
553
00:35:28.280 --> 00:35:32.360
the horizontal, and they basically all the structuralists agreed on this.
554
00:35:32.920 --> 00:35:37.119
But mister Weatherby here says, if you actually, if you tell,
555
00:35:38.159 --> 00:35:41.199
if you tell, this isn't even Chagibt three yet, this
556
00:35:41.320 --> 00:35:43.440
was just the neural net. If you tell the neural
557
00:35:43.519 --> 00:35:46.960
net to follow these rules, it actually ends up doing worse.
558
00:35:47.440 --> 00:35:51.199
The best way is to let it do. Let it
559
00:35:51.239 --> 00:35:56.079
do predictions based on its training data set, and it
560
00:35:56.159 --> 00:36:01.280
will figure out the word order without syntagmatic without a
561
00:36:01.280 --> 00:36:06.400
syntagmatic access. So he's he basically says contra Jakobsen, and
562
00:36:06.760 --> 00:36:11.920
so sir, there's only one procedure that can do both
563
00:36:12.760 --> 00:36:16.039
syntax and semantics at the same time, which I mean
564
00:36:16.039 --> 00:36:19.199
that was totally mind blowing to me. I was like,
565
00:36:19.280 --> 00:36:22.920
holy shit, that's crazy. So that means there's like one
566
00:36:24.639 --> 00:36:29.360
there's one axis in language, one access through one thousand
567
00:36:29.400 --> 00:36:35.480
dimensional space. That's crazy. So it doesn't There was a
568
00:36:35.519 --> 00:36:38.320
guy I can't remember the guy, but he's training this
569
00:36:38.320 --> 00:36:41.639
this neural net, and he has a funny quip that says,
570
00:36:41.679 --> 00:36:45.400
every time I fired a linguist, the neural net did
571
00:36:45.400 --> 00:36:49.800
better at spitting out a coherent natural language because all
572
00:36:49.840 --> 00:36:52.400
the things that humans think language is about all the
573
00:36:52.480 --> 00:36:55.599
rules we infer, all the groupings of verbs versus ads
574
00:36:56.039 --> 00:37:00.599
versus noun. It gets in the way because rules can
575
00:37:00.639 --> 00:37:02.960
be broken and even if you follow the rules sometimes
576
00:37:02.960 --> 00:37:05.119
like don't end a sentence with a preposition. If you
577
00:37:05.360 --> 00:37:08.000
if you follow the rules to a t, you sound
578
00:37:08.079 --> 00:37:10.400
weird because we don't follow the rules to a t.
579
00:37:10.960 --> 00:37:14.760
So the more the more guidelines he took off of
580
00:37:14.800 --> 00:37:18.519
the neural net, the more natural it sounded because it
581
00:37:18.559 --> 00:37:20.800
was just copying speech.
582
00:37:21.599 --> 00:37:23.239
It was able to it was able to figure out.
583
00:37:23.280 --> 00:37:26.159
It was able to figure out patterns like yeah, they
584
00:37:26.199 --> 00:37:27.840
don't even know that. We don't even know that we're
585
00:37:27.840 --> 00:37:31.039
doing them, so we don't even have words for you know,
586
00:37:31.280 --> 00:37:33.559
it's not like, yeah, we break rules all the time,
587
00:37:33.599 --> 00:37:36.519
but there is there are these patterns that are discernible.
588
00:37:37.199 --> 00:37:38.920
We just don't know what they are. And if you
589
00:37:39.039 --> 00:37:41.199
just like let let a neural net figure them out,
590
00:37:41.199 --> 00:37:42.920
they'll figure them out and do better than if we
591
00:37:42.960 --> 00:37:48.400
try to impose rules, like like kind of yeah, like
592
00:37:48.559 --> 00:37:51.920
if we if we constrain it with rules first, it's
593
00:37:51.960 --> 00:37:53.039
just it's going to miss all that.
594
00:37:53.960 --> 00:37:57.000
So like when to say boy versus lad, they are
595
00:37:57.119 --> 00:37:59.480
almost identical, meaning and if you ask an English speaker
596
00:38:00.119 --> 00:38:03.039
to use the word lad versus boy, probably you wouldn't
597
00:38:03.440 --> 00:38:07.119
even know. You don't know the rules. Yeah, you know
598
00:38:07.159 --> 00:38:10.840
what the boys or if you call someone boy, it
599
00:38:10.880 --> 00:38:12.840
can be like disparaging.
600
00:38:12.239 --> 00:38:16.079
Right, yeah, exactly. There's actually there's a there's a a
601
00:38:16.119 --> 00:38:21.400
professor at what's it called. There was like the this
602
00:38:21.519 --> 00:38:24.960
college at uf T University of Toronto. It's like graduate
603
00:38:25.000 --> 00:38:27.119
student college. You have to like pay to be part
604
00:38:27.119 --> 00:38:29.079
of it. It's like a weird like elite club. And
605
00:38:29.559 --> 00:38:32.760
at one of their dinners some professor called like a
606
00:38:33.000 --> 00:38:35.119
like a student of color boy and it became like
607
00:38:35.159 --> 00:38:40.639
a huge controversy and yeah, especially my drink boy.
608
00:38:41.679 --> 00:38:42.400
Yeah garsol.
609
00:38:43.159 --> 00:38:53.360
So, so one one operation can perform both the syntactical
610
00:38:53.400 --> 00:38:57.039
and the semantic, or can map it out. You're just
611
00:38:57.079 --> 00:39:01.159
mapping out this, uh this like and if well, I
612
00:39:01.159 --> 00:39:04.440
guess it's again it goes back to Solcier's idea of
613
00:39:04.920 --> 00:39:09.559
value though, like it's it's mapping out linguistic value. But
614
00:39:09.599 --> 00:39:16.000
that is powerful enough on its own to do the
615
00:39:16.039 --> 00:39:20.920
work of both syntactics and semantics. So how you might
616
00:39:21.159 --> 00:39:23.960
combine the words together and how what they might mean,
617
00:39:24.119 --> 00:39:26.199
Like the equivalences between the words.
618
00:39:27.599 --> 00:39:29.599
Well, that's what it says. It does. It does all
619
00:39:29.639 --> 00:39:31.880
that before it selects the word. So we think of
620
00:39:31.920 --> 00:39:36.320
it like, oh, I said, I said boy, But I
621
00:39:36.400 --> 00:39:38.920
could have said garson, or I could have said lad,
622
00:39:39.000 --> 00:39:40.360
or I could have said bro, or I could have
623
00:39:40.400 --> 00:39:43.719
said a different word. But that's not the way that
624
00:39:44.599 --> 00:39:48.639
LLLM does it. The LLLM. So this is what happens. Okay,
625
00:39:50.239 --> 00:39:54.119
if I understand it correctly, I very likely I'm incorrect
626
00:39:54.159 --> 00:39:58.000
on some of the details here. But for every word,
627
00:39:58.039 --> 00:40:00.440
and it doesn't even think in words, it thinks into tokens.
628
00:40:00.719 --> 00:40:05.599
But for every token it can put down, it compares
629
00:40:05.679 --> 00:40:08.920
that token not just to every other part of the
630
00:40:08.960 --> 00:40:12.639
training data where that token appears, So not just every
631
00:40:12.679 --> 00:40:15.679
sentence in which the word it appears, it also compares
632
00:40:15.679 --> 00:40:19.159
it to every sentence that it doesn't appear in and
633
00:40:19.400 --> 00:40:24.280
inferentially understands why that word was not used in that sentence.
634
00:40:24.679 --> 00:40:27.519
So it's going to have a bunch of sentences with
635
00:40:27.599 --> 00:40:29.519
the word boy in it. It's also going to have
636
00:40:29.519 --> 00:40:32.039
a bunch of sentences with the word lad in it,
637
00:40:32.440 --> 00:40:35.280
So it differentiates based on that. But then it's also
638
00:40:35.360 --> 00:40:37.320
going to have a bunch of training data that has
639
00:40:37.360 --> 00:40:39.840
the word scissors in it and go, Okay, they're not
640
00:40:39.960 --> 00:40:44.159
using They're not using scissors in any of these sentences,
641
00:40:44.199 --> 00:40:45.880
So I'm not going to use scissors when I'm talking
642
00:40:45.920 --> 00:40:49.440
about boy. But that's how it decides whether to use
643
00:40:49.519 --> 00:40:52.840
boy or lad. And this is this is not giving
644
00:40:53.719 --> 00:40:56.360
I don't want to sound like I'm hyping CHATGYBT up
645
00:40:56.400 --> 00:40:58.960
for being so ingenious, Like it's if anything, it's more
646
00:40:59.000 --> 00:41:02.599
amazing that we can do all this stuff instantaneously without
647
00:41:02.599 --> 00:41:03.519
even thinking about it.
648
00:41:04.239 --> 00:41:04.440
Yeah.
649
00:41:05.159 --> 00:41:09.199
It compares, It compares any potential usage to a term
650
00:41:09.360 --> 00:41:12.760
before it puts the term down to every sentence that
651
00:41:12.800 --> 00:41:16.320
the term is not in. So that's why it's not
652
00:41:16.519 --> 00:41:19.760
the same as the paradigmatic syntigmatic, because we in a
653
00:41:19.800 --> 00:41:24.199
paradigmatic syntigmatic you put it down thinking, Okay, I'm choosing
654
00:41:24.360 --> 00:41:26.639
from the other words that I could have put there instead.
655
00:41:27.199 --> 00:41:34.239
This does that prior just by comparing one trillion instances
656
00:41:34.320 --> 00:41:36.239
or actually it's five what is it, five hundred million,
657
00:41:36.519 --> 00:41:40.519
five hundred million other sentences and seeing how those sentences
658
00:41:40.559 --> 00:41:43.480
were built in order to decide whether to use the
659
00:41:43.480 --> 00:41:47.639
word boy or lad. It's not choosing between the two. Right,
660
00:41:47.679 --> 00:41:48.880
it's a pre made decision.
661
00:41:49.280 --> 00:41:52.840
And that's why a lot of people there's like discussion
662
00:41:53.920 --> 00:41:58.320
among lllm P like engineers that they don't exactly understand
663
00:41:58.320 --> 00:42:03.400
why makes the choices that it makes because it's comparing
664
00:42:03.480 --> 00:42:06.000
so many different things and then kind of like making
665
00:42:06.079 --> 00:42:11.079
pattern judgments based on so many vectors that then it
666
00:42:11.119 --> 00:42:14.400
comes up with this decision based on obviously feedback it
667
00:42:14.440 --> 00:42:18.039
gets when it's spitting stuff out, so there is like
668
00:42:18.079 --> 00:42:21.079
a feedback element also bringing humanism back in. There's a
669
00:42:21.119 --> 00:42:23.199
feedback element where the humans saying that's right, that's wrong,
670
00:42:23.239 --> 00:42:27.599
that's right, that's wrong. But ultimately it is like detecting
671
00:42:27.639 --> 00:42:30.360
these patterns at like a crazy level and making connections
672
00:42:30.360 --> 00:42:32.239
that we would never make. And still some of these
673
00:42:32.239 --> 00:42:34.920
engineers don't really understand. And I think it's in that
674
00:42:34.960 --> 00:42:37.800
sense that you're saying it kind of goes beyond structural
675
00:42:37.880 --> 00:42:39.559
because it's just figuring out things.
676
00:42:40.199 --> 00:42:45.599
Oh I meant very specifically, the lowest denominator that a
677
00:42:45.679 --> 00:42:48.840
structuralist could imagine is two axes. Oh yes, this has
678
00:42:49.440 --> 00:42:52.719
this has no axes, or it has one axis, but
679
00:42:52.760 --> 00:42:56.519
the axis is like five hundred million comparison points.
680
00:42:56.599 --> 00:42:57.760
Yes, yes, exactly.
681
00:42:57.800 --> 00:43:01.400
So the implication of this and in normal speaking terms,
682
00:43:01.559 --> 00:43:05.119
is there's no such thing as a word without a
683
00:43:05.519 --> 00:43:08.760
without a sentence. And then there's no such thing without
684
00:43:08.800 --> 00:43:12.639
a word without a sentence compared to every other sentence
685
00:43:12.639 --> 00:43:17.719
in the language. And that's not exactly. I don't I
686
00:43:17.719 --> 00:43:20.440
don't know how to say this properly, but it's we
687
00:43:21.280 --> 00:43:25.159
do this without knowing that we're doing it.
688
00:43:25.599 --> 00:43:26.280
Yeah, exactly.
689
00:43:26.639 --> 00:43:29.440
We know when a word is correct to say versus
690
00:43:29.599 --> 00:43:35.719
incorrect to say, because for the first fifteen twenty years
691
00:43:35.760 --> 00:43:38.800
of your life, you're having people around you saying, yes,
692
00:43:38.840 --> 00:43:41.039
that's the correct word. No, that's not the correct word. Yes,
693
00:43:41.079 --> 00:43:42.800
that's the correct word. No you mean this, Yes you
694
00:43:42.800 --> 00:43:44.360
mean this, No you mean this, No you mean yes,
695
00:43:44.360 --> 00:43:46.440
I mean this. Your parents start with that. Then you
696
00:43:46.480 --> 00:43:48.079
go out and you say something stupid at school and
697
00:43:48.159 --> 00:43:50.760
you get made fun of. You're correcting your you're correcting
698
00:43:50.760 --> 00:43:56.480
your terms. So you're comparing it unconsciously to this entire structure,
699
00:43:57.480 --> 00:44:00.360
and the structure doesn't belong to you. The structurable is
700
00:44:01.239 --> 00:44:06.400
held in common nowhere, it's virtually real. But yeah, I mean,
701
00:44:06.880 --> 00:44:09.480
the implication that there's no such thing as a term
702
00:44:09.920 --> 00:44:11.880
without a sentence, and there's no such thing as this
703
00:44:12.440 --> 00:44:16.519
as a sentence without all the other sentences that not
704
00:44:16.599 --> 00:44:19.320
just have been uttered but could be uttered. That just
705
00:44:19.840 --> 00:44:23.760
it's pretty incredible that all that's going on this.
706
00:44:23.880 --> 00:44:26.960
Is But then that's long, isn't it. That's what so
707
00:44:27.000 --> 00:44:32.239
sir recommended we study like as as linguists study long.
708
00:44:33.320 --> 00:44:37.800
Not not parole, not the not the actual act of speaking,
709
00:44:38.599 --> 00:44:43.960
but long, the totality of language, not not as it
710
00:44:44.000 --> 00:44:47.559
exists in any like single person's head, but as a
711
00:44:47.559 --> 00:44:51.559
social thing, as a as a social or a collective entity,
712
00:44:52.760 --> 00:44:57.159
of which we all have like some imperfect copy of
713
00:44:57.199 --> 00:45:00.119
it in our heads, based on like our vocabulary, in
714
00:45:00.159 --> 00:45:04.199
our rules and everything. And so it sounds like he's
715
00:45:04.199 --> 00:45:06.599
saying there's a sort of I could be wrong on this,
716
00:45:06.719 --> 00:45:08.719
but it sounds like he's saying that there's a kind
717
00:45:08.719 --> 00:45:14.440
of analogy there between long, what the linguistics should study,
718
00:45:15.320 --> 00:45:18.440
what the linguist should study, and and what AI has
719
00:45:18.920 --> 00:45:22.400
trained on. In a way, what AI has in its
720
00:45:22.519 --> 00:45:24.840
training and what it has access to seems to be
721
00:45:24.960 --> 00:45:28.639
like a version of this long and then it's specific.
722
00:45:28.719 --> 00:45:31.159
You prompt it, it gives you an answer like that's parole.
723
00:45:31.280 --> 00:45:33.800
That's not what we're studying, but that's what generates the
724
00:45:33.920 --> 00:45:37.440
data that turns into that that we infer long from.
725
00:45:37.480 --> 00:45:41.280
Because long is that virtual thing. It's it's not anywhere
726
00:45:41.320 --> 00:45:45.519
and no one has it like AI gets that, AI
727
00:45:45.920 --> 00:45:50.199
gets that trained, or AI gets to that thing that
728
00:45:50.239 --> 00:45:54.159
the linguists are trying to study. That are that that's so, sir,
729
00:45:54.320 --> 00:45:57.559
theorized that linguistics should study.
730
00:45:57.800 --> 00:46:01.320
You're right that long is virtual and it's every where.
731
00:46:01.400 --> 00:46:07.320
It's in it's behind every expression, but not in the
732
00:46:07.400 --> 00:46:12.039
pieces of the expression. So what what training data is
733
00:46:12.039 --> 00:46:19.000
is just five hundred million paroles. And we a human
734
00:46:19.039 --> 00:46:22.239
looks at a powerhole and if you look at it carefully,
735
00:46:22.239 --> 00:46:25.239
you can go, Okay, I'm starting to see the structure.
736
00:46:25.599 --> 00:46:28.000
I'm starting to see the structure of meaning. And the
737
00:46:28.000 --> 00:46:31.199
structure of meaning is called long. And we have a
738
00:46:31.199 --> 00:46:34.599
certain way of doing that, like building semiotic squares is
739
00:46:34.639 --> 00:46:38.199
one way, or comparing a whole bunch of fashion outfits.
740
00:46:38.199 --> 00:46:41.679
That's one thing bar Barth did is saying we ring
741
00:46:41.760 --> 00:46:47.480
up lexicons, probably right, and and Chad GBT looks at
742
00:46:47.639 --> 00:46:53.920
five hundred million examples of parole and just buy the
743
00:46:54.000 --> 00:46:57.960
amount of them, like we humans don't do this by amount.
744
00:46:58.079 --> 00:46:59.679
It doesn't seem like you know, I don't know how
745
00:46:59.719 --> 00:47:01.920
humans do it. I don't know if I don't know
746
00:47:01.960 --> 00:47:05.039
if that's known. But we don't have five hundred million
747
00:47:05.039 --> 00:47:10.239
comparisons in our head clearly. Maybe maybe it's unconscious, maybe
748
00:47:10.280 --> 00:47:13.000
it's stored in some way. I don't want to get
749
00:47:13.000 --> 00:47:17.800
into this, into the ram metaphors, but it figures out
750
00:47:17.880 --> 00:47:21.800
long by doing five hundred million comparisons each time it's
751
00:47:21.840 --> 00:47:25.079
about to say something, so we both get to la
752
00:47:25.159 --> 00:47:30.199
long whatever's going on in people. It's pretty crazy. If
753
00:47:30.239 --> 00:47:33.679
that's how much. That's why that's why CHGBT takes so
754
00:47:33.800 --> 00:47:38.599
much compute power, because you've got to do that many
755
00:47:38.599 --> 00:47:42.000
comparisons for every single token that you're going to land there.
756
00:47:42.679 --> 00:47:44.280
I don't know much to say about that. You guys
757
00:47:44.320 --> 00:47:46.880
are more the structuralist guys.
758
00:47:47.159 --> 00:47:48.400
What did you pull out of this? What do you
759
00:47:48.440 --> 00:47:49.199
want to talk about?
760
00:47:50.039 --> 00:47:51.719
I don't know. I don't I mean, I said, like
761
00:47:51.800 --> 00:47:56.199
my main thing. I mean I don't really have it. Uh,
762
00:47:56.400 --> 00:47:57.679
I mean, let me look at my notes, but I
763
00:47:57.679 --> 00:47:59.960
don't have anything that comes to mind.
764
00:48:00.119 --> 00:48:02.000
The outcomes of all this. This might be interesting to
765
00:48:02.039 --> 00:48:06.159
people is because you know there's like tells that something's
766
00:48:06.159 --> 00:48:10.199
written by oh yeah chgbt. One of them is. One
767
00:48:10.199 --> 00:48:12.519
of them is M dashes, which I'm not happy about
768
00:48:12.519 --> 00:48:16.639
because I I've long been a proponent and a utilizer
769
00:48:16.679 --> 00:48:17.639
of the M dash.
770
00:48:18.320 --> 00:48:21.159
I can I can I say something quick about the
771
00:48:21.239 --> 00:48:23.480
M dash because it's funny. I also have been a
772
00:48:23.480 --> 00:48:26.239
longtime user of the M dash. But I realized because
773
00:48:26.280 --> 00:48:27.960
of the age of a that I'd been I'd been
774
00:48:27.960 --> 00:48:32.039
writing them wrong in word, so I've just been using.
775
00:48:31.840 --> 00:48:34.079
The dash all zero, one, five one.
776
00:48:34.679 --> 00:48:36.440
Well. I also realized that you could you could just
777
00:48:36.519 --> 00:48:38.559
press the dash button twice and it would turn into
778
00:48:38.559 --> 00:48:40.000
an M dash. I never did that. I would just
779
00:48:40.039 --> 00:48:41.719
press it once and then I would press space and
780
00:48:41.719 --> 00:48:43.559
start writing the word and it would kind of get bigger.
781
00:48:43.920 --> 00:48:45.760
But it wasn't the full M dash. It's like I was,
782
00:48:45.800 --> 00:48:47.320
I was making the M dash wrong.
783
00:48:47.760 --> 00:48:48.920
It was an M dash.
784
00:48:49.320 --> 00:48:51.119
Yeah, maybe it was, I guess so, So I was
785
00:48:51.159 --> 00:48:53.199
just doing so anyway, just like just like a funny
786
00:48:53.199 --> 00:48:55.119
aside that I was like and I and I and
787
00:48:55.159 --> 00:48:57.599
I remember like pre ll m h I would see
788
00:48:57.599 --> 00:49:00.440
people with like the long M dash with connect the words,
789
00:49:00.440 --> 00:49:01.679
and I'd be like God, I don't even know how
790
00:49:01.679 --> 00:49:03.440
to make that in word. And I was like, oh whatever,
791
00:49:04.239 --> 00:49:06.599
I'll just keep using my one. And then this whole
792
00:49:06.599 --> 00:49:08.159
thing came out and people were like, oh, M dash
793
00:49:08.199 --> 00:49:09.440
is our a tell And I was like, not for me,
794
00:49:09.480 --> 00:49:11.159
because I've been writing them wrong the whole time.
795
00:49:11.480 --> 00:49:12.800
You've been using M dashes.
796
00:49:13.159 --> 00:49:14.119
Yeah, yeah, exactly.
797
00:49:15.159 --> 00:49:19.519
So I guess what he says is now the poetics
798
00:49:19.559 --> 00:49:22.719
of language. I mean, we should go into the poetics
799
00:49:22.920 --> 00:49:27.400
point because it's important, but Language Chat GPT is literally
800
00:49:27.480 --> 00:49:31.280
changing language for the rest of us, because now the
801
00:49:31.320 --> 00:49:34.559
rest of us are like, oh, well, should I actually
802
00:49:34.639 --> 00:49:36.039
use an M dash here or is it going to
803
00:49:36.119 --> 00:49:39.119
seem like I used I used cloud to do this?
804
00:49:39.719 --> 00:49:43.679
Or rhetorical it's the rhetorical device of like not this,
805
00:49:43.880 --> 00:49:47.920
not that, but this like arguments yeah, yeah, you can't
806
00:49:48.000 --> 00:49:49.320
use that anymore, or.
807
00:49:49.360 --> 00:49:52.360
Using too much florid language, like it's better to be
808
00:49:52.480 --> 00:49:55.119
like Hemingway, you know, and get get to the point
809
00:49:55.480 --> 00:49:58.000
so you don't sound like So it's going to change
810
00:49:58.639 --> 00:50:01.719
human writing, so that because humans are trying not to
811
00:50:01.760 --> 00:50:03.400
sound like it. But then it's going to copy that
812
00:50:03.480 --> 00:50:05.519
writing if it's added to the new data set. So
813
00:50:05.599 --> 00:50:09.000
it's going to be a system of co evolution.
814
00:50:10.119 --> 00:50:12.960
I mean, I'm guessing that eventually lms are just going
815
00:50:13.000 --> 00:50:17.920
to become like so diverse in their ability to change
816
00:50:17.920 --> 00:50:20.400
their style that it's going to be pretty indistinguishable, because
817
00:50:20.440 --> 00:50:22.639
right now there is like an LM style. But I imagine,
818
00:50:22.639 --> 00:50:26.360
I mean, for example, I mean, I just can't imagine
819
00:50:26.400 --> 00:50:28.880
that there's not going to be a world where people
820
00:50:28.920 --> 00:50:31.079
could just be like, hey, here's stuff that I've actually
821
00:50:31.119 --> 00:50:35.159
written like myself. Can you just like like learn what
822
00:50:35.199 --> 00:50:39.239
my style is and just write just like me? And like,
823
00:50:39.440 --> 00:50:41.280
I'll but that has to be possible, and then everyone's
824
00:50:41.320 --> 00:50:42.960
just going to have like their own LLLM R like
825
00:50:43.000 --> 00:50:45.000
writing for them that writes and then they'll change it
826
00:50:45.079 --> 00:50:48.119
or whatever like to how they like. But they'll be like,
827
00:50:48.159 --> 00:50:50.599
I obviously have ticks and patterns with how I write,
828
00:50:50.599 --> 00:50:52.719
like you need to copy my pattern and then it's
829
00:50:52.719 --> 00:50:53.880
just going to be indistinguishable.
830
00:50:54.199 --> 00:50:56.599
Well, I think I think one of the points is
831
00:50:56.679 --> 00:51:00.119
like that, you know, insofar as human beings can and
832
00:51:00.639 --> 00:51:05.679
we can generate language like authentically, and we could imitate, right,
833
00:51:05.719 --> 00:51:08.320
we could imitate someone else. We can do an impression,
834
00:51:08.360 --> 00:51:11.400
We can imitate someone else's writing style, or we can
835
00:51:11.480 --> 00:51:14.480
imitate some style and then that I guess that's kind
836
00:51:14.480 --> 00:51:15.280
of inauthentic.
837
00:51:15.320 --> 00:51:15.679
I don't know.
838
00:51:15.800 --> 00:51:19.400
James Joyce really good at imitation and sort of writing
839
00:51:19.400 --> 00:51:20.480
at different registers.
840
00:51:21.199 --> 00:51:22.760
But then, well, when you're learning how to write, you're
841
00:51:22.760 --> 00:51:25.320
always imitating. I mean, when when you start writing, you're
842
00:51:25.360 --> 00:51:27.920
just like looking, you write from what you read, or
843
00:51:27.960 --> 00:51:31.159
you have someone who like reads your I remember when
844
00:51:31.159 --> 00:51:32.719
I was like learning how to write good essays, I
845
00:51:32.760 --> 00:51:35.119
like had my stepbrother and he would like correct my essays,
846
00:51:35.119 --> 00:51:37.599
and I felt like I started internalizing his tics.
847
00:51:37.920 --> 00:51:40.840
Yeah, I heard like Hunter S. Thompson used to type
848
00:51:40.880 --> 00:51:43.239
out the great gat speech just to get like that
849
00:51:43.280 --> 00:51:45.920
writing style down and become a better writer.
850
00:51:46.440 --> 00:51:47.239
But we can't.
851
00:51:47.400 --> 00:51:49.599
I guess one of the points is we can't really
852
00:51:49.679 --> 00:51:53.159
just say that, like, human beings can either produce language
853
00:51:53.280 --> 00:51:58.159
or imitate, but lms can only imitate. No, now we're
854
00:51:58.199 --> 00:52:05.960
saying that lms are using language in a way that
855
00:52:06.119 --> 00:52:07.800
like that. These are some of the I found some
856
00:52:07.840 --> 00:52:11.079
of the quotes I was thinking about when like the
857
00:52:11.079 --> 00:52:16.280
the outputs of llms are valuations of words in Sosure's sense.
858
00:52:17.159 --> 00:52:19.760
So there, when we were talking about so surre and
859
00:52:20.159 --> 00:52:25.639
value as as the sort of differential value of terms
860
00:52:25.639 --> 00:52:28.880
in a system. Then he says, this, I think is
861
00:52:28.920 --> 00:52:32.400
why l ms cannot be said to be doing something
862
00:52:32.519 --> 00:52:38.079
other than language in a strict sense, because when so
863
00:52:38.239 --> 00:52:40.840
Surah talks about language, that's what he says. He says,
864
00:52:41.400 --> 00:52:46.679
language is there are only differences without positive terms. That's
865
00:52:46.679 --> 00:52:50.039
where value comes from, the differences between terms. There's no
866
00:52:50.199 --> 00:52:53.760
positive sort of values in that sense. Then he says,
867
00:52:53.800 --> 00:52:57.239
when we write, now we write in a language system
868
00:52:57.239 --> 00:53:01.360
that includes l l M language. This is not a delusion,
869
00:53:02.159 --> 00:53:05.719
not just the Eliza effect. This is language in the
870
00:53:05.719 --> 00:53:09.000
age of its algorithmic reproducibility.
871
00:53:09.519 --> 00:53:10.920
There's a little little.
872
00:53:10.679 --> 00:53:18.199
Nod towards Benumine there technological reproducibility of language, which ironically
873
00:53:18.280 --> 00:53:24.079
came way after the technological reproducibility of sound and image.
874
00:53:24.239 --> 00:53:28.440
Through photography, we could we could reproduce like the world
875
00:53:28.800 --> 00:53:32.360
with a camera before we could build something that could
876
00:53:32.360 --> 00:53:36.719
reproduce language like a person could. That's that's kind of
877
00:53:37.159 --> 00:53:42.559
another that's that's a tangent anyway. Yeah, yeah, but we
878
00:53:42.559 --> 00:53:46.000
were going towards poetics. But I just want to flag that, Yes,
879
00:53:46.840 --> 00:53:50.000
I think I think lms are sort of realizing and
880
00:53:50.199 --> 00:53:53.679
exceeding so Sorre's vision of like what language is and
881
00:53:53.760 --> 00:53:54.559
how to study it.
882
00:53:55.920 --> 00:53:58.800
One of our first takes that I I think we
883
00:53:58.800 --> 00:54:01.639
were right on. It was probably from a year ago
884
00:54:01.800 --> 00:54:07.280
or two years ago. Even we said, nah, yeah, Chat GBT,
885
00:54:07.400 --> 00:54:09.239
it's going to get better, but it's going to replace
886
00:54:09.280 --> 00:54:14.679
the bullshit writing, and that's the writing. It brings up
887
00:54:14.679 --> 00:54:17.079
the question like why why do we write? Like, sure,
888
00:54:17.159 --> 00:54:19.920
it can write an email for you, or a resume
889
00:54:20.039 --> 00:54:22.760
for you, or a draft for you, and that's great,
890
00:54:22.960 --> 00:54:29.920
Like I welcome that until Chat, until open AI and
891
00:54:29.920 --> 00:54:33.239
and Anthropic go bankrupt because it's it's way too expensive.
892
00:54:33.559 --> 00:54:35.599
But it's nice to have them pick up the bullshit
893
00:54:35.719 --> 00:54:39.119
for you. But it leaves out I don't want to
894
00:54:39.119 --> 00:54:41.559
even say intentionality because that's adding to it, but it
895
00:54:42.239 --> 00:54:44.880
does take away the whatever you want to call the
896
00:54:44.880 --> 00:54:55.519
the epiphany, the gratification, the learning process that I think there's.
897
00:54:53.760 --> 00:54:56.119
A productive labor, and there's a productive labor that can
898
00:54:56.159 --> 00:54:57.519
come from from trying to figure.
899
00:54:57.280 --> 00:54:59.960
Out how I mean practice, not even not even poisse,
900
00:55:00.039 --> 00:55:05.400
it's just practice, like the feeling of creating coherence by
901
00:55:05.599 --> 00:55:08.039
learning to write, and not most writing is not that
902
00:55:08.360 --> 00:55:10.480
like most writing is if you're trying to write an
903
00:55:10.480 --> 00:55:12.719
academic paper or you're trying to write a book to
904
00:55:12.719 --> 00:55:17.119
get published or so much of writing is like poetic
905
00:55:17.480 --> 00:55:19.440
in that you're trying to produce something at the end.
906
00:55:19.960 --> 00:55:27.519
But let Chadgibd cannot for you create the coherence that
907
00:55:27.559 --> 00:55:32.559
it gets from like sitting and working out ideas via language.
908
00:55:32.599 --> 00:55:36.199
That's that's something that's irreproducible by something else because you
909
00:55:36.239 --> 00:55:38.480
have to do it. And I wish, I wish I
910
00:55:38.559 --> 00:55:40.679
was still doing it by hand. I usually do it
911
00:55:40.760 --> 00:55:44.880
right by computer now, But now it's like an excuse
912
00:55:44.920 --> 00:55:48.000
to go back to what what is writing actually for?
913
00:55:48.159 --> 00:55:51.599
Why do we actually write? And it's not to write
914
00:55:51.639 --> 00:55:52.840
bullshit emails? You know?
915
00:55:54.159 --> 00:55:56.559
Yeah, yeah, I remember, I remember us having that take
916
00:55:56.599 --> 00:55:59.199
that it's going to replace like that bullshit like the
917
00:55:59.360 --> 00:56:01.960
that is actually like the real laborious writing is just
918
00:56:02.000 --> 00:56:05.159
writing an email to someone or yeah, like just the
919
00:56:05.440 --> 00:56:11.880
like like kind of menial menial tasks of it all. Yeah,
920
00:56:11.920 --> 00:56:12.920
I forgot what I was going to say. I was
921
00:56:12.920 --> 00:56:14.159
going to sais I forgot?
922
00:56:14.760 --> 00:56:17.039
Well? I think I think.
923
00:56:18.840 --> 00:56:23.199
Why why we write or produce language in any way?
924
00:56:23.239 --> 00:56:26.199
I guess including speech I mean, it's not really a why,
925
00:56:26.360 --> 00:56:31.079
but Jakobsen gives us that answer. The six functions that
926
00:56:31.519 --> 00:56:37.400
language kind of performs, right, which are really good. I
927
00:56:37.880 --> 00:56:40.639
enjoy those six functions. I think they're I think they're
928
00:56:40.679 --> 00:56:41.519
pretty complete.
929
00:56:42.039 --> 00:56:43.199
Do you want to list them real quick?
930
00:56:43.760 --> 00:56:44.039
Is you can?
931
00:56:44.159 --> 00:56:46.320
You can give a kind of example of each one
932
00:56:46.400 --> 00:56:47.039
really easily.
933
00:56:47.320 --> 00:56:47.639
Yeah.
934
00:56:47.760 --> 00:56:52.360
You have the the referential function, which is sort of
935
00:56:52.440 --> 00:56:55.599
talking about things in the world context.
936
00:56:56.199 --> 00:56:56.480
Uh.
937
00:56:56.519 --> 00:57:00.000
You have the emotive function, which is trying to express
938
00:57:00.119 --> 00:57:03.880
how the addresser feels. You have the like I'm hungry whatever.
939
00:57:04.400 --> 00:57:07.599
You have the conative function, which is which is aimed
940
00:57:07.599 --> 00:57:12.400
at the address e. Uh, so you know what what's
941
00:57:12.599 --> 00:57:17.280
what's the example gives of the conative function here, oh,
942
00:57:17.320 --> 00:57:20.639
from like the epic poetry, right when you start with
943
00:57:20.760 --> 00:57:23.880
like oh, and you're using this evocative case and you're
944
00:57:23.880 --> 00:57:28.039
like addressing someone directly listen up you Uh, that's kind
945
00:57:28.039 --> 00:57:31.079
of the aimed at the address ee. You have the
946
00:57:31.480 --> 00:57:36.239
fatic function, which which tests contact, right like when someone
947
00:57:36.280 --> 00:57:37.920
taps a mic and says, hey, is.
948
00:57:37.880 --> 00:57:38.360
This thing on?
949
00:57:38.760 --> 00:57:41.639
Or you're breaking up? Can you hear me? Can you
950
00:57:41.679 --> 00:57:44.159
can you hear me? It's like you're testing the connection
951
00:57:44.480 --> 00:57:49.960
of the of the line nice, there's a well, that's
952
00:57:50.039 --> 00:57:52.559
that's I guess, I guess yeah, you could say that.
953
00:57:52.559 --> 00:57:54.440
That's sort of you're you're just trying to keep the
954
00:57:54.480 --> 00:58:00.280
conversation going, maintain contact. Uh, there's the there's the there's
955
00:58:00.320 --> 00:58:04.159
the meta lingual function, which is which is, uh, you're
956
00:58:04.159 --> 00:58:07.800
talking this, you're talking about the code, which is basically
957
00:58:07.840 --> 00:58:12.480
like you're you're basically trying to determine. I mean, one
958
00:58:12.480 --> 00:58:14.840
good way to think about it is when you're clarifying
959
00:58:14.840 --> 00:58:18.119
the meaning of something, Uh, like what did if you
960
00:58:18.159 --> 00:58:20.599
ask someone like they just said something to you and
961
00:58:20.639 --> 00:58:22.559
you say, like, what did you mean by this?
962
00:58:23.679 --> 00:58:24.920
Or Oh?
963
00:58:24.960 --> 00:58:27.960
The example the author gives is is are pugnacious and
964
00:58:28.000 --> 00:58:32.159
pugilistic synonyms. So you're you're kind of like talking You're
965
00:58:32.159 --> 00:58:35.280
you're talking about the code of language directly. You're not
966
00:58:35.360 --> 00:58:39.639
using language, using metal language, using language to talk about language.
967
00:58:40.440 --> 00:58:42.199
That's this whole podcast.
968
00:58:43.599 --> 00:58:47.400
In a in a sense, yeah, theory theory uh is
969
00:58:47.440 --> 00:58:51.480
often meta lingual. And then like the last one, Okay,
970
00:58:51.519 --> 00:58:56.800
that's six of them. The most important one is the
971
00:58:56.840 --> 00:59:03.920
poetic function here, which is actually where actually where Yakobsen sticks.
972
00:59:04.480 --> 00:59:09.079
So Serti's conception of value corresponds to most strongly to
973
00:59:09.280 --> 00:59:13.480
the poetic function, which is a little weird. The poetic
974
00:59:13.519 --> 00:59:19.039
functions described as the set of messages about the message.
975
00:59:19.960 --> 00:59:21.639
It's about the message itself.
976
00:59:23.000 --> 00:59:24.519
What's the example that it has.
977
00:59:26.760 --> 00:59:32.639
Where I'm not sure. The poetic function is the diametrical opposition.
978
00:59:33.239 --> 00:59:36.960
As Jakobsen says it here he's quoting Jakobsen, poetry uses
979
00:59:36.960 --> 00:59:40.719
an equation to build a sequence. It projects the principle
980
00:59:40.760 --> 00:59:45.199
of equivalence from the axis of selection into the axis
981
00:59:45.199 --> 00:59:45.960
of combination.
982
00:59:47.119 --> 00:59:49.039
Okay, you can do some meta language and tell me
983
00:59:49.039 --> 00:59:51.800
if I'm right about this, because it's been a while
984
00:59:51.880 --> 00:59:56.039
since I've read it. But Jakubsin's poetics, as far as
985
00:59:56.079 --> 01:00:00.960
I understand, it's when the words are message is chosen
986
01:00:01.320 --> 01:00:07.679
for the message's sake rather than the sake of effective
987
01:00:07.719 --> 01:00:13.159
communication or something like that. So each word matters much
988
01:00:13.239 --> 01:00:15.800
more in poetry than in any of the other six
989
01:00:15.840 --> 01:00:18.960
types of language that you mentioned, because each word, you know,
990
01:00:19.000 --> 01:00:23.239
if it's a poem, you can you can infer that
991
01:00:23.440 --> 01:00:27.679
each word is chosen, and in poetry they're chosen for
992
01:00:27.760 --> 01:00:31.400
different reasons. Some are chose like if you're trying to
993
01:00:31.440 --> 01:00:34.480
do iambic pentameter, you have to have like a bunch
994
01:00:34.519 --> 01:00:37.719
of two or three syllable words. If you're choosing for
995
01:00:37.880 --> 01:00:41.599
rhyme and you want internal rhyme and you want end rhyme,
996
01:00:42.159 --> 01:00:45.639
then you are choosing them for their sounds. So as
997
01:00:45.679 --> 01:00:48.239
far as I is that right, that's that's the poetics, right.
998
01:00:48.719 --> 01:00:51.760
Yeah, like meta meta language does seem to be about
999
01:00:52.079 --> 01:00:59.880
making semantic correspondences, whereas poetry about the message means sent
1000
01:01:00.079 --> 01:01:04.239
seems to be. Yeah, like metalanguage is about the code.
1001
01:01:04.880 --> 01:01:10.039
You're using the sequence to establish equivalence, poetry is about
1002
01:01:10.039 --> 01:01:14.159
the message. You're using equivalents to generate a sequence, So
1003
01:01:14.400 --> 01:01:19.760
equivalence can mean here. Here's another example poetry. In poetry,
1004
01:01:19.880 --> 01:01:23.840
one syllable is equalized with any other of the same sequence.
1005
01:01:23.880 --> 01:01:27.599
Word stress is assumed to equal word stress, and unstress
1006
01:01:27.960 --> 01:01:31.920
equals unstress, and so on. So here's the example. You know,
1007
01:01:32.039 --> 01:01:35.280
here's ed grallan pose the raven. He builds this sequence
1008
01:01:35.320 --> 01:01:39.800
based on the equivalence's aka the end rhyme. Okay, when
1009
01:01:39.800 --> 01:01:43.679
I nodded nearly napping, suddenly there came a tapping as
1010
01:01:43.719 --> 01:01:48.719
of someone gently wrapping, right, So you have those not
1011
01:01:48.880 --> 01:01:53.320
those aren't semantic equivalences, So.
1012
01:01:53.199 --> 01:01:57.159
You're using the words as themselves instead of the words
1013
01:01:57.920 --> 01:01:59.280
as a reference to other.
1014
01:01:59.719 --> 01:02:04.000
St Rhythm and sound become the focus on which you're
1015
01:02:04.039 --> 01:02:10.559
building the sequence, so it's more like the physicality the
1016
01:02:10.599 --> 01:02:14.360
materiality of the language, rather than the semantic aspect of it.
1017
01:02:15.039 --> 01:02:16.480
You could say, maybe.
1018
01:02:16.559 --> 01:02:21.960
Yeah, why then, is why are lms the most poetic?
1019
01:02:22.920 --> 01:02:28.559
Because I think another way to think about it is contiguity,
1020
01:02:29.000 --> 01:02:34.039
which is like the physical or like the physical correspondences
1021
01:02:34.079 --> 01:02:39.400
between words, closeness, contiguity, Like AI is really good at
1022
01:02:39.440 --> 01:02:44.039
mapping that out, And that's the part that the poetic
1023
01:02:44.079 --> 01:02:48.599
function takes over as well, because when it says that
1024
01:02:49.159 --> 01:02:52.360
for poetry, for example, when you're using the poetic function,
1025
01:02:52.960 --> 01:02:57.639
projects the principle of equivalence from the axis of selection
1026
01:02:57.960 --> 01:03:01.360
into the axis of combination. Now, the axis of combination
1027
01:03:01.519 --> 01:03:05.719
is normally operating on the principle of contiguity, but contiguity
1028
01:03:05.840 --> 01:03:11.880
is then displaced by equivalents, so equivalence becomes sort of
1029
01:03:11.920 --> 01:03:16.480
the generative principle of combination. And so I guess AI
1030
01:03:16.880 --> 01:03:23.079
is really really good at mapping out those sort of
1031
01:03:23.840 --> 01:03:30.599
material correspondences contiguities between yeah, and maybe I'm not saying it.
1032
01:03:30.639 --> 01:03:32.039
Maybe I'm not saying it right.
1033
01:03:32.679 --> 01:03:35.679
What I got out of it. He doesn't say that
1034
01:03:35.800 --> 01:03:38.840
it's good at doing this. He says, this is literally
1035
01:03:38.880 --> 01:03:43.480
what an LM is doing is poetics. Yeah, And what
1036
01:03:43.519 --> 01:03:46.760
I think, and what I take that to mean is
1037
01:03:47.480 --> 01:03:51.679
that when all the other forms of language that you mentioned,
1038
01:03:51.679 --> 01:03:53.719
like one is about saying what your feelings are, one
1039
01:03:53.760 --> 01:03:55.639
is about making sure you're able to talk to somebody,
1040
01:03:55.679 --> 01:03:58.559
one is about communicating, Hey, look over there, there's a
1041
01:03:58.639 --> 01:04:02.360
stop sign. All those things are about the world, whereas
1042
01:04:02.840 --> 01:04:06.440
the poetic function of language, in Jakobsen sense is just
1043
01:04:06.559 --> 01:04:11.719
about the words themselves, and the reference to the world
1044
01:04:11.880 --> 01:04:14.599
is secondary. It's not the primary purpose. You're not trying
1045
01:04:14.599 --> 01:04:17.039
to communicate about the world. You're trying to communicate about
1046
01:04:17.119 --> 01:04:22.639
language itself. So that's I see what the LLM is doing.
1047
01:04:22.679 --> 01:04:25.079
Because the LLM doesn't know anything about the world. It
1048
01:04:25.119 --> 01:04:27.679
doesn't know a single thing about any word that it
1049
01:04:27.760 --> 01:04:31.559
is ever uttered, except that it's a word, if you
1050
01:04:31.559 --> 01:04:35.800
want to call that knowing. So there's nothing that's referential.
1051
01:04:35.920 --> 01:04:41.000
All it knows is words, word combinations, statistical probability, how
1052
01:04:41.039 --> 01:04:44.679
words connect, So it's doing the poetic function because it
1053
01:04:44.800 --> 01:04:49.519
is nothing. There's nothing else. There's no secondary layer, there's
1054
01:04:49.519 --> 01:04:54.480
no layer of reference with a T, there's just the
1055
01:04:54.519 --> 01:04:56.800
words themselves. So that's what I took him to me
1056
01:04:57.360 --> 01:05:01.239
when he's saying this is just poetical speech. It's like
1057
01:05:01.719 --> 01:05:05.480
it's words chosen for the internal properties of the words,
1058
01:05:05.880 --> 01:05:08.400
not for the things that they represent.
1059
01:05:08.719 --> 01:05:13.639
Yeah, like the quality of the words themselves. Yeah, not
1060
01:05:13.639 --> 01:05:16.880
not what they represent, not like the content or the meaning,
1061
01:05:18.079 --> 01:05:22.239
and not the sort of rules or the code by
1062
01:05:22.320 --> 01:05:28.199
which they're supposed to be interpreted. Yeah, that's sort of
1063
01:05:28.239 --> 01:05:32.519
what the message means too. The message means like language.
1064
01:05:32.559 --> 01:05:37.039
It's as such, And I guess I don't know. Yeah,
1065
01:05:36.760 --> 01:05:40.039
lms do a different version of this. They're not just
1066
01:05:40.079 --> 01:05:45.440
stuck making poetry, but they do, you they do enact
1067
01:05:45.519 --> 01:05:46.840
the poetic function.
1068
01:05:49.119 --> 01:05:51.480
Through statistical modeling as opposed to vibes.
1069
01:05:52.400 --> 01:05:55.360
Yeah. Yeah, it's a very it's like a it's like
1070
01:05:55.400 --> 01:05:58.559
a banal type of poetry where they start sounded like
1071
01:05:58.960 --> 01:06:00.000
fucking LinkedIn posts.
1072
01:06:01.199 --> 01:06:03.119
But maybe, but but what it points to is maybe
1073
01:06:03.199 --> 01:06:08.800
underlying all all great air quotes art is some kind
1074
01:06:08.800 --> 01:06:11.360
of pattern that maybe an advanced enough LM could actually
1075
01:06:11.400 --> 01:06:14.000
figure out what that pattern is that makes it good,
1076
01:06:14.199 --> 01:06:16.519
like through a statistical model, and give some kind of
1077
01:06:16.559 --> 01:06:19.880
facsimile of it that is satisfactory. I mean, there were
1078
01:06:19.880 --> 01:06:22.000
all these short stories that are apparently one hundred percent
1079
01:06:22.039 --> 01:06:24.559
AI that have been winning awards recently. We shouldn't read
1080
01:06:24.599 --> 01:06:27.480
one of those on the podcast. It'd be funny, but
1081
01:06:27.880 --> 01:06:29.719
we're not doing anymore AI. But I think it would
1082
01:06:29.760 --> 01:06:32.039
be interesting to I never know what I said.
1083
01:06:32.159 --> 01:06:36.000
I said. Increasingly several people have said we should stop talking.
1084
01:06:36.039 --> 01:06:38.079
Okay, I see, I see, but I don't know.
1085
01:06:38.199 --> 01:06:40.000
I think I I think maybe we brought them back
1086
01:06:40.039 --> 01:06:40.480
with this one.
1087
01:06:40.519 --> 01:06:43.079
But if you two, if especially if you pills could abstain,
1088
01:06:43.119 --> 01:06:44.719
because I think there was a recent one in the news.
1089
01:06:44.760 --> 01:06:47.639
From looking into which one won an award, maybe I
1090
01:06:47.679 --> 01:06:50.840
could give us two short stories and you have to
1091
01:06:50.840 --> 01:06:52.840
figure out which one was written by the LM and
1092
01:06:52.880 --> 01:06:55.039
won the award and which one is authentic.
1093
01:06:56.760 --> 01:07:00.199
I I find it likely that i'd be due. I
1094
01:07:00.239 --> 01:07:05.320
don't have my poetic reading skills are not the snuff
1095
01:07:05.320 --> 01:07:10.480
I did say. It's not easy. Yeah, so they are
1096
01:07:10.559 --> 01:07:14.679
poetic poetic function of language. Don't confuse it with poetry,
1097
01:07:14.960 --> 01:07:19.400
but they use words for words sake, and we do
1098
01:07:19.480 --> 01:07:21.199
that sometimes too. Well.
1099
01:07:21.239 --> 01:07:26.480
The this is this line hits the hardest done in
1100
01:07:26.519 --> 01:07:31.360
this argument is he says language as such takes structure
1101
01:07:31.639 --> 01:07:37.960
from poetic function. So that so that's it. I remember
1102
01:07:38.000 --> 01:07:44.000
when I said, basically Jakobsen locates, so Sirr's idea of
1103
01:07:44.039 --> 01:07:48.159
like linguistic value in the poetic function. So now we
1104
01:07:48.239 --> 01:07:50.599
get to the point where we're saying language takes its
1105
01:07:50.639 --> 01:07:52.400
structure from the poetic function.
1106
01:07:53.079 --> 01:07:56.400
If you ask someone, the common sense answer is that
1107
01:07:57.400 --> 01:08:00.880
the the referential function of language, the first function of
1108
01:08:00.960 --> 01:08:04.840
language is like in the most indicative of the whole.
1109
01:08:04.920 --> 01:08:08.079
The thing we use language for is to point to
1110
01:08:08.119 --> 01:08:10.159
out things in the world to other people so that
1111
01:08:10.199 --> 01:08:13.280
they can see the same thing we're saying. But this argument,
1112
01:08:13.639 --> 01:08:17.279
this argument is that the poetic function, which is language,
1113
01:08:17.359 --> 01:08:22.760
that is language and nothing else that is catchy. It's
1114
01:08:23.199 --> 01:08:25.600
it's catchy, it's sonorous.
1115
01:08:25.840 --> 01:08:29.079
That's its most important thing. Language has to be catchy.
1116
01:08:29.159 --> 01:08:31.640
That's the most important thing, first and foremost.
1117
01:08:31.920 --> 01:08:34.079
That's what seduces us, that's what that's what.
1118
01:08:34.079 --> 01:08:37.279
Draws us into anything. We wouldn't use language if it
1119
01:08:37.319 --> 01:08:38.680
weren't catchy and seductive?
1120
01:08:39.720 --> 01:08:42.279
I mean, interestingly, this is what Lacan says about the
1121
01:08:42.359 --> 01:08:46.920
unconscious as well, that like in dreams, you need to
1122
01:08:46.920 --> 01:08:51.760
look at like the connotations of words, because one thing
1123
01:08:51.840 --> 01:08:55.439
might be substituted for another word that that that sounds
1124
01:08:55.520 --> 01:08:58.000
like it, or that has a similar meaning in a
1125
01:08:58.000 --> 01:09:01.760
different context, that kind of stuff. Yeah, I guess sexual pleasure.
1126
01:09:02.279 --> 01:09:07.479
Sexual pleasure is to reproduction as as as catchiness and
1127
01:09:08.359 --> 01:09:11.239
funness is to communication.
1128
01:09:11.720 --> 01:09:15.560
Well, I guess we still haven't answered the question of
1129
01:09:15.680 --> 01:09:20.359
why why can I not for the opposite of being disappointed?
1130
01:09:20.439 --> 01:09:23.239
Why can't I say I'm very appointed today about this?
1131
01:09:26.079 --> 01:09:28.680
Well? Is that a false what's that? There's a word for, Hey,
1132
01:09:28.760 --> 01:09:33.960
metal language question, there's a word for like false etymology.
1133
01:09:35.319 --> 01:09:36.600
Oh, I don't know what that word is, but yeah,
1134
01:09:36.760 --> 01:09:38.600
I think I think I know what you're talking about.
1135
01:09:39.079 --> 01:09:40.960
Is that one a false homology or is it just
1136
01:09:41.000 --> 01:09:42.600
a drift? The genetic drift?
1137
01:09:43.119 --> 01:09:48.680
Disappointed? I've been appointed or someone fired? You've been disappointed
1138
01:09:48.800 --> 01:09:49.640
from this position?
1139
01:09:51.479 --> 01:09:56.560
False cognate? More more metal language false cognate? Was that
1140
01:09:56.640 --> 01:09:58.680
what I was thinking of? Okay, but now I want
1141
01:09:58.680 --> 01:10:00.399
to look up disappointed to see if it is in
1142
01:10:00.479 --> 01:10:05.399
fact a false cognate or not. Okay, So here's what
1143
01:10:05.439 --> 01:10:08.600
I mean. Here's what I mean. It's victor from Middle French.
1144
01:10:09.800 --> 01:10:15.000
That's appointe. The word originally meant to dispossess of an
1145
01:10:15.000 --> 01:10:15.920
appointed office.
1146
01:10:16.319 --> 01:10:19.399
Ah, but isn't it to lose, to lose a position? Oh?
1147
01:10:19.399 --> 01:10:21.439
I get Yeah, you take away, you take away. There,
1148
01:10:21.600 --> 01:10:23.159
you take away a position.
1149
01:10:22.920 --> 01:10:25.439
You take away. So we still say you've been appointed
1150
01:10:25.479 --> 01:10:29.880
and especially yeah, but I feel like appointed is never used.
1151
01:10:29.960 --> 01:10:32.840
It's only used at like high levels. Like I feel
1152
01:10:32.880 --> 01:10:35.079
like whenever I see announcements from a university, it'll be
1153
01:10:35.159 --> 01:10:38.279
like professor so and so has been appointed the dean
1154
01:10:39.239 --> 01:10:41.920
of arts and science or whatever. Right, it'll be like
1155
01:10:41.920 --> 01:10:45.800
they've been appointed. Uh, It's like it's like when you're yeah,
1156
01:10:45.800 --> 01:10:47.720
I don't know, I feel because you would never say,
1157
01:10:47.760 --> 01:10:49.800
for example, oh, I've been appointed to be a sales
1158
01:10:49.840 --> 01:10:51.560
associate at Rogers Wireless.
1159
01:10:51.640 --> 01:10:53.840
Right, you'd never like, although you would say I have
1160
01:10:53.920 --> 01:10:55.640
a dentist appointment.
1161
01:10:56.239 --> 01:10:59.680
That's true, that's true. So I was so the relationship
1162
01:10:59.680 --> 01:11:02.920
between but then how did disappoint it become its different meaning?
1163
01:11:02.920 --> 01:11:05.399
I don't know where you're about to explain that just slipped.
1164
01:11:06.119 --> 01:11:10.199
Uh, well, it's Middle French, so we're talking disappoint was
1165
01:11:10.560 --> 01:11:14.079
is a several hundred year old word, so I guess,
1166
01:11:14.800 --> 01:11:18.960
oh he got. It just means you got rejected.
1167
01:11:18.479 --> 01:11:19.600
I guess.
1168
01:11:19.920 --> 01:11:21.560
And then we took it to make it to become
1169
01:11:21.600 --> 01:11:24.479
like an internal feeling instead of an expense circumstance.
1170
01:11:24.560 --> 01:11:25.880
I see, that makes sense. So it would be the
1171
01:11:25.880 --> 01:11:28.279
feeling from being from from losing a position.
1172
01:11:29.079 --> 01:11:32.119
Yeah, it did. It did use to It did used
1173
01:11:32.159 --> 01:11:34.800
to mean that. It used to mean, yes, strip someone
1174
01:11:34.840 --> 01:11:37.880
of their position, like you appoint them to a position
1175
01:11:38.199 --> 01:11:40.479
and then you But then that.
1176
01:11:40.479 --> 01:11:42.960
Makes so it used to be, oh, why is he
1177
01:11:43.119 --> 01:11:46.680
sad because he was disappointed exactly, and then you just
1178
01:11:46.720 --> 01:11:48.359
got rid of the sad, So now you just are
1179
01:11:48.520 --> 01:11:50.760
disappointed exactly having been it.
1180
01:11:51.359 --> 01:11:55.079
So yeah, it got like this psychological or effective meaning
1181
01:11:55.159 --> 01:11:59.079
because of semantic drift executive focus on.
1182
01:11:59.000 --> 01:12:02.199
The because it was the feeling that people had when
1183
01:12:02.239 --> 01:12:04.800
they were literally disappointed from a position, and they had
1184
01:12:04.800 --> 01:12:08.640
this feeling, and that feeling just became applied to other situations.
1185
01:12:08.680 --> 01:12:12.239
You use the same word to refer to other situations.
1186
01:12:12.039 --> 01:12:15.439
But now you can say like they were unappointed.
1187
01:12:17.079 --> 01:12:19.680
You just say they were fired. But yeah, you can
1188
01:12:19.720 --> 01:12:21.880
say unappointed. You could be like, yeah, they lost.
1189
01:12:21.840 --> 01:12:24.640
People, wouldn't you know what it means. But it's not
1190
01:12:24.760 --> 01:12:26.479
the most common way to say that.
1191
01:12:26.520 --> 01:12:28.800
But then it's funny. It's interesting that that drift didn't
1192
01:12:28.800 --> 01:12:33.159
happen with appointed, where you're just like, oh, I'm stoked today,
1193
01:12:33.199 --> 01:12:34.159
I'm very appointed.
1194
01:12:34.880 --> 01:12:35.840
We can make it happen.
1195
01:12:36.119 --> 01:12:37.000
We can make it happen.
1196
01:12:38.039 --> 01:12:38.399
There are.
1197
01:12:40.239 --> 01:12:41.439
Very appointed right now.
1198
01:12:42.319 --> 01:12:46.279
I think this is uh. These are words that are
1199
01:12:46.319 --> 01:12:51.319
like orphans negatives linguistics call the linguists call them where
1200
01:12:51.359 --> 01:12:55.439
they're negative words that don't have like an obvious You
1201
01:12:55.479 --> 01:12:58.399
can't just chop off like the prefix and make it
1202
01:12:58.439 --> 01:12:59.479
into the opposite.
1203
01:13:00.159 --> 01:13:02.079
Yeah, unless we unless we worked really hard.
1204
01:13:02.399 --> 01:13:06.720
Like disembowel, you can't embowel somebody, yeah exactly.
1205
01:13:06.600 --> 01:13:10.159
Or when somebody, uh yeah, maybe you could embowel them.
1206
01:13:10.199 --> 01:13:12.960
You stick their guts, stick their guts back inside.
1207
01:13:13.359 --> 01:13:16.159
Maybe surgeons in bowel all the time. Maybe they disembowel
1208
01:13:16.159 --> 01:13:18.560
and then embowel. Maybe they like you know, they take
1209
01:13:18.600 --> 01:13:21.840
the the they take the the intestines out and then
1210
01:13:21.840 --> 01:13:23.920
put them back in and they're emboweling them again.
1211
01:13:24.520 --> 01:13:27.439
You can you can call someone nonchalant when they like
1212
01:13:27.640 --> 01:13:29.720
don't give a fuck. But you can't really call them
1213
01:13:29.800 --> 01:13:32.199
chalant when they when they're worried about stuff.
1214
01:13:33.600 --> 01:13:38.119
Yeah, I've started putting ingenuous back into my vocabulary. You
1215
01:13:38.399 --> 01:13:42.880
don't never hear ingenuous, just disingenuous. But oh right, ingenuous
1216
01:13:42.920 --> 01:13:43.640
can be used.
1217
01:13:44.840 --> 01:13:45.319
I like it.
1218
01:13:45.640 --> 01:13:48.880
Okay, here's interesting. M bowel. I looked it up and
1219
01:13:49.079 --> 01:13:54.439
bowel means to remove the bowels and also disembowel means
1220
01:13:54.560 --> 01:13:58.720
to remove the bowels. Oh oh that's choice. So it's
1221
01:13:58.760 --> 01:13:59.920
like inflammable and flat.
1222
01:14:01.000 --> 01:14:02.239
Oh yeah, wait.
1223
01:14:04.079 --> 01:14:10.000
The same don't inflammable in the Simpsons? Oh yeah, I can't.
1224
01:14:10.039 --> 01:14:14.760
They both mean it can catch on fire. That's the
1225
01:14:15.560 --> 01:14:20.680
very dangerous shifts there. I can't believe you guys know
1226
01:14:20.840 --> 01:14:22.640
Star Trek and not the Simpson.
1227
01:14:22.520 --> 01:14:24.239
Well, I know that. I used to watch the Simpsons.
1228
01:14:24.239 --> 01:14:26.520
I just can't. I don't have the references.
1229
01:14:26.000 --> 01:14:31.760
As doctor it's been too long. Oh doctor, Nick, don't worry.
1230
01:14:31.920 --> 01:14:37.159
It's inflammable. Okay.
1231
01:14:38.159 --> 01:14:41.119
Yeah, and A and A l M could pick up
1232
01:14:41.159 --> 01:14:44.439
all this perfectly well, because you could just ask it
1233
01:14:44.960 --> 01:14:49.079
why why those mean that, and they'll give you some
1234
01:14:49.159 --> 01:14:51.560
convincing answer. I don't know if it's right. You better
1235
01:14:51.640 --> 01:14:56.199
consultant etymological dictionary, just to be sure, but probably knows well.
1236
01:14:56.239 --> 01:15:01.720
I'm very satisfied with my new appointment into disappointment knowledge.
1237
01:15:02.439 --> 01:15:03.920
So you're appointed today.
1238
01:15:04.039 --> 01:15:05.359
I feel very appointed by it.
1239
01:15:05.920 --> 01:15:09.560
I was appointed by this episode. I was appointed by
1240
01:15:09.560 --> 01:15:14.000
this book. I bet we could yo, mister Weatherby, come
1241
01:15:14.079 --> 01:15:15.279
come talk to us about it.
1242
01:15:15.800 --> 01:15:16.000
Yeah.
1243
01:15:16.039 --> 01:15:20.159
But now now you're using appointed as the opposite of
1244
01:15:20.199 --> 01:15:22.920
the semantically drifted disappointment.
1245
01:15:23.359 --> 01:15:25.600
We're trying to semantically We just said we're gonna we're
1246
01:15:25.600 --> 01:15:28.800
going to change la long back. We're trying to semantically
1247
01:15:28.880 --> 01:15:31.279
drift appointments appointed.
1248
01:15:31.680 --> 01:15:32.680
But then appointed.
1249
01:15:32.720 --> 01:15:36.560
Would you have to put disappointed back into its original meaning.
1250
01:15:36.920 --> 01:15:40.039
It's just to strip someone of their office.
1251
01:15:40.319 --> 01:15:43.479
We don't have to do anything. We're bringing it back.
1252
01:15:43.720 --> 01:15:47.119
Eric, you can there's no there's no rules.
1253
01:15:47.600 --> 01:15:48.880
You can be on board or not.
1254
01:15:49.399 --> 01:15:51.960
There's just patterns. There's no rules here. Like we we
1255
01:15:51.960 --> 01:15:53.560
we invent the rules and we just have to get
1256
01:15:53.560 --> 01:15:56.920
people to buy into it. And then that's it. That's
1257
01:15:57.039 --> 01:15:57.840
that's the reality.
1258
01:15:57.880 --> 01:16:01.760
Now you write it in a Reddit post, then CHATGBT
1259
01:16:01.920 --> 01:16:06.319
uses it as training data. Done, I making it a rule.
1260
01:16:06.479 --> 01:16:12.880
I'm talking meta language, all right, Take care Guys take.
1261
01:16:12.720 --> 01:16:13.159
Care of one.
1262
01:16:13.279 --> 01:16:17.000
Goodbye,