The AI Stack After LLMs: Local Inference, World Models, and Logic | Dr. Matthew Molineaux

1
00:00:00,000 --> 00:00:03,940
All right, it's another session of our new podcast.
2
00:00:04,360 --> 00:00:07,060
We are so excited to have Matt Molyneux here.
3
00:00:07,720 --> 00:00:09,320
It's a big pleasure for me.
4
00:00:09,320 --> 00:00:13,060
So Matt is co-founder of Delegance.
5
00:00:13,760 --> 00:00:16,960
And I first met Matt back in 2012.
6
00:00:17,760 --> 00:00:19,980
We were both doing autonomy research.
7
00:00:20,300 --> 00:00:22,880
And I actually had found one of his papers first.
8
00:00:23,760 --> 00:00:26,380
He did some awesome work called Discover History,
9
00:00:26,380 --> 00:00:30,460
which is like causal explanations for things in,
10
00:00:31,040 --> 00:00:34,100
when figuring out diagnosis for when autonomous systems fail.
11
00:00:34,740 --> 00:00:38,480
And this is like, you know, early 2010s, I think,
12
00:00:38,560 --> 00:00:40,400
or that this work was published.
13
00:00:41,000 --> 00:00:43,580
I first published a Discover History paper in 2010.
14
00:00:44,620 --> 00:00:49,000
Yeah, and so I read that as I was getting ready to go to grad school.
15
00:00:49,280 --> 00:00:53,640
And then we met and yeah, now we're, you know,
16
00:00:54,060 --> 00:00:55,480
15 years later starting a company.
17
00:00:55,480 --> 00:00:56,580
It's so exciting.
18
00:00:57,680 --> 00:00:57,960
Amazing.
19
00:00:59,000 --> 00:01:04,900
So, yeah, I mean, you've been in the AI field, you know, before it was cool, I guess.
20
00:01:06,080 --> 00:01:11,780
Yeah, so I spent a bit more than 20 years mostly defense contracting.
21
00:01:12,780 --> 00:01:21,200
I started back in 2003, and I got my PhD close to the time when Dustin did, actually, in 2017,
22
00:01:21,200 --> 00:01:24,780
because I was just working in the field and doing my PhD alongside for years.
23
00:01:25,480 --> 00:01:29,120
But a lot of the things that I have done
24
00:01:29,120 --> 00:01:30,500
have been about autonomous agents.
25
00:01:30,800 --> 00:01:32,480
Agent to KI isn't really new.
26
00:01:32,880 --> 00:01:35,100
It's just kind of being rediscovered, I think.
27
00:01:36,360 --> 00:01:38,700
And autonomous agents that act in environments
28
00:01:38,700 --> 00:01:40,780
like video games at first was a thing that we did.
29
00:01:40,920 --> 00:01:42,940
Warcraft 2 was one of the early domains
30
00:01:42,940 --> 00:01:46,200
that I worked on agents in.
31
00:01:47,520 --> 00:01:51,380
And I've made agents that play chess,
32
00:01:51,380 --> 00:01:54,460
that play Unreal Tournament,
33
00:01:55,580 --> 00:01:59,340
just a whole different range of things.
34
00:01:59,520 --> 00:02:02,120
And having environments where you can act
35
00:02:02,120 --> 00:02:05,640
and get better at things
36
00:02:05,640 --> 00:02:08,720
has always been really interesting to me
37
00:02:08,720 --> 00:02:11,000
to find things that can interact with us
38
00:02:11,000 --> 00:02:16,300
on a really clear level
39
00:02:16,300 --> 00:02:19,120
where they understand the consequences of their action.
40
00:02:19,420 --> 00:02:20,820
And that's really blown up a lot.
41
00:02:21,380 --> 00:02:23,260
with conversational AI.
42
00:02:23,460 --> 00:02:26,520
There are a lot more domains where an agent can understand
43
00:02:26,520 --> 00:02:29,080
the consequences of its actions because they can understand.
44
00:02:30,380 --> 00:02:33,980
And when you were working on these agents,
45
00:02:34,140 --> 00:02:37,120
were they LLM-based or this was like pre-LLM?
46
00:02:37,560 --> 00:02:38,760
Yeah, very pre-LLM.
47
00:02:39,140 --> 00:02:43,440
So LLMs, I mean, I first learned about deep learning
48
00:02:43,440 --> 00:02:44,900
in I think 2016.
49
00:02:46,660 --> 00:02:47,900
I might be about right.
50
00:02:48,160 --> 00:02:51,200
And that was sort of a precursor technology.
51
00:02:51,380 --> 00:02:55,080
to what we think of as LLMs now or foundation models.
52
00:02:57,020 --> 00:03:00,020
And, you know, ChatGPT,
53
00:03:00,120 --> 00:03:01,820
does anybody know when ChatGPT came out?
54
00:03:01,880 --> 00:03:03,660
It feels like a COVID thing.
55
00:03:04,320 --> 00:03:06,140
Yeah, you can't really tell, right?
56
00:03:06,200 --> 00:03:09,520
It's like the before times before ChatGPT.
57
00:03:10,900 --> 00:03:12,180
When was it?
58
00:03:12,760 --> 00:03:15,140
I think end of 2022.
59
00:03:15,600 --> 00:03:15,860
Okay.
60
00:03:16,480 --> 00:03:18,460
When it hit the first public release?
61
00:03:18,820 --> 00:03:20,180
I think that was the first open to the public.
62
00:03:20,180 --> 00:03:20,660
Yeah.
63
00:03:21,380 --> 00:03:26,560
pretty big moment what did that feel like for you i mean you've been in the industry for so long and
64
00:03:26,560 --> 00:03:32,920
then suddenly the public is obsessed with ai seemingly overnight i gotta tell you first i
65
00:03:32,920 --> 00:03:42,100
hated it uh so common reaction well to me so i've worked for years on explainability stuff
66
00:03:42,100 --> 00:03:49,160
and sort of neural networks which is the technology which a lot of um all llms are based on at some
67
00:03:49,160 --> 00:03:53,400
level is fundamentally a technology that is black box.
68
00:03:53,400 --> 00:03:54,400
Mm-hmm.
69
00:03:54,400 --> 00:04:00,020
And what worries me about that kind of technology is that it's not as
70
00:04:00,020 --> 00:04:04,520
governable as a word that we use a lot these days.
71
00:04:04,520 --> 00:04:13,960
But it's harder for it to be transparent or, and I feel like we don't,
72
00:04:13,960 --> 00:04:17,520
sorry, I'm not talking into the microphone as much.
73
00:04:19,160 --> 00:04:25,580
I have worked on technologies that are much more understandable and that a person can really
74
00:04:25,580 --> 00:04:29,240
explain their goals to and know that those goals are being followed.
75
00:04:29,240 --> 00:04:42,200
And I really wanted that to be where the first, you know, really general and interesting AI
76
00:04:42,200 --> 00:04:46,160
that everybody, you know, wanted to use for everything came from.
77
00:04:46,160 --> 00:04:50,140
And I was really disappointed that neural networks got there first.
78
00:04:51,400 --> 00:04:56,500
And, you know, over time, I have come to find lots of different ways
79
00:04:56,500 --> 00:04:58,200
where we can make these things work together.
80
00:04:58,720 --> 00:05:03,360
And sort of, I think we've come to understand better
81
00:05:03,360 --> 00:05:07,540
what the strengths and weaknesses of these different technologies are
82
00:05:07,540 --> 00:05:12,740
and, you know, why maybe these things are good for language
83
00:05:12,740 --> 00:05:18,100
and good for, I guess, very unstructured data.
84
00:05:18,920 --> 00:05:21,420
Whereas I still want to use a lot of stuff
85
00:05:21,420 --> 00:05:22,880
that works with structured data,
86
00:05:22,980 --> 00:05:28,080
but it needs to be at a higher, more abstract level.
87
00:05:28,360 --> 00:05:30,880
And abstraction is something that Dustin and I
88
00:05:30,880 --> 00:05:32,240
have worked on a lot in AI.
89
00:05:32,840 --> 00:05:33,280
Nice.
90
00:05:34,020 --> 00:05:35,780
One of the things that we've talked about
91
00:05:35,780 --> 00:05:37,480
in the past couple episodes
92
00:05:37,480 --> 00:05:41,100
is it's not necessarily either or,
93
00:05:41,100 --> 00:05:44,440
or we've talked about world models and LLMs
94
00:05:44,440 --> 00:05:47,300
or deterministic and non-deterministic
95
00:05:47,300 --> 00:05:50,740
and that rather than just choosing
96
00:05:50,740 --> 00:05:54,780
one type of AI and going all in on that,
97
00:05:55,100 --> 00:05:56,860
that there's maybe synergies
98
00:05:56,860 --> 00:05:59,680
where the LLM is particularly suited for something
99
00:05:59,680 --> 00:06:02,180
and then other types of AI is suited for something else
100
00:06:02,180 --> 00:06:03,840
and then they can kind of work together
101
00:06:03,840 --> 00:06:05,580
to get the best of both worlds.
102
00:06:05,760 --> 00:06:08,860
Is that something you think makes sense
103
00:06:08,860 --> 00:06:10,260
or do you have a view there?
104
00:06:10,260 --> 00:06:32,380
Absolutely. I really believe it. A world model fundamentally is a structured representation that is made through interpretation. And interpretation is something that LLMs are much, much better at in terms of interpretation of structured data.
105
00:06:32,380 --> 00:06:34,840
And the human brain does a lot of this stuff.
106
00:06:34,920 --> 00:06:37,420
Our senses take unstructured data
107
00:06:37,420 --> 00:06:39,560
and make structured representations inside.
108
00:06:39,680 --> 00:06:42,660
At least that's what we kind of think is happening, right?
109
00:06:43,240 --> 00:06:45,660
And then there are parts of our minds,
110
00:06:45,780 --> 00:06:47,300
the conscious parts in particular,
111
00:06:47,440 --> 00:06:49,320
that I think are really good at manipulating
112
00:06:49,320 --> 00:06:50,720
these more structured things
113
00:06:50,720 --> 00:06:54,880
and making creative use,
114
00:06:56,160 --> 00:06:59,720
manipulating them in ways that are less,
115
00:06:59,720 --> 00:07:02,780
I want to say
116
00:07:02,780 --> 00:07:03,480
so
117
00:07:03,480 --> 00:07:06,840
the word I want to use is inductive
118
00:07:06,840 --> 00:07:08,540
and I don't know whether everybody's going to understand
119
00:07:08,540 --> 00:07:09,620
what I mean by that
120
00:07:09,620 --> 00:07:12,240
so when we have things that
121
00:07:12,240 --> 00:07:14,260
learn through
122
00:07:14,260 --> 00:07:16,480
a lot of repetition, a lot of examples
123
00:07:16,480 --> 00:07:17,940
then
124
00:07:17,940 --> 00:07:20,260
they sort of figure out what the median
125
00:07:20,260 --> 00:07:21,500
of something is
126
00:07:21,500 --> 00:07:24,240
and you see when you use
127
00:07:24,240 --> 00:07:25,920
LLMs that they're going to
128
00:07:25,920 --> 00:07:28,040
tend toward what lots of humans
129
00:07:28,040 --> 00:07:28,980
are doing
130
00:07:28,980 --> 00:07:31,100
and
131
00:07:31,100 --> 00:07:33,640
this is good
132
00:07:33,640 --> 00:07:36,140
in you know being able to
133
00:07:36,140 --> 00:07:38,380
sort out these unstructured representations
134
00:07:38,380 --> 00:07:40,240
but when you're working with a structured representation
135
00:07:40,240 --> 00:07:41,380
like a world model
136
00:07:41,380 --> 00:07:44,220
then you can extrapolate
137
00:07:44,220 --> 00:07:46,420
better you can go outside the bounds
138
00:07:46,420 --> 00:07:48,280
of you know what is in
139
00:07:48,280 --> 00:07:49,780
what is most common
140
00:07:49,780 --> 00:07:51,680
and you can start to
141
00:07:51,680 --> 00:07:54,340
make predictions what would happen if we did
142
00:07:54,340 --> 00:07:56,420
something fundamentally different and that's what science
143
00:07:56,420 --> 00:07:58,240
is about to a large extent
144
00:07:58,240 --> 00:08:00,360
is pushing the boundaries,
145
00:08:01,060 --> 00:08:03,900
figuring out, well, what haven't we tried yet?
146
00:08:04,880 --> 00:08:08,140
And technology, invention,
147
00:08:08,460 --> 00:08:09,500
all of this stuff, I think,
148
00:08:09,500 --> 00:08:13,500
is sort of based on this ability to think,
149
00:08:14,440 --> 00:08:16,720
I mean, I guess outside the box is one way to put it,
150
00:08:16,820 --> 00:08:21,840
but to go beyond the boundaries of what is common.
151
00:08:22,620 --> 00:08:26,100
And I think that the world models,
152
00:08:26,100 --> 00:08:29,560
the kind of goal reasoning that Dustin and I have worked on,
153
00:08:30,160 --> 00:08:34,140
they're good for doing that in a way that LLMs aren't necessarily.
154
00:08:35,480 --> 00:08:36,440
Super interesting.
155
00:08:36,440 --> 00:08:44,760
I mean, the way I feel about it is we didn't really have very general AI components before,
156
00:08:44,900 --> 00:08:45,760
in a sense.
157
00:08:47,700 --> 00:08:49,580
We've been working on this for a long time.
158
00:08:49,960 --> 00:08:52,500
We've made really powerful systems,
159
00:08:52,900 --> 00:08:55,540
but in some ways they were always a little brittle.
160
00:08:55,540 --> 00:09:09,940
If there was a little mistake in a certain place or it wasn't typed in a certain way or you couldn't even have free form entry where you could kind of talk to it or have a user ask a question.
161
00:09:10,140 --> 00:09:14,020
You might even have structured parts of questions you could ask.
162
00:09:14,680 --> 00:09:20,980
Or if it did look like natural language behind the scenes, it was really like a bunch of hacks to get it to be structured.
163
00:09:20,980 --> 00:09:23,900
so now we have this very general
164
00:09:23,900 --> 00:09:26,000
you know
165
00:09:26,000 --> 00:09:27,920
it's much more flexible in general
166
00:09:27,920 --> 00:09:29,860
because it's kind of learned
167
00:09:29,860 --> 00:09:31,440
these general patterns but
168
00:09:31,440 --> 00:09:33,500
in some ways that also
169
00:09:33,500 --> 00:09:35,740
you know
170
00:09:35,740 --> 00:09:37,480
sometimes you want flexibility and generality
171
00:09:37,480 --> 00:09:39,640
and sometimes you want
172
00:09:39,640 --> 00:09:41,600
proofs and guarantees
173
00:09:41,600 --> 00:09:43,840
you want to know that if you are
174
00:09:43,840 --> 00:09:45,700
about to deploy this to go
175
00:09:45,700 --> 00:09:47,880
achieve something it's going to do what you
176
00:09:47,880 --> 00:09:49,640
want not what you don't want
177
00:09:49,640 --> 00:10:03,020
And so I think that's the thing is where, and the other thing about this is that, so these generative AI models are trained from unstructured data.
178
00:10:03,840 --> 00:10:16,740
And that is really great for figuring out representations of the world, like kind of, you know, creating some kind of understanding.
179
00:10:16,740 --> 00:10:19,980
But then when you want to go use it, right?
180
00:10:19,980 --> 00:10:21,840
When you want to act in the world,
181
00:10:22,320 --> 00:10:24,920
which by the way, a book is not the world.
182
00:10:25,320 --> 00:10:25,760
It's not?
183
00:10:26,180 --> 00:10:27,000
It's my world.
184
00:10:28,540 --> 00:10:32,740
I mean, like a computer file system
185
00:10:32,740 --> 00:10:35,940
where you can run execute code is also not the world.
186
00:10:36,720 --> 00:10:39,700
And so the thing is, is that what all these people are,
187
00:10:39,900 --> 00:10:41,700
what language is so good for,
188
00:10:42,960 --> 00:10:44,260
it's just not necessarily the same.
189
00:10:44,260 --> 00:10:45,880
You can't just like take that immediately
190
00:10:45,880 --> 00:10:48,840
and go use it to run agents in the world.
191
00:10:50,340 --> 00:10:52,140
Yeah, I mean, it makes me think about
192
00:10:52,140 --> 00:10:56,440
how a lot of programmers went into this line of business
193
00:10:56,440 --> 00:10:59,840
because of all the complexities and nuance
194
00:10:59,840 --> 00:11:01,820
that goes with communicating with humans,
195
00:11:02,460 --> 00:11:06,240
whereas the beauty of at least the way code used to work
196
00:11:06,240 --> 00:11:12,960
is it is entirely unambiguous what things mean.
197
00:11:14,240 --> 00:11:15,280
At least if you wrote it.
198
00:11:15,280 --> 00:11:17,300
If somebody else wrote it, it might be ambiguous.
199
00:11:17,740 --> 00:11:22,380
It might be ambiguous, but to the compiler, it's unambiguous, right?
200
00:11:22,500 --> 00:11:25,240
Like interpreting why did they write this code is ambiguous,
201
00:11:25,240 --> 00:11:31,380
but it is this subset of language where we've kind of defined it down
202
00:11:31,380 --> 00:11:35,480
to have this exact action or result.
203
00:11:35,860 --> 00:11:38,880
Whereas when we're programming with an LLM,
204
00:11:39,320 --> 00:11:43,420
we don't have any idea with any kind of certainty
205
00:11:43,420 --> 00:11:45,860
what that result is going to be.
206
00:11:46,700 --> 00:11:51,200
And all the complexities of dealing with human beings,
207
00:11:51,520 --> 00:11:55,100
which is there's misunderstandings, there's nuance,
208
00:11:56,100 --> 00:11:59,600
and semantic meaning is inherently open-ended.
209
00:11:59,820 --> 00:12:01,920
So like we see this, for example, with prompt injection
210
00:12:01,920 --> 00:12:04,280
where people will try to filter it,
211
00:12:04,620 --> 00:12:06,760
and then there's just a way to rephrase it
212
00:12:06,760 --> 00:12:09,160
or re-encode the meaning in a different way,
213
00:12:09,200 --> 00:12:11,480
and it's hard to constrain that entire space.
214
00:12:11,480 --> 00:12:24,080
or situations where the AI will achieve the benchmark
215
00:12:24,080 --> 00:12:28,540
by hacking the machine and faking the result or something like that,
216
00:12:28,660 --> 00:12:31,840
where now you're dealing with English, you're dealing with meaning,
217
00:12:32,240 --> 00:12:34,800
and you're saying, accomplish this goal,
218
00:12:34,800 --> 00:12:38,080
and you have to then think about, well, how is that going to be interpreted?
219
00:12:38,600 --> 00:12:40,700
And there's many ways it can be interpreted.
220
00:12:41,480 --> 00:12:48,380
So it's both the kind of greatest strength is that it has this open-endedness to it,
221
00:12:48,640 --> 00:12:51,960
where it's general and you don't have to know in advance,
222
00:12:52,480 --> 00:12:57,280
like some big, you know, if branch tree for every possible thing that can happen.
223
00:12:58,020 --> 00:13:02,460
But on the other hand, we introduce this kind of unpredictability
224
00:13:02,460 --> 00:13:04,760
into the system at the same time.
225
00:13:05,300 --> 00:13:09,400
I'll say it's just been wild to last few years coding,
226
00:13:09,400 --> 00:13:12,640
especially coding a production system that touches users
227
00:13:12,640 --> 00:13:15,920
when you have to have a function call
228
00:13:15,920 --> 00:13:17,900
that is non-deterministic
229
00:13:17,900 --> 00:13:19,340
oh my god
230
00:13:19,340 --> 00:13:23,580
what I mean by that is whenever you call an LLM API
231
00:13:23,580 --> 00:13:25,1000
so in some ways it's gotten better
232
00:13:25,1000 --> 00:13:29,140
and there's these tools to refine prompts
233
00:13:29,140 --> 00:13:31,700
and you can try setting temperature equals to zero
234
00:13:31,700 --> 00:13:35,440
but I actually caught
235
00:13:35,440 --> 00:13:39,320
I actually was kind of even getting blamed
236
00:13:39,320 --> 00:13:40,460
for something not working.
237
00:13:40,880 --> 00:13:43,180
But it was like a rare bug that was happening.
238
00:13:43,260 --> 00:13:45,000
It would only happen, honestly,
239
00:13:45,180 --> 00:13:48,040
like it only happened every couple weeks or something.
240
00:13:48,660 --> 00:13:50,980
And I was sure that I had,
241
00:13:51,080 --> 00:13:52,620
I was literally running evals, everything.
242
00:13:52,760 --> 00:13:55,400
I was sure that the model provider
243
00:13:55,400 --> 00:13:56,740
was doing something funny.
244
00:13:58,120 --> 00:13:59,180
Temperature was zero.
245
00:13:59,560 --> 00:14:00,820
And, you know, temperature being zero
246
00:14:00,820 --> 00:14:03,540
is not a guarantee that you're going to get the same thing.
247
00:14:04,060 --> 00:14:06,640
But, you know, most of the time it would work.
248
00:14:06,740 --> 00:14:08,600
And that's kind of one of the best controls you had
249
00:14:08,600 --> 00:14:09,160
if you're not,
250
00:14:09,320 --> 00:14:10,540
running the model yourself.
251
00:14:11,260 --> 00:14:13,740
Or you can't like set other parameters.
252
00:14:14,480 --> 00:14:19,340
So, but anyway, one day I finally got hard proof.
253
00:14:19,480 --> 00:14:22,740
I got like hard traces of calling this model provider
254
00:14:22,740 --> 00:14:25,320
with the same thing and getting different answers back.
255
00:14:25,760 --> 00:14:28,660
And I was mostly happy because I was like going insane.
256
00:14:28,840 --> 00:14:30,520
I was like, I agree.
257
00:14:30,920 --> 00:14:35,040
Right. I mean, I've been seeing this a lot on X recently.
258
00:14:35,420 --> 00:14:37,820
And I feel like I've been seeing it over the last few months,
259
00:14:37,820 --> 00:14:40,540
but it seems like it's kind of increasing in intensity
260
00:14:40,540 --> 00:14:44,980
where people are convinced that Anthropic
261
00:14:44,980 --> 00:14:49,440
dumbed down the model or they quantized it
262
00:14:49,440 --> 00:14:53,720
and Opus 4.6 isn't the way it used to be a month ago.
263
00:14:53,920 --> 00:14:54,860
It's not as smart.
264
00:14:55,460 --> 00:14:58,680
Maybe they need more compute for Mythos or whatever.
265
00:14:59,000 --> 00:15:00,920
I've seen similar accusations.
266
00:15:01,440 --> 00:15:03,860
A lot of times when new models drop,
267
00:15:03,940 --> 00:15:06,240
whether from OpenAI or Anthropic,
268
00:15:06,240 --> 00:15:09,600
There's like this kind of honeymoon phase where everybody's mind is blown.
269
00:15:09,820 --> 00:15:12,880
And then people insist they dumbed it down.
270
00:15:13,400 --> 00:15:21,820
And it's really hard to know what if is that happening or is that some kind of human hallucination.
271
00:15:22,220 --> 00:15:35,840
And another wildcard factor, at least I found in my own coding, is like even if the model is exactly the same, the way I approach the model in a coding session might be different if my mood is different.
272
00:15:36,240 --> 00:16:01,600
So the language I use, the words I input into the machine are going to change based upon how I change. And that's another kind of non-deterministic factor. So how much of this is people got a good result, then they got frustrated, then they get in a bad mood, then they start yelling at their AI, then their AI starts spinning out?
273
00:16:01,600 --> 00:16:04,920
or how much is it that Anthropic is actually doing this
274
00:16:04,920 --> 00:16:06,380
or some combination of the both.
275
00:16:06,860 --> 00:16:11,800
And it's another example of how the open-endedness
276
00:16:11,800 --> 00:16:14,280
and the non-determinism creates a lot of confusion
277
00:16:14,280 --> 00:16:15,900
about what the system is actually doing.
278
00:16:16,620 --> 00:16:17,880
I have a hypothesis here.
279
00:16:18,100 --> 00:16:18,840
You mind if I talk?
280
00:16:18,860 --> 00:16:19,600
Please jump in.
281
00:16:19,720 --> 00:16:19,980
Yeah, yeah.
282
00:16:21,340 --> 00:16:24,280
So every time you retrain a model,
283
00:16:25,260 --> 00:16:29,000
you are fundamentally getting a different hypothesis
284
00:16:29,000 --> 00:16:33,140
about a connection between language and meaning.
285
00:16:34,760 --> 00:16:38,000
And I think that each of us,
286
00:16:38,320 --> 00:16:40,060
each person you know,
287
00:16:40,780 --> 00:16:45,240
has a model of some kind in their minds
288
00:16:45,240 --> 00:16:48,800
that is connecting language to meaning.
289
00:16:48,920 --> 00:16:50,260
But we don't all have the same one,
290
00:16:50,320 --> 00:16:52,280
and that's what creates Wilson communication.
291
00:16:52,280 --> 00:16:54,060
In part, there's probably more to it than that.
292
00:16:56,040 --> 00:16:57,480
But...
293
00:16:59,000 --> 00:17:04,300
When that shifts, it's fundamentally like you're talking to a different person.
294
00:17:04,460 --> 00:17:09,280
To the extent that we anthropomorphize these tools like Claude,
295
00:17:10,260 --> 00:17:15,380
we think of it as being reliable.
296
00:17:15,720 --> 00:17:17,980
We make a relationship with it.
297
00:17:17,980 --> 00:17:26,700
And when they introduce a new model, really it's like somebody went through a fundamental change.
298
00:17:26,820 --> 00:17:27,820
They're not the same person.
299
00:17:27,820 --> 00:17:29,420
they were before.
300
00:17:30,500 --> 00:17:32,040
And, you know, it's going to be
301
00:17:32,040 --> 00:17:33,860
fascinating at first, oh, well,
302
00:17:34,520 --> 00:17:36,240
you seem to have
303
00:17:36,240 --> 00:17:38,080
gone away and learned all this new stuff.
304
00:17:38,660 --> 00:17:39,820
Right. But, you know,
305
00:17:39,900 --> 00:17:42,000
as you talk with them more, you're
306
00:17:42,000 --> 00:17:43,640
going to say, well, I
307
00:17:43,640 --> 00:17:46,020
miss things. Yes.
308
00:17:46,440 --> 00:17:47,120
And
309
00:17:47,120 --> 00:17:50,700
it is,
310
00:17:52,060 --> 00:17:53,840
I mean, you can go too far in
311
00:17:53,840 --> 00:17:55,440
anthropomorphizing this stuff.
312
00:17:57,060 --> 00:17:57,620
But,
313
00:17:57,820 --> 00:18:02,320
I don't actually think it's necessarily the wrong thing to do
314
00:18:02,320 --> 00:18:09,640
because MLMs are like humans in a lot of important ways
315
00:18:09,640 --> 00:18:15,120
in the way we process meaning and we share meaning with one another.
316
00:18:16,400 --> 00:18:23,440
And so thinking of it as another person with a viewpoint is correct, I think.
317
00:18:23,440 --> 00:18:29,360
that there is a viewpoint inherent in the way we process language.
318
00:18:31,940 --> 00:18:36,740
You know, linguists, I think, talk about something called the Sapir-Whorf hypothesis, right,
319
00:18:36,800 --> 00:18:42,460
where the words you have access to, the language you have influences how you think.
320
00:18:44,600 --> 00:18:50,240
And I think that we're seeing that with these agents,
321
00:18:50,240 --> 00:18:53,800
and that regardless of how much Anthropic is doing
322
00:18:53,800 --> 00:18:57,860
to put something on top of the model
323
00:18:57,860 --> 00:19:01,780
that makes it easier for people to use
324
00:19:01,780 --> 00:19:05,400
and changes how it works,
325
00:19:05,700 --> 00:19:06,660
when they retrain the model,
326
00:19:06,760 --> 00:19:08,300
they don't have the power
327
00:19:08,300 --> 00:19:12,840
to ensure that it's going to give the same experience
328
00:19:12,840 --> 00:19:16,060
because it already thinks differently
329
00:19:16,060 --> 00:19:17,800
every time you retrain it.
330
00:19:17,800 --> 00:19:19,800
Mm-hmm.
331
00:19:20,240 --> 00:19:21,200
Right, yeah.
332
00:19:21,700 --> 00:19:25,460
I mean, I think I've heard certain people refer to it as like growing a model.
333
00:19:25,640 --> 00:19:28,560
Like there's almost like this organic process to it.
334
00:19:28,560 --> 00:19:31,940
And it has the complexity of an organic process.
335
00:19:32,180 --> 00:19:36,240
And it has the, I want to say epiphenomenality.
336
00:19:37,200 --> 00:19:40,900
I don't know whether the word epiphenomenon is particularly recognizable.
337
00:19:41,440 --> 00:19:46,340
But people know about fractals and stuff like that.
338
00:19:46,340 --> 00:19:51,220
how you get emergent properties from something that grows from little bits.
339
00:19:51,700 --> 00:19:57,040
And I think that's the organicness that we're talking about there,
340
00:19:57,540 --> 00:20:01,840
that it is sufficiently complex and it's small rules
341
00:20:01,840 --> 00:20:08,140
that the whole picture is unpredictable.
342
00:20:08,720 --> 00:20:10,500
It looks very different.
343
00:20:11,520 --> 00:20:15,400
And we can't, I mean, this is what I've heard from other people,
344
00:20:15,400 --> 00:20:22,120
that if you try to go in its brain and look at the weights and understand how it's doing
345
00:20:22,120 --> 00:20:29,040
what it's doing or the specific models or logical constructions, it's not totally transparent
346
00:20:29,040 --> 00:20:30,100
to us either.
347
00:20:31,000 --> 00:20:36,940
And that would also kind of, I think, maybe contribute to that sense of that it's something
348
00:20:36,940 --> 00:20:41,360
that's grown or it's something that we can't fully predict how it's going to behave.
349
00:20:41,360 --> 00:20:44,960
and to your point about personalities
350
00:20:44,960 --> 00:20:47,640
you know I remember I think this happened with 4.0
351
00:20:47,640 --> 00:20:49,000
where it was like particularly
352
00:20:49,000 --> 00:20:50,420
this was a big deal for a lot of people
353
00:20:50,420 --> 00:20:52,020
they were very emotionally upset
354
00:20:52,020 --> 00:20:54,300
it was very sycophantic
355
00:20:54,300 --> 00:20:57,680
and people got really attached to it
356
00:20:57,680 --> 00:21:01,300
and then I think some of the open AI engineers
357
00:21:01,300 --> 00:21:03,180
were saying like 4.0 was a mistake
358
00:21:03,180 --> 00:21:05,500
it's like creating all this AI psychosis
359
00:21:05,500 --> 00:21:07,100
like we never should have put it out there
360
00:21:07,100 --> 00:21:08,940
and then they added a new one
361
00:21:08,940 --> 00:21:10,280
and they got rid of 4.0
362
00:21:10,280 --> 00:21:13,080
and then all the 4-0 people freaked out and they added it back.
363
00:21:13,840 --> 00:21:19,500
But it is like whether you anthropomorphize it or not
364
00:21:19,500 --> 00:21:23,760
or whether you think of it as alive or not,
365
00:21:24,580 --> 00:21:33,040
it creates an illusion of an other that has a personality
366
00:21:33,040 --> 00:21:35,340
that you develop a relationship with.
367
00:21:36,200 --> 00:21:38,980
And if suddenly that personality shifts
368
00:21:38,980 --> 00:21:44,380
or its brain changes and you're used to relating to it,
369
00:21:44,680 --> 00:21:46,320
it can be very jarring.
370
00:21:46,980 --> 00:21:52,760
And I think this is why I'm really bullish on open weight models
371
00:21:52,760 --> 00:21:57,220
because I want to have a model where I don't have to wonder,
372
00:21:57,560 --> 00:21:59,680
like, did it change, right?
373
00:21:59,760 --> 00:22:01,900
It changes when I decide it changes
374
00:22:01,900 --> 00:22:05,840
and I get to choose the weights or fine-tune them
375
00:22:05,840 --> 00:22:07,940
or switch the model when I want.
376
00:22:08,980 --> 00:22:11,720
And nobody can tell me,
377
00:22:11,860 --> 00:22:14,020
oh, you shouldn't use 4.0 anymore
378
00:22:14,020 --> 00:22:15,640
because it's bad for you or whatever.
379
00:22:15,940 --> 00:22:17,720
If I like the way it's relating,
380
00:22:18,380 --> 00:22:21,300
it stays that way until I decide to change it.
381
00:22:23,080 --> 00:22:26,280
And I think, I mean, there are a lot of problems with,
382
00:22:26,280 --> 00:22:28,920
at this stage, we're still very early on
383
00:22:28,920 --> 00:22:34,340
in understanding how people will use conversational AI
384
00:22:34,340 --> 00:22:35,980
over the long term.
385
00:22:35,980 --> 00:22:48,000
And I think that one of the problems with having a lot of it controlled by these private companies at the moment is, you know, there's a lack of transparency.
386
00:22:49,520 --> 00:22:55,200
There's a lack of understanding of what is part of the trained model and what isn't.
387
00:22:55,200 --> 00:23:07,240
And there's a lack of security fundamentally in that if you want the good performance, you have to give up the certainty that it's in your sphere of control where the information is going.
388
00:23:07,240 --> 00:23:30,460
And so I think that over the long term, open models, local models will become the norm. It's just we have not gotten to the point where that is efficient enough or, I guess, really, really practical. But it seems like everybody wants that, right?
389
00:23:30,460 --> 00:24:01,358
I think so yeah I mean as we get more dependent on these systems and we put more and more private and intimate information in them it really across at least those two dimensions right That it like you know if I using this thing as my therapist I don exactly want everybody to know that Or even as a lawyer like I saw you know people use ChatGPT as a lawyer And then all of those conversations are subject to discovery in court because there no attorney privilege
390
00:24:01,358 --> 00:24:07,378
and talk about a backfire oh my god can i do this no it's illegal can i do this no it's illegal
391
00:24:07,378 --> 00:24:12,618
and it's like okay so what and then you know if you don't follow the ai and you do something else
392
00:24:12,618 --> 00:24:18,998
and then you might be able to say you you knew it was illegal yeah or you're you know all these
393
00:24:18,998 --> 00:24:25,178
ways of interpreting the intention of the contract or the circumstance or your motivation or everything
394
00:24:25,178 --> 00:24:29,218
else is now available in litigation.
395
00:24:30,598 --> 00:24:32,918
I mean, there's just countless examples of this.
396
00:24:33,698 --> 00:24:37,138
And so I think for that reason, there's going to be a big desire, whether
397
00:24:37,138 --> 00:24:41,258
for individuals or businesses that want to have control over their
398
00:24:41,258 --> 00:24:44,738
data. But then the other thing about
399
00:24:44,738 --> 00:24:49,258
kind of, I think anything that we
400
00:24:49,258 --> 00:24:53,038
grow really dependent on, there's a natural
401
00:24:53,038 --> 00:24:56,058
desire to have control over it.
402
00:24:56,538 --> 00:25:00,198
Because you just feel vulnerable where you're like, if this thing was taken away all of a
403
00:25:00,198 --> 00:25:04,258
sudden, I'm lost in the world.
404
00:25:04,558 --> 00:25:10,198
Imagine after 10 and 20 years of us using AI every single day, and then all of a sudden
405
00:25:10,198 --> 00:25:11,398
it's taken away from you.
406
00:25:11,398 --> 00:25:17,178
Your ability to work in the economy, or who knows what kind of downstream impacts that
407
00:25:17,178 --> 00:25:17,578
would have.
408
00:25:17,578 --> 00:25:29,998
So I think as it becomes something that is so essential to human life, humans want to have control and possession and ownership of something that is so essential to how they live.
409
00:25:29,998 --> 00:25:31,058
and rights.
410
00:25:31,738 --> 00:25:33,858
I mean, an editor friend of mine,
411
00:25:34,018 --> 00:25:34,798
Catherine Forrest,
412
00:25:35,298 --> 00:25:37,198
was recently posting about how
413
00:25:37,198 --> 00:25:40,218
there are court cases right now
414
00:25:40,218 --> 00:25:41,338
about what is copyrightable
415
00:25:41,338 --> 00:25:46,738
if you work with an LLN to generate content.
416
00:25:47,358 --> 00:25:48,838
And it really matters, I think,
417
00:25:49,238 --> 00:25:51,018
how you're doing it, right?
418
00:25:51,058 --> 00:25:52,578
If you ask it to write something
419
00:25:52,578 --> 00:25:54,238
and you quit that out,
420
00:25:55,478 --> 00:25:58,818
then it's definitely not copyrightable,
421
00:25:58,818 --> 00:25:59,738
I guess.
422
00:25:59,998 --> 00:26:04,658
If you write something and ask it to revise it, yeah, that's copyrightable.
423
00:26:05,058 --> 00:26:10,338
If you ask it to write something and then you edit it,
424
00:26:10,858 --> 00:26:15,158
then your modifications are copyrightable, right?
425
00:26:15,878 --> 00:26:20,758
So I think people don't understand their rights with respect to it.
426
00:26:21,438 --> 00:26:27,898
And I think that, I mean, I don't know what the right answer is,
427
00:26:27,898 --> 00:26:29,418
what these things should do in the future,
428
00:26:29,718 --> 00:26:31,878
but fundamentally to me,
429
00:26:32,078 --> 00:26:36,718
it feels like people are using it as a tool
430
00:26:36,718 --> 00:26:40,318
and they expect that if they do something with it,
431
00:26:40,878 --> 00:26:45,878
and they put their credibility behind it,
432
00:26:45,878 --> 00:26:48,038
they should be able to treat it as theirs.
433
00:26:49,018 --> 00:26:51,458
And I think there's a strong argument there.
434
00:26:51,838 --> 00:26:54,738
I also think that people shouldn't just put things out
435
00:26:54,738 --> 00:26:57,878
from an LLM without reading and considering
436
00:26:57,878 --> 00:26:59,138
whether they really believe them,
437
00:26:59,178 --> 00:27:00,598
which a lot of people are doing right now.
438
00:27:00,678 --> 00:27:01,198
It's problematic.
439
00:27:02,898 --> 00:27:05,078
But, you know,
440
00:27:05,098 --> 00:27:07,358
if somebody is willing to stand behind something
441
00:27:07,358 --> 00:27:09,558
and nobody else, you know,
442
00:27:09,778 --> 00:27:11,978
bothered to take the time
443
00:27:11,978 --> 00:27:13,938
to decide what it should be,
444
00:27:14,858 --> 00:27:16,438
I mean, that is a product of them.
445
00:27:17,478 --> 00:27:19,438
I mean, it's the verifier step, right?
446
00:27:19,558 --> 00:27:22,238
So even like the way these things are trained
447
00:27:22,238 --> 00:27:25,118
or at least in post-training and other things,
448
00:27:25,118 --> 00:27:28,158
there is this notion of the verifier.
449
00:27:28,398 --> 00:27:33,218
Like you need some end thing, human something,
450
00:27:33,438 --> 00:27:37,838
some function to say this is success or this is true
451
00:27:37,838 --> 00:27:39,838
or this is useful and this is not.
452
00:27:39,958 --> 00:27:41,098
Because with these things,
453
00:27:41,198 --> 00:27:44,678
you can generate infinite amount of copy and text
454
00:27:44,678 --> 00:27:45,898
and marketing and all that.
455
00:27:46,258 --> 00:27:50,898
So even if it is just simply you choosing this one
456
00:27:50,898 --> 00:27:54,558
out of five or 10 or 100 or 1,000 samples that's generated,
457
00:27:54,558 --> 00:28:04,798
that alone is a very in some ways that's the most valuable special capability because
458
00:28:04,798 --> 00:28:13,058
like fundamentally it cannot do that right yeah i mean it i'd love to dig into the case law on this
459
00:28:13,058 --> 00:28:19,918
a little bit because it makes me think about other examples like if i take if i take a paintbrush
460
00:28:19,918 --> 00:28:23,198
and I create like a machine where it just spins around
461
00:28:23,198 --> 00:28:26,738
and I'm trying to create some kind of like Jackson Pollock thing
462
00:28:26,738 --> 00:28:28,698
which is like streaks of paint, right?
463
00:28:28,778 --> 00:28:32,798
And I just hit the go button and it does that and I hit stop.
464
00:28:33,438 --> 00:28:36,238
Is that not copyrightable art
465
00:28:36,238 --> 00:28:39,398
because I used a machine to spin a bunch of paint around?
466
00:28:39,818 --> 00:28:41,478
And does it matter who created the machine?
467
00:28:42,658 --> 00:28:45,378
I think it's, if I had to say,
468
00:28:45,458 --> 00:28:47,858
I think it's you pressing the stop button
469
00:28:47,858 --> 00:28:50,358
when you decided to press the stop button
470
00:28:50,358 --> 00:28:55,138
and you deciding whether or not to even take it
471
00:28:55,138 --> 00:28:56,698
and do anything else with it.
472
00:28:57,458 --> 00:28:59,878
Right, but how is that any different than an LLM?
473
00:28:59,978 --> 00:29:04,458
Because I'm deciding to, I put the prompt in, right?
474
00:29:04,558 --> 00:29:07,998
I may have a few back and forths decide when it's done.
475
00:29:07,998 --> 00:29:13,758
And so where does the line come in?
476
00:29:13,918 --> 00:29:17,598
If we think about using tools or automation
477
00:29:17,598 --> 00:29:19,478
or machines or robots
478
00:29:19,478 --> 00:29:21,758
assisting in the process of
479
00:29:21,758 --> 00:29:23,338
creating something,
480
00:29:24,558 --> 00:29:26,598
where do we draw that line?
481
00:29:26,658 --> 00:29:28,158
Because in a way, it's almost as if
482
00:29:28,158 --> 00:29:29,878
people are kind of
483
00:29:29,878 --> 00:29:31,578
mythologizing the LLM.
484
00:29:31,818 --> 00:29:33,878
It's just this magical
485
00:29:33,878 --> 00:29:35,918
AI box, and so the normal rules
486
00:29:35,918 --> 00:29:37,758
don't apply. Someone should name a new
487
00:29:37,758 --> 00:29:38,698
LM mythos.
488
00:29:38,858 --> 00:29:40,638
Something myth-based.
489
00:29:40,638 --> 00:29:43,338
AI box right there.
490
00:29:44,318 --> 00:29:45,538
No editing necessary.
491
00:29:46,038 --> 00:29:47,058
Just the full thing.
492
00:29:47,058 --> 00:29:49,038
Maybe like operating system plus myth.
493
00:29:49,358 --> 00:29:50,498
So there's something there, I think.
494
00:29:50,878 --> 00:29:51,178
Right.
495
00:29:51,418 --> 00:29:51,778
Exactly.
496
00:29:52,478 --> 00:29:52,658
Yeah.
497
00:29:53,438 --> 00:29:58,458
But the, or like I think William S. Burroughs in Naked Lunch,
498
00:29:58,638 --> 00:30:03,178
there was something where he like cut up a bunch of pages of something
499
00:30:03,178 --> 00:30:05,298
and like randomly rearranged them.
500
00:30:06,038 --> 00:30:07,858
Is that not copyrightable?
501
00:30:07,918 --> 00:30:10,138
I mean, that seems like a very artistic expression.
502
00:30:10,138 --> 00:30:16,538
So, yeah, I'm curious in these copyright cases,
503
00:30:17,058 --> 00:30:23,358
how these nuances are getting adjudicated and where the judges are drawing the line,
504
00:30:23,438 --> 00:30:32,998
and if there is principled, informed reasons, or if it's kind of the early state of the law here,
505
00:30:33,118 --> 00:30:38,578
and by the time it hits the Supreme Court or, you know, higher courts,
506
00:30:38,718 --> 00:30:41,498
that maybe they're going to come to a different conclusion about this.
507
00:30:41,498 --> 00:30:44,898
I would suspect that it would change over time,
508
00:30:44,958 --> 00:30:47,278
but I am really not an expert on the law,
509
00:30:47,378 --> 00:30:49,678
and I'm just, you know, talking half-assed about something
510
00:30:49,678 --> 00:30:54,438
that I read online that was filtered through a friend of mine.
511
00:30:55,598 --> 00:30:56,878
Well, that's what we're all about here.
512
00:30:57,038 --> 00:30:57,198
Yeah.
513
00:31:00,078 --> 00:31:03,378
But yeah, I think that it's really interesting,
514
00:31:03,538 --> 00:31:06,818
and it's important where these rights decisions land.
515
00:31:08,718 --> 00:31:09,118
And...
516
00:31:11,498 --> 00:31:13,158
That process is always messy.
517
00:31:13,518 --> 00:31:17,018
Yeah, I mean, there's another big legal,
518
00:31:17,518 --> 00:31:19,758
I love to nerd out on legal stuff
519
00:31:19,758 --> 00:31:20,698
because I used to be a lawyer.
520
00:31:21,458 --> 00:31:23,398
So there's one, there's a case,
521
00:31:23,498 --> 00:31:26,338
I mean, there's all these lawsuits,
522
00:31:27,798 --> 00:31:30,618
you know, oftentimes very tragic circumstances
523
00:31:30,618 --> 00:31:33,358
where a teenager commits suicide
524
00:31:33,358 --> 00:31:36,138
after talking to a chat bot or these kinds of things.
525
00:31:36,258 --> 00:31:39,838
And so these companies are getting sued left and right.
526
00:31:39,838 --> 00:31:47,938
And one of the arguments that is being mounted in court is that the LLM speech is protected under the First Amendment.
527
00:31:49,098 --> 00:31:57,978
And what's interesting there is not, because at first it sounds like, oh, the AI has First Amendment rights to speak, which is a little dubious.
528
00:31:58,398 --> 00:32:00,378
But that's not actually the argument.
529
00:32:00,618 --> 00:32:06,538
The argument is that the First Amendment also protects the right to hear speech.
530
00:32:06,538 --> 00:32:09,418
The listener has First Amendment protection.
531
00:32:09,838 --> 00:32:31,498
So if there's speech in the marketplace, however it was produced, if I want to hear it and the government is getting in the way of that, that's a violation of my First Amendment right. And so they are situating this in that theory where they're saying the person who wants to hear the chatbot, they have a right to that.
532
00:32:31,498 --> 00:32:38,698
and you could imagine kind of how what the consequences might be if the case law goes the
533
00:32:38,698 --> 00:32:46,218
other way if the government has the power to pass a law that says what your chatbot is allowed to say
534
00:32:46,218 --> 00:32:51,578
or not allowed to say i mean first there's like the technical problem of no one really knows how
535
00:32:51,578 --> 00:32:58,218
to implement that but also um we're going to get these ridiculous laws where it's going to be like
536
00:32:58,218 --> 00:33:05,418
your AI is not allowed to tell you that this policy or this politician makes sense. Or every
537
00:33:05,418 --> 00:33:11,218
politician in the world is going to want to get their fingers in there and say, for safety or for
538
00:33:11,218 --> 00:33:18,758
the children, the AI has to say this or can never say this. And that to me seems like, at least in
539
00:33:18,758 --> 00:33:23,818
violation of the spirit of the First Amendment, that when we think about the marketplace of ideas
540
00:33:23,818 --> 00:33:30,778
and freedom of thought that it kind of seems inappropriate for the state to be dictating that.
541
00:33:31,478 --> 00:33:38,918
But I think that's another place to watch to see how the law plays out with these issues.
542
00:33:40,318 --> 00:33:47,158
I mean, one thing I feel about this is, you know, this is kind of early technology.
543
00:33:47,158 --> 00:33:48,698
it's very powerful.
544
00:33:49,798 --> 00:33:54,038
But we, you know, there's been a lot of hype.
545
00:33:54,298 --> 00:33:55,578
I mean, you have to generate,
546
00:33:55,738 --> 00:33:58,378
I mean, if you want to get the word out there
547
00:33:58,378 --> 00:34:00,398
and get people excited, you've got to hype things up.
548
00:34:00,458 --> 00:34:04,538
And at the same time, you know, it's early technology.
549
00:34:04,658 --> 00:34:05,398
It's new technology.
550
00:34:05,878 --> 00:34:08,538
There's inherent, you know, limits on it.
551
00:34:09,138 --> 00:34:15,798
And so, you know, there is the disclaimer, right,
552
00:34:15,798 --> 00:34:17,718
that like it can make mistakes.
553
00:34:18,298 --> 00:34:19,038
Maybe that's not,
554
00:34:19,158 --> 00:34:20,678
maybe that should be elaborated on.
555
00:34:23,818 --> 00:34:26,398
Yeah, it's kind of just a delicate thing.
556
00:34:26,538 --> 00:34:28,318
And I also just wonder like
557
00:34:28,318 --> 00:34:30,698
where personal responsibility comes into this.
558
00:34:30,798 --> 00:34:33,218
I guess, is this mostly an issue
559
00:34:33,218 --> 00:34:36,378
because of people who are under 18?
560
00:34:36,878 --> 00:34:39,218
I think that is where the majority
561
00:34:39,218 --> 00:34:41,518
of the issue is right now.
562
00:34:41,919 --> 00:34:43,458
And I think that's at least
563
00:34:43,458 --> 00:34:45,658
the strongest case you could make.
564
00:34:45,798 --> 00:34:49,358
is to say, you know, minors, you know,
565
00:34:49,398 --> 00:34:52,118
don't have access to the full spectrum of rights.
566
00:34:52,138 --> 00:34:53,578
Although there is case about, you know,
567
00:34:54,138 --> 00:34:58,738
minors do have First Amendment protection to some extent.
568
00:34:58,858 --> 00:35:03,678
But I think, you know, if it only applies to minors,
569
00:35:03,778 --> 00:35:06,919
the impact on the broader underlying policy concerns
570
00:35:06,919 --> 00:35:09,198
of the First Amendment are not as impacted
571
00:35:09,198 --> 00:35:10,758
as if it applied to everyone.
572
00:35:11,218 --> 00:35:11,238
Okay.
573
00:35:11,958 --> 00:35:14,939
So, yeah, but, you know,
574
00:35:14,939 --> 00:35:21,919
I think all these different nuances and distinctions are going to play themselves out in the courts to see where they draw the line.
575
00:35:23,978 --> 00:35:31,138
But yeah, there's going to be, you know, another situation could be, you know,
576
00:35:31,218 --> 00:35:36,138
what if somebody uses an AI to hack into something or build a virus or whatever?
577
00:35:36,358 --> 00:35:39,158
Like, should the company that created it be responsible?
578
00:35:39,318 --> 00:35:44,038
And this kind of relates to what we were alluding to before, Mythos.
579
00:35:44,038 --> 00:35:46,178
So Mythos is the new
580
00:35:46,178 --> 00:35:48,258
Anthropic model that apparently
581
00:35:48,258 --> 00:35:50,398
Is too dangerous to release to the public
582
00:35:50,398 --> 00:35:51,678
And
583
00:35:51,678 --> 00:35:54,078
And you know
584
00:35:54,078 --> 00:35:56,758
People are accusing Anthropic of overhyping
585
00:35:56,758 --> 00:35:58,638
And some people are saying
586
00:35:58,638 --> 00:35:59,838
Hey you know I took
587
00:35:59,838 --> 00:36:02,138
An open weight model or open source
588
00:36:02,138 --> 00:36:04,419
And I pointed it at the same code base
589
00:36:04,419 --> 00:36:05,919
And I found all these zero days
590
00:36:05,919 --> 00:36:07,658
And I didn't need Mythos to do it
591
00:36:07,658 --> 00:36:10,678
But on the other hand
592
00:36:10,678 --> 00:36:12,638
Like the benchmarks that they're publishing
593
00:36:12,638 --> 00:36:15,578
do seem like a step function up from before.
594
00:36:15,858 --> 00:36:18,298
So, you know, can we discount it completely?
595
00:36:18,478 --> 00:36:20,738
I'm just kind of curious what you guys think.
596
00:36:20,858 --> 00:36:22,338
Like, is it real?
597
00:36:22,498 --> 00:36:23,398
Is it overhyped?
598
00:36:23,938 --> 00:36:25,678
I mean, the funny thing is,
599
00:36:26,498 --> 00:36:28,458
yeah, I'm just thinking about this right now,
600
00:36:28,858 --> 00:36:32,718
but did they have to announce it the way they did?
601
00:36:32,838 --> 00:36:35,058
Like, what if in an alternate timeline,
602
00:36:35,058 --> 00:36:37,338
they just didn't say anything publicly about Mythos?
603
00:36:37,458 --> 00:36:39,218
They still had the private meetings
604
00:36:39,218 --> 00:36:42,098
with all the top companies, let them know.
605
00:36:42,098 --> 00:36:48,338
almost like the way you would disclose any kind of security issue right and maybe they're like well
606
00:36:48,338 --> 00:36:55,439
not only did we find these we actually built a system we think will find even more or could run
607
00:36:55,439 --> 00:37:00,838
it and you know they could just be offering that as a private service or all those so you know you
608
00:37:00,838 --> 00:37:08,078
gotta wonder why make all the buzz around the you know just telling everybody you know you're not
609
00:37:08,078 --> 00:37:13,438
going to release it to them so it's not like you're or maybe this maybe this is right but you
610
00:37:13,438 --> 00:37:18,778
know these days right isn't it all about going viral like anything you can go to go viral and
611
00:37:18,778 --> 00:37:26,378
that alone i think is why before we've talked about like doom and and like you know p doom and
612
00:37:26,378 --> 00:37:32,778
like ai and you know fear of the future that somebody you know i i wonder that you read a new
613
00:37:32,778 --> 00:37:38,818
headline about you know when when things are going to get bad like there's all these predictions
614
00:37:38,818 --> 00:37:46,078
about specific years or you hear all this stuff i wonder if that's just so engaging i mean you see
615
00:37:46,078 --> 00:37:50,938
this anything that kind of baits you to have like fear or anger this is what the social media
616
00:37:50,938 --> 00:37:57,118
algorithms are using to like keep you scrolling and keep you engaged so i almost wonder if it's
617
00:37:57,118 --> 00:37:59,898
just trying to get attention.
618
00:37:59,998 --> 00:38:00,998
Like that's the new marketing.
619
00:38:01,238 --> 00:38:09,198
Like you just marketing via getting your audience to feel negative emotions is, you know, tempting
620
00:38:09,198 --> 00:38:14,778
or at least seems to have initial high metrics, which you can then show and be like, hey, it's
621
00:38:14,778 --> 00:38:15,138
working.
622
00:38:15,218 --> 00:38:18,678
We're getting more attention and more people are talking about our company and can use
623
00:38:18,678 --> 00:38:19,218
our product.
624
00:38:19,478 --> 00:38:24,798
And yeah, I guess that's just the way business is done these days, you know, get people and
625
00:38:24,798 --> 00:38:25,638
I hope there's an alternative.
626
00:38:26,638 --> 00:38:27,358
And cash in.
627
00:38:28,398 --> 00:38:31,278
I think there's maybe another thing on it.
628
00:38:32,958 --> 00:38:36,238
Anthropic recently has had all the press about, you know,
629
00:38:36,458 --> 00:38:42,378
not wanting the U.S. to use their stuff for war, right?
630
00:38:42,638 --> 00:38:42,978
Mm-hmm.
631
00:38:42,978 --> 00:38:50,658
So they have clearly some kind of desire to appear ethical.
632
00:38:50,658 --> 00:38:57,378
and that may well be based on a desire to be ethical at some level
633
00:38:57,378 --> 00:39:01,338
and you know any business is just so many moving pieces
634
00:39:01,338 --> 00:39:03,958
that often get out of sync with one another
635
00:39:03,958 --> 00:39:06,478
that it's quite possible that the people who want to be ethical
636
00:39:06,478 --> 00:39:08,638
and who want to appear ethical are different people
637
00:39:08,638 --> 00:39:17,758
and so it makes me think that this is messaging
638
00:39:17,758 --> 00:39:21,338
not just that they are,
639
00:39:23,398 --> 00:39:25,658
I mean, dangerous sounds like power.
640
00:39:26,558 --> 00:39:28,858
And I think we're focusing on this idea.
641
00:39:28,958 --> 00:39:30,238
Are they trying to sound powerful?
642
00:39:30,618 --> 00:39:31,298
I'm thinking,
643
00:39:31,638 --> 00:39:33,578
what if they're trying to sound ethical?
644
00:39:33,818 --> 00:39:35,218
What if they're trying to sound like
645
00:39:35,218 --> 00:39:38,878
they care where the other guys don't?
646
00:39:40,098 --> 00:39:41,978
I'm not entirely certain that
647
00:39:41,978 --> 00:39:44,898
a business at a certain size can care.
648
00:39:45,478 --> 00:39:45,598
Right.
649
00:39:45,598 --> 00:39:47,098
but
650
00:39:47,098 --> 00:39:49,238
you're just so
651
00:39:49,238 --> 00:39:50,898
at the whims of the market
652
00:39:50,898 --> 00:39:53,078
they're definitely
653
00:39:53,078 --> 00:39:55,278
virtue signaling at some level I think
654
00:39:55,278 --> 00:39:57,018
yeah right and
655
00:39:57,018 --> 00:39:59,278
I think it's important and
656
00:39:59,278 --> 00:40:01,439
you know I think I wish
657
00:40:01,439 --> 00:40:03,278
other companies would virtue signal
658
00:40:03,278 --> 00:40:05,218
more if only because
659
00:40:05,218 --> 00:40:07,278
it creates an incentive not
660
00:40:07,278 --> 00:40:09,518
to be total assets all the time
661
00:40:09,518 --> 00:40:10,878
right yeah
662
00:40:10,878 --> 00:40:12,998
even if you're not doing it for the right reasons
663
00:40:12,998 --> 00:40:15,178
as long as you you know arrive
664
00:40:15,178 --> 00:40:18,618
there, it doesn't really matter how you got there.
665
00:40:19,018 --> 00:40:22,998
If we're in a world where
666
00:40:22,998 --> 00:40:27,198
the most profitable thing is to do the most ethical thing, I think that would
667
00:40:27,198 --> 00:40:30,858
be a pretty nice place to end up, even if the motives aren't pure.
668
00:40:32,098 --> 00:40:35,278
But yeah, I think that's kind of been Anthropik's shtick
669
00:40:35,278 --> 00:40:38,558
from the very beginning, when they broke off from OpenAI.
670
00:40:39,118 --> 00:40:41,978
Is there kind of the more ethical, more
671
00:40:41,978 --> 00:40:47,338
conscientious. That's how they've been trying to define themselves
672
00:40:47,338 --> 00:40:49,618
and define their brand.
673
00:40:50,838 --> 00:40:54,578
But there's a dark side to that as well, which is
674
00:40:54,578 --> 00:40:58,078
they're constantly pushing
675
00:40:58,078 --> 00:41:02,118
centralized control. We need more regulations.
676
00:41:03,198 --> 00:41:06,318
The government needs to... It's kind of funny because on the one hand they're saying we better not
677
00:41:06,318 --> 00:41:10,378
sell this stuff to the Pentagon. On the other hand, hey, the government better take over
678
00:41:10,378 --> 00:41:16,098
everything or you know the world's gonna end so which one is it exactly i'm not sure but um
679
00:41:16,098 --> 00:41:22,718
whenever anybody says ethical or even governance and i'm hearing a lot of conversations about ai
680
00:41:22,718 --> 00:41:30,798
governance i have to wonder who's ethics right yeah because it matters a lot i've done i've been
681
00:41:30,798 --> 00:41:36,858
doing work on on human alignment and ethics over the last few years and you know one of the ways
682
00:41:36,858 --> 00:41:39,618
in my research, we approach this as to say,
683
00:41:39,718 --> 00:41:40,958
okay, I'm going to take a person
684
00:41:40,958 --> 00:41:44,558
and look at their individual preferences
685
00:41:44,558 --> 00:41:47,538
and try to make the same decision that person would.
686
00:41:47,578 --> 00:41:52,419
Because that person is at least somewhat consistent, right?
687
00:41:52,419 --> 00:41:55,678
Whereas if I want to look at humanity's ethics,
688
00:41:55,858 --> 00:41:58,078
they're all over the place, right?
689
00:41:58,458 --> 00:42:00,598
So is it the ethics of, you know,
690
00:42:00,638 --> 00:42:02,198
making money for your stockholders,
691
00:42:02,478 --> 00:42:06,038
which is what, you know, a lot of businesses even,
692
00:42:06,038 --> 00:42:10,258
they say that out loud. It's not even the quiet part at all.
693
00:42:12,678 --> 00:42:15,738
Or is it the ethics of freedom
694
00:42:15,738 --> 00:42:19,538
and to do what you want with your data,
695
00:42:20,378 --> 00:42:23,598
your rights to privacy,
696
00:42:25,198 --> 00:42:30,018
right to use the things that you purchased
697
00:42:30,018 --> 00:42:34,658
when we talk about software rights and stuff like that.
698
00:42:34,658 --> 00:42:45,198
different people clearly, I mean, so a person's right to use the music that they purchase
699
00:42:45,198 --> 00:42:53,198
is clearly bumping up against the company's right to charge for the music they own, right?
700
00:42:54,238 --> 00:42:54,658
Totally.
701
00:42:54,778 --> 00:42:56,838
Whose right wins in these situations?
702
00:42:57,358 --> 00:43:01,838
There isn't actually a clear ethical answer, and just saying that something is ethical
703
00:43:01,838 --> 00:43:03,858
doesn't actually mean that you agree with it
704
00:43:03,858 --> 00:43:04,919
or it's in your best interest.
705
00:43:05,578 --> 00:43:06,678
Right, 100%.
706
00:43:06,678 --> 00:43:09,019
And I mean, there's the category of people
707
00:43:09,019 --> 00:43:12,038
where they're not thinking about ethics at all.
708
00:43:12,459 --> 00:43:14,638
And, you know, it's cynical
709
00:43:14,638 --> 00:43:18,038
or if they talk about it, it's disingenuous.
710
00:43:18,038 --> 00:43:21,258
They're pretending to care about it, but they don't.
711
00:43:21,398 --> 00:43:22,778
There's like that bucket of people.
712
00:43:23,258 --> 00:43:26,698
Then there's the bucket of people who genuinely do care
713
00:43:26,698 --> 00:43:29,718
or are trying to think about ethical, what's ethical.
714
00:43:29,718 --> 00:43:37,718
But within that bucket, like you're saying, that doesn't mean that that person's ethics is your ethics.
715
00:43:38,298 --> 00:43:38,378
Right.
716
00:43:38,538 --> 00:43:43,198
And there's that saying, you know, the road to hell is paved in good intentions.
717
00:43:43,818 --> 00:43:46,098
Good intentions aren't enough.
718
00:43:46,558 --> 00:43:46,758
Right.
719
00:43:46,838 --> 00:43:51,818
It's not enough for Anthropic or whoever to say, we want to do the right thing.
720
00:43:51,959 --> 00:43:57,798
There's no guarantee that something really bad isn't going to come out of their desire to do what's good.
721
00:43:57,798 --> 00:44:04,439
and so and all you know oftentimes in these safety arguments that's what we see right so like
722
00:44:04,439 --> 00:44:11,939
after um september 11th the patriot act is put in and you know you could say one group of people
723
00:44:11,939 --> 00:44:16,218
tried to do that maybe were cynical they just want more power they don't really think it you know
724
00:44:16,218 --> 00:44:21,838
uh is for safety or whatever and then there's another bucket of you know national security
725
00:44:21,838 --> 00:44:27,939
people that are genuinely concerned about terrorism and the safety of the American public.
726
00:44:29,078 --> 00:44:35,858
And in the name of safety, we could easily lose all of our rights and have a totalitarian
727
00:44:35,858 --> 00:44:36,338
state.
728
00:44:36,459 --> 00:44:42,078
And somebody can make a very powerful ethical argument for it to say, these freedoms are
729
00:44:42,078 --> 00:44:47,338
putting us in danger and we have to get rid of them for the greater good.
730
00:44:47,979 --> 00:44:50,939
And they might really believe that and maybe they might be right.
731
00:44:50,939 --> 00:44:57,218
so you know that is kind of at play with this stuff and i don't know if you guys are familiar
732
00:44:57,218 --> 00:45:05,378
with um tristan harris he's been um uh he was one of the people and uh i think he worked at
733
00:45:05,378 --> 00:45:11,158
facebook and he was in this social network documentary and kind of a very powerful voice
734
00:45:11,158 --> 00:45:18,338
about how social media is used to monetize people's attention and and you know um that
735
00:45:18,338 --> 00:45:26,378
there was some unethical uses of it without even, you know, trying to be ethical. But now he's sort
736
00:45:26,378 --> 00:45:35,898
of pivoted to AI and he has this organization and he goes around, he's like, was in the AI
737
00:45:35,898 --> 00:45:42,278
documentary. And I heard recently about a talk he gave, cause he's kind of like a celebrity in some
738
00:45:42,278 --> 00:45:51,638
circles and you know he's giving and he has these very dire predictions about ai um you know even
739
00:45:51,638 --> 00:45:57,298
in this ai documentary the documentarian was asking him like are you gonna have kids like
740
00:45:57,298 --> 00:46:01,098
should people have kids and he was like i don't know you know the world's gonna end in 10 years
741
00:46:01,098 --> 00:46:07,578
like really extreme oh my god and people what i heard is he was giving this talk and people in
742
00:46:07,578 --> 00:46:16,118
audience are crying and getting really upset and in deep despair because they believe what
743
00:46:16,118 --> 00:46:16,678
he's saying.
744
00:46:17,518 --> 00:46:20,738
And then he's telling these people, that's how you should feel.
745
00:46:22,479 --> 00:46:26,999
And they literally have this campaign they're going to do.
746
00:46:27,098 --> 00:46:30,198
They're going to put all these billboards up and all of this stuff, like the kind of
747
00:46:30,198 --> 00:46:35,338
anti-smoking campaigns, but it's going to be like an AI thing to shock people, make
748
00:46:35,338 --> 00:46:37,499
people upset, make people afraid.
749
00:46:37,578 --> 00:46:45,818
And of course, the solution that they're prescribing is, you know, call your senator and, you know, get the government to do something about it.
750
00:46:45,818 --> 00:46:53,358
To the point that I almost wonder if he's like an intelligence operative that's like trying to corral everybody into this.
751
00:46:53,358 --> 00:47:05,158
But like, I think that is really scary to me when there are people out there who want people to be in despair and not to have hope.
752
00:47:05,158 --> 00:47:08,138
Because I think that's incredibly disempowering.
753
00:47:08,858 --> 00:47:14,419
And it literally causes people to give their power away when they're in that state.
754
00:47:14,658 --> 00:47:17,558
Yeah, you literally cannot make good decisions.
755
00:47:17,939 --> 00:47:18,058
Right.
756
00:47:18,258 --> 00:47:24,558
I mean, if you, I mean, do you, you know, you, when you do something when you're angry or you're afraid,
757
00:47:25,218 --> 00:47:29,478
in the long run, if you look back, that decision ends up not being a very good decision.
758
00:47:29,959 --> 00:47:30,158
Right?
759
00:47:30,598 --> 00:47:30,818
Absolutely.
760
00:47:30,818 --> 00:47:42,917
A lot of times people make decisions to I don know stay at their job or or stay in a bad relationship out of fear And then you know later they look back and they absolutely feel
761
00:47:42,917 --> 00:47:47,317
And it's also like you...
762
00:47:47,317 --> 00:47:49,317
Have you ever done that exercise
763
00:47:49,317 --> 00:47:54,557
where you count how many circles you can see in the room?
764
00:47:55,117 --> 00:47:55,377
No.
765
00:47:55,657 --> 00:47:58,157
So there's this exercise where you can...
766
00:47:58,157 --> 00:48:00,237
You know, wherever you are now,
767
00:48:00,297 --> 00:48:01,077
you can take 30 seconds
768
00:48:01,077 --> 00:48:03,097
and look around the room and count how many circles.
769
00:48:03,617 --> 00:48:07,117
And what you'll notice is there's so many other things you did not see.
770
00:48:07,597 --> 00:48:10,517
So what you're thinking and what you're feeling
771
00:48:10,517 --> 00:48:13,897
colors your whole experience and all of your work
772
00:48:13,897 --> 00:48:16,857
and ends up being what you produce.
773
00:48:17,537 --> 00:48:20,137
So it's almost, it's where you get into this self-fulfilling prophecy.
774
00:48:20,817 --> 00:48:24,637
And this is why, you know, healthy habits and meditation
775
00:48:24,637 --> 00:48:27,717
and I mean, talk about getting out of AI psychosis, you know,
776
00:48:27,717 --> 00:48:44,057
Like you got to step back and clear your mind, clear your emotions, especially now more than ever, because especially now we're in a world where information and perspective is is the battleground.
777
00:48:44,717 --> 00:48:48,617
Like I hate, you know, and maybe it's been like this for a while, but it's even more so.
778
00:48:48,797 --> 00:48:49,757
And it's not even us.
779
00:48:49,797 --> 00:48:50,557
It's the agents.
780
00:48:50,757 --> 00:48:51,197
Right.
781
00:48:51,337 --> 00:48:54,417
I mean, you know, it's not even SEO anymore.
782
00:48:54,477 --> 00:48:55,037
It's AEO.
783
00:48:55,037 --> 00:48:59,897
It's like, if people are going to ChatGPT, how are they talking about your product?
784
00:49:00,497 --> 00:49:14,417
So now we're just living in this massive weighted cloud of information agendas.
785
00:49:15,037 --> 00:49:17,517
And so it's so hard to think clearly.
786
00:49:17,957 --> 00:49:20,137
And I've been saying this last couple of episodes.
787
00:49:20,137 --> 00:49:27,097
I feel like going on certain websites or social platforms is just straight up adversarial.
788
00:49:27,657 --> 00:49:30,857
Like I want to surround myself with what I'm calling the agentic aura,
789
00:49:31,237 --> 00:49:35,137
which is like I want all information filtered through my own local,
790
00:49:35,877 --> 00:49:43,317
to use your term, fiduciary agents that are like for me to help me have more clarity
791
00:49:43,317 --> 00:49:48,557
and help me have like happier state of mind, more relaxed state of mind,
792
00:49:48,557 --> 00:49:49,597
because then I'm more powerful.
793
00:49:49,597 --> 00:49:53,877
I can achieve more. I'm more efficient. And also, life is just better.
794
00:49:54,977 --> 00:50:02,457
Absolutely. Yeah, I think that's really true. When you log on to the internet or social media,
795
00:50:02,957 --> 00:50:10,437
you're immediately subject to countless psychological operation campaigns, quite literally.
796
00:50:10,437 --> 00:50:29,857
You know, whether it's by intelligence agencies, private actors, individuals, bots, businesses, advertising, tracking, all of it, it's all trying to control your mind, get you to do something, get you to think something.
797
00:50:30,217 --> 00:50:31,437
And it is highly adversarial.
798
00:50:32,137 --> 00:50:37,857
And I think that's one of the big reasons where you get a weird feeling when you're done logging off.
799
00:50:37,857 --> 00:50:40,497
You know, you don't feel good.
800
00:50:40,717 --> 00:50:45,717
You kind of feel on some level that your mind just got, like, extracted or manipulated,
801
00:50:46,237 --> 00:50:48,117
and you don't feel whole.
802
00:50:48,257 --> 00:50:52,317
You don't feel connected to your humanity and other people's humanity.
803
00:50:52,877 --> 00:51:01,217
A lot of times people feel depressed or in a bad mood or, you know, all of these things.
804
00:51:01,217 --> 00:51:07,137
I don't think it's new in the sense that for thousands of years,
805
00:51:07,857 --> 00:51:18,197
People have been using narratives and trying to control people through fear, through being able to influence what they believe.
806
00:51:19,397 --> 00:51:36,397
Not trying to get too controversial, but certainly there are some religions which have been exploited, at the very least, to attempt to control and gain power quite successfully.
807
00:51:37,857 --> 00:51:39,277
and authoritarian governments.
808
00:51:40,497 --> 00:51:45,517
I think that what has changed is the ability to craft a message
809
00:51:45,517 --> 00:51:50,317
and deploy it much more effectively in the past.
810
00:51:50,477 --> 00:51:52,917
People are all kind of looking at the same places.
811
00:51:54,697 --> 00:51:59,337
And it's always really important, I think, to look at the incentives.
812
00:52:00,157 --> 00:52:02,637
So for OpenAI and Anthropic,
813
00:52:02,637 --> 00:52:06,797
they have an incentive to get you to use their tools.
814
00:52:06,797 --> 00:52:20,637
So if they tell you that something is governable and everybody seems to actually think this is possible at the moment, it's because they want you to believe that it's governable so you'll use it.
815
00:52:21,277 --> 00:52:25,617
My point of view, it's a black box.
816
00:52:26,317 --> 00:52:28,777
You don't know what its motivation is.
817
00:52:28,777 --> 00:52:37,977
There is no way for any LLM to be governable.
818
00:52:38,177 --> 00:52:41,057
It is total snake oil.
819
00:52:42,117 --> 00:52:49,217
At the same time, if you don't give it power, then it can't do something to you.
820
00:52:49,337 --> 00:52:51,777
I mean, you interact with people all the time.
821
00:52:51,997 --> 00:52:54,997
You don't govern the people you interact with.
822
00:52:54,997 --> 00:53:21,197
I don't know why you expect to be able to govern the LLM, but I'm not saying don't use it, and I'm not saying be scared of it because, you know, you don't give out your social security number and your credit card number to everybody because, you know, you're willing to trust them with a certain amount of power over you.
823
00:53:21,197 --> 00:53:24,237
don't give the LLM power over you.
824
00:53:24,877 --> 00:53:27,097
Certainly, I mean, if anybody out there
825
00:53:27,097 --> 00:53:28,877
is putting an LLM in charge of missiles
826
00:53:28,877 --> 00:53:30,897
or nuclear codes,
827
00:53:31,157 --> 00:53:33,237
I have a big beef with them.
828
00:53:33,937 --> 00:53:36,217
But the fault is not the LLM.
829
00:53:36,337 --> 00:53:38,077
I used to have a boss who liked to say,
830
00:53:38,657 --> 00:53:40,757
if the king gives a sword to a monkey
831
00:53:40,757 --> 00:53:41,897
and the monkey stabs him,
832
00:53:41,997 --> 00:53:43,617
it's not the monkey's fault.
833
00:53:47,617 --> 00:53:49,117
And so, anyway,
834
00:53:49,117 --> 00:53:52,757
We have extreme people who are saying,
835
00:53:53,297 --> 00:53:56,177
Oh, this AI is great. Trust everything.
836
00:53:56,937 --> 00:53:58,377
And people are overtrusting.
837
00:53:59,517 --> 00:54:01,157
We have other people who are saying,
838
00:54:01,597 --> 00:54:03,677
AI is terrible. Don't use it at all.
839
00:54:03,857 --> 00:54:05,677
Take it completely out of your life.
840
00:54:07,637 --> 00:54:09,557
They have incentives too.
841
00:54:10,157 --> 00:54:12,457
There are also people out there who are undertrusting.
842
00:54:13,457 --> 00:54:16,397
And, you know, years worth of media
843
00:54:16,397 --> 00:54:20,097
that focuses more on evil AI like the Terminator
844
00:54:20,097 --> 00:54:22,757
rather than good AI like Data from Star Trek
845
00:54:22,757 --> 00:54:25,497
tends to bias people.
846
00:54:26,537 --> 00:54:28,497
And I've always thought that
847
00:54:28,497 --> 00:54:31,797
it's all in how you use it.
848
00:54:33,117 --> 00:54:37,097
But pay attention to, you know,
849
00:54:37,837 --> 00:54:39,637
if you see an extreme narrative,
850
00:54:39,637 --> 00:54:43,217
then somebody's got an extreme incentive.
851
00:54:44,777 --> 00:54:45,517
And...
852
00:54:46,397 --> 00:54:47,597
I don't know.
853
00:54:48,317 --> 00:54:50,077
Fear mongering on...
854
00:54:50,077 --> 00:54:53,297
Fear mongering is, you know, a thing on one side,
855
00:54:53,397 --> 00:54:56,837
but the seductive promise of new technology
856
00:54:56,837 --> 00:54:59,457
is kind of the opposite to me, right?
857
00:54:59,697 --> 00:55:01,137
Where everybody says,
858
00:55:01,297 --> 00:55:03,097
I'm going to use LLMs for everything.
859
00:55:03,937 --> 00:55:08,137
I'm going to write my emails to my friends with LLMs.
860
00:55:08,357 --> 00:55:10,437
I mean, you don't need to do that.
861
00:55:12,857 --> 00:55:14,697
They might stop being your friends.
862
00:55:14,697 --> 00:55:18,197
Do you really want to save time talking to your friends?
863
00:55:18,377 --> 00:55:18,897
I don't know.
864
00:55:21,077 --> 00:55:22,437
Get back to vibe coding.
865
00:55:22,817 --> 00:55:23,977
That's right.
866
00:55:25,317 --> 00:55:27,177
Which is disappointing in some ways
867
00:55:27,177 --> 00:55:29,937
because we love working with these really
868
00:55:29,937 --> 00:55:32,937
explicit, non-ambiguous languages, right?
869
00:55:34,057 --> 00:55:35,017
I say,
870
00:55:36,657 --> 00:55:39,377
and I don't think that I'll work with Claude Cody as vibe coding,
871
00:55:39,957 --> 00:55:43,377
but I am disappointed
872
00:55:43,377 --> 00:55:45,417
that in order to be efficient,
873
00:55:45,577 --> 00:55:47,837
I have to use it these days in some ways,
874
00:55:48,257 --> 00:55:51,797
but I'm also elated that I am incredibly productive, right?
875
00:55:51,797 --> 00:55:52,437
Oh, yeah.
876
00:55:53,577 --> 00:55:57,797
And, yeah, there are downsides, but...
877
00:56:00,437 --> 00:56:03,077
Yeah, I mean, it's...
878
00:56:03,077 --> 00:56:06,797
I just worry about, like, how...
879
00:56:06,797 --> 00:56:10,337
When people talk about, oh, we, you know,
880
00:56:10,637 --> 00:56:12,717
we're worried about not being able to govern the AI,
881
00:56:12,717 --> 00:56:19,317
that in some ways that's a cipher for we're worried about not being able to govern the individual.
882
00:56:20,757 --> 00:56:24,677
You know, that they're actually want to control humans.
883
00:56:25,297 --> 00:56:32,117
And if we give these AIs to humans and we can't control the AI, then we can't control the humans.
884
00:56:32,817 --> 00:56:38,397
And so we need to have these guardrails and it needs to be centralized.
885
00:56:38,397 --> 00:56:42,337
and only the big tech companies can get mythos
886
00:56:42,337 --> 00:56:45,737
and the plebs, we couldn't give it to the plebs
887
00:56:45,737 --> 00:56:47,956
because imagine what they would do.
888
00:56:48,637 --> 00:56:52,137
And behind all of this is this institutional bias,
889
00:56:52,297 --> 00:56:55,057
whether it's the government needs to take over
890
00:56:55,057 --> 00:56:57,777
or the big tech companies need to take over.
891
00:56:58,317 --> 00:56:59,897
But none of these voices are saying,
892
00:57:00,037 --> 00:57:02,337
actually what we need to do is push the power
893
00:57:02,337 --> 00:57:06,077
to the individual and the institutions are the danger.
894
00:57:06,597 --> 00:57:07,977
Centralization is the danger.
895
00:57:08,397 --> 00:57:11,377
That is a perspective that is underrepresented.
896
00:57:12,017 --> 00:57:14,217
You know, I saw this new AI doc,
897
00:57:14,277 --> 00:57:15,577
which is a really good documentary.
898
00:57:15,777 --> 00:57:16,917
I recommend seeing it.
899
00:57:17,277 --> 00:57:19,417
And they had this very wide range of perspectives
900
00:57:19,417 --> 00:57:21,317
from the extreme pessimists,
901
00:57:21,417 --> 00:57:23,537
the extreme optimists, people in the middle.
902
00:57:23,936 --> 00:57:27,177
But nobody had the viewpoint of
903
00:57:27,177 --> 00:57:30,177
the danger is actually centralization
904
00:57:30,177 --> 00:57:32,237
and we have to push this to the individual.
905
00:57:32,417 --> 00:57:33,577
We have to decentralize it.
906
00:57:33,677 --> 00:57:34,956
That wasn't represented at all,
907
00:57:35,037 --> 00:57:36,677
which I thought was very surprising.
908
00:57:38,397 --> 00:57:42,477
The people with the money, their viewpoints are the ones that are getting heard, right?
909
00:57:42,537 --> 00:57:48,137
The danger of centralization, I think, is kind of the concern of the little guy who isn't well organized enough.
910
00:57:48,917 --> 00:57:54,277
And it doesn't fit, you know, the contrary viewpoints of any of the people with the big money.
911
00:57:54,357 --> 00:57:56,517
So they just don't talk about it.
912
00:57:57,517 --> 00:57:59,837
Or they make it seem really dangerous.
913
00:57:59,837 --> 00:58:06,237
They're like, that's the last thing we would ever want to do, you know, is let the individual get access to this.
914
00:58:06,237 --> 00:58:19,237
And it kind of reminds me of maybe what, or makes me think about maybe what it would have felt like when the Gutenberg printing press was first invented and all the translations of the Bible were in Latin.
915
00:58:20,417 --> 00:58:26,437
And, you know, Tinsdale had one of the first translations of the Bible in English.
916
00:58:27,717 --> 00:58:34,037
And that is still one of the most influential works in the English language.
917
00:58:34,037 --> 00:58:36,797
but he was burned at the stake for that.
918
00:58:37,877 --> 00:58:40,917
And, you know, a lot of the arguments for the church was
919
00:58:40,917 --> 00:58:45,937
only the elite should be able to have direct access to this
920
00:58:45,937 --> 00:58:50,357
and it's harmful for the average person to interpret it themselves.
921
00:58:50,357 --> 00:58:51,956
They're going to get all these bad ideas.
922
00:58:52,157 --> 00:58:53,637
The devil's going to get in their head
923
00:58:53,637 --> 00:58:57,317
and it should all be mediated through the church
924
00:58:57,317 --> 00:59:02,017
to kind of deliver the safe, proper interpretation of the text.
925
00:59:02,017 --> 00:59:09,257
and in a lot of the debates today seem kind of like the same dynamic just dressed up in new
926
00:59:09,257 --> 00:59:15,477
clothing you know what i feel like this is we're not giving humans enough credit i i remember kind
927
00:59:15,477 --> 00:59:20,117
of from a very early very early age when you start to get a bigger perspective of the world
928
00:59:20,117 --> 00:59:30,717
i just felt like it's so amazing how almost everyone on the planet like gets along and i
929
00:59:30,717 --> 00:59:33,417
obviously that might sound very naive.
930
00:59:33,936 --> 00:59:35,337
Of course, there's all these conflicts.
931
00:59:36,017 --> 00:59:37,917
In some ways, peace is very fragile.
932
00:59:38,917 --> 00:59:41,057
But at the same time,
933
00:59:42,097 --> 00:59:44,997
it's incredible what we have on the planet.
934
00:59:44,997 --> 00:59:48,837
The fact that you walk into a grocery store
935
00:59:48,837 --> 00:59:51,637
and you see the diversity of high-quality goods
936
00:59:51,637 --> 00:59:54,197
and you walk into other...
937
00:59:54,197 --> 00:59:57,497
In some ways, there's a lot of improvements to make.
938
00:59:57,497 --> 01:00:05,317
But this, I don't think people give the individual enough credit on the whole across all of humanity.
939
01:00:06,397 --> 01:00:09,177
And I would actually go a little further.
940
01:00:09,357 --> 01:00:14,697
I would say that at least 99% of conflict is driven by centralized power dynamics.
941
01:00:15,257 --> 01:00:20,417
Even if you go down to people who are arguing about their sports team, right?
942
01:00:20,417 --> 01:00:27,956
There's a centralized power behind each of the different teams that is motivating them.
943
01:00:28,917 --> 01:00:33,297
The things that are motivating them are not their problems themselves.
944
01:00:33,517 --> 01:00:41,237
I mean, occasionally people have problems with their neighbors, you know, fences being in the wrong place or something like that is the archetypal example.
945
01:00:42,097 --> 01:00:43,937
But they don't go to war.
946
01:00:44,097 --> 01:00:46,877
The only people who go to war are governments.
947
01:00:46,877 --> 01:00:49,957
and it's not, you know,
948
01:00:50,337 --> 01:00:54,057
all the major source of conflict in the world is
949
01:00:54,057 --> 01:00:56,397
I don't have enough power.
950
01:00:56,677 --> 01:00:59,317
I want the power that guy has as well.
951
01:01:00,277 --> 01:01:03,017
And, you know, yeah, I agree.
952
01:01:03,857 --> 01:01:05,877
Individuals are not the source of the problem.
953
01:01:05,997 --> 01:01:06,917
They never have been.
954
01:01:06,997 --> 01:01:09,437
It's the people who, you know,
955
01:01:09,797 --> 01:01:10,857
try to get all,
956
01:01:10,977 --> 01:01:13,257
they try to suck up all the power from the individuals
957
01:01:13,257 --> 01:01:17,717
so that they can be the vast determiners.
958
01:01:18,817 --> 01:01:20,217
And just one more thing there is,
959
01:01:20,497 --> 01:01:23,137
if you've looked into Bitcoin,
960
01:01:23,416 --> 01:01:26,097
Bitcoin has a very long-term view, right?
961
01:01:26,177 --> 01:01:28,237
I mean, literally, in the white paper,
962
01:01:28,357 --> 01:01:32,397
it says the year 2140 is when the last thing will be mined.
963
01:01:32,517 --> 01:01:35,097
I mean, to build something and then have it,
964
01:01:35,217 --> 01:01:38,097
you know, you've set a date that's like 150 years
965
01:01:38,097 --> 01:01:40,997
or like 120, 30 years in the future.
966
01:01:40,997 --> 01:01:47,097
he's like wild long-term thinking there's an organization here in san francisco they it's a
967
01:01:47,097 --> 01:01:52,097
really cool bar if you want to go to the interval but it's run by the long now foundation and i don't
968
01:01:52,097 --> 01:01:54,857
know too much about them but i've been there a couple it's really cool spot if you haven't been
969
01:01:54,857 --> 01:02:03,337
there but they their whole thing is how do we get humanity to think on such long time scales like
970
01:02:03,337 --> 01:02:11,357
the next thousand years, 10,000 years. And I just see if you want robust systems, decentralized
971
01:02:11,357 --> 01:02:16,617
is so much better. And, you know, we mentioned this before, but like we take for granted
972
01:02:16,617 --> 01:02:23,037
all the decentralized systems in our life, right? Like you're, you're like literally local and state
973
01:02:23,037 --> 01:02:27,877
governments as a form of decentralization, your family and your house. Like we, it's not like
974
01:02:27,877 --> 01:02:33,537
everyone in the whole country lives in one building, right? It's very decentralized. And so
975
01:02:33,537 --> 01:02:42,317
all of that makes for very resilient systems. And so I think if we want to build a future that is
976
01:02:42,317 --> 01:02:48,197
better for everyone, that works for more people, just decentralization in general is the goal.
977
01:02:49,457 --> 01:02:55,637
Yeah, I totally agree. And I love the point you made about all the human goodness that's happening
978
01:02:55,637 --> 01:02:57,717
every single day, every single moment
979
01:02:57,717 --> 01:02:59,956
when you think about all the
980
01:02:59,956 --> 01:03:01,997
cooperation that's
981
01:03:01,997 --> 01:03:03,956
happening all the time that is
982
01:03:03,956 --> 01:03:06,097
making the world go round. It's all invisible
983
01:03:06,097 --> 01:03:07,657
no one talks about it. I mean it's like
984
01:03:07,657 --> 01:03:09,877
this would not exist. We couldn't be here
985
01:03:09,877 --> 01:03:11,797
we wouldn't have to say like the amount of
986
01:03:11,797 --> 01:03:13,697
cooperation that happens in
987
01:03:13,697 --> 01:03:15,737
the oh we just lost the light I think
988
01:03:15,737 --> 01:03:17,917
it got a little dark here as
989
01:03:17,917 --> 01:03:19,817
we're talking about positive. I know we're trying to
990
01:03:19,817 --> 01:03:20,877
raise the light but
991
01:03:20,877 --> 01:03:23,657
you know it's just
992
01:03:23,657 --> 01:03:26,617
I feel like it's almost like a miracle.
993
01:03:28,117 --> 01:03:34,137
I think it's something that it's something obviously is going on
994
01:03:34,137 --> 01:03:40,337
where 99.9% of people and individuals are good
995
01:03:40,337 --> 01:03:45,877
and they want outcomes that work for not only themselves
996
01:03:45,877 --> 01:03:47,657
but other people around them.
997
01:03:48,557 --> 01:03:51,877
And I think if that wasn't the case,
998
01:03:51,877 --> 01:03:54,436
maybe none of this would exist.
999
01:03:55,057 --> 01:03:55,857
We wouldn't even be here.
1000
01:03:55,956 --> 01:03:56,757
There's no way.
1001
01:03:57,637 --> 01:04:00,137
And then, I mean, how many people
1002
01:04:00,137 --> 01:04:01,717
not only are cooperating,
1003
01:04:02,117 --> 01:04:04,837
how many people are sacrificing
1004
01:04:04,837 --> 01:04:07,517
for somebody else every day,
1005
01:04:07,977 --> 01:04:08,977
for somebody they love,
1006
01:04:09,077 --> 01:04:10,757
somebody they care about, right?
1007
01:04:10,837 --> 01:04:12,517
That's happening all the time.
1008
01:04:12,877 --> 01:04:15,677
But that's not what we hear about in the news.
1009
01:04:16,657 --> 01:04:20,797
And people get so cynical about human nature.
1010
01:04:20,797 --> 01:04:22,517
humans are bad
1011
01:04:22,517 --> 01:04:25,017
and a lot of it I think comes from
1012
01:04:25,017 --> 01:04:26,697
looking at an
1013
01:04:26,697 --> 01:04:28,717
individual's inability
1014
01:04:28,717 --> 01:04:30,777
to solve something that they have
1015
01:04:30,777 --> 01:04:32,877
no ability to solve on their
1016
01:04:32,877 --> 01:04:34,977
own so people will say you know
1017
01:04:34,977 --> 01:04:36,777
no one cares about climate
1018
01:04:36,777 --> 01:04:38,777
change so people are bad
1019
01:04:38,777 --> 01:04:40,577
or you know racism
1020
01:04:40,577 --> 01:04:42,897
exists so you know you're
1021
01:04:42,897 --> 01:04:44,637
a bad person because you haven't solved
1022
01:04:44,637 --> 01:04:46,977
racism or you haven't solved climate
1023
01:04:46,977 --> 01:04:48,456
change or you're not doing enough
1024
01:04:48,456 --> 01:04:50,357
to battle these like
1025
01:04:50,797 --> 01:04:58,896
incredibly complicated problems that nobody has a great answer for. And there's no easy solutions.
1026
01:04:59,197 --> 01:05:02,777
And meanwhile, you know, somebody has got their hands full just getting through the day
1027
01:05:02,777 --> 01:05:08,896
and meeting their immediate responsibilities. But then we blame that individual somehow for
1028
01:05:08,896 --> 01:05:15,497
climate change or for, you know, all the social problems. It's this bad individual doesn't care
1029
01:05:15,497 --> 01:05:21,677
enough isn't doing enough. And that also reminds me of tactics that have been used in religion,
1030
01:05:21,677 --> 01:05:30,237
where it's like, you're bad for being human. You're sinful, you're bad. And that is why you
1031
01:05:30,237 --> 01:05:36,697
have to give me the power over you, because you're no good. And it's a similar thing, you know,
1032
01:05:36,857 --> 01:05:45,057
with the with this rhetoric about humanity is bad, the individual is bad. And therefore,
1033
01:05:45,057 --> 01:05:49,497
we need to take that power from the individual because they're not using it correctly.
1034
01:05:50,416 --> 01:05:55,097
When so often the opposite is true, the individual is fine.
1035
01:05:55,857 --> 01:06:00,617
Like you were saying, it's these institutions or these large organizations or these groups
1036
01:06:00,617 --> 01:06:02,677
where things start to break down.
1037
01:06:03,376 --> 01:06:08,237
And even then, you can't even necessarily blame the individual because there are all
1038
01:06:08,237 --> 01:06:12,617
these peoples in a corporation or in a government and there's incentives and there's structure
1039
01:06:12,617 --> 01:06:16,637
and there's no one person that you can point to and say it's their fault.
1040
01:06:16,797 --> 01:06:20,117
They're the bad guy. They made the wrong decision.
1041
01:06:21,157 --> 01:06:23,817
A lot of it is just about system structure.
1042
01:06:25,117 --> 01:06:28,877
But we love to find that bad guy.
1043
01:06:29,436 --> 01:06:31,537
We want the villain. We want somebody to blame.
1044
01:06:31,617 --> 01:06:33,197
We want to say it was their fault.
1045
01:06:33,197 --> 01:06:35,777
They did the wrong thing for the wrong reason.
1046
01:06:35,777 --> 01:06:47,337
And the more nuanced perspective where it's kind of more about oftentimes the difficulty of the underlying problem.
1047
01:06:47,337 --> 01:06:55,597
like i think a lot of times people think that you know there's somebody who you know when it comes to
1048
01:06:55,597 --> 01:07:02,057
income inequality or climate change or any of these big issues ai you know social media there's
1049
01:07:02,057 --> 01:07:09,357
somebody who is has access to some lever where they could just make everything good and they're
1050
01:07:09,357 --> 01:07:15,537
just choosing not to pull the lever but like i think usually when you like kind of unpeel what's
1051
01:07:15,537 --> 01:07:17,537
going on, there's these crazy
1052
01:07:17,537 --> 01:07:19,356
trade-offs and risks and
1053
01:07:19,356 --> 01:07:21,217
uncertainties. And it's like, this lever
1054
01:07:21,217 --> 01:07:23,577
has these risks and this one has these risks.
1055
01:07:23,916 --> 01:07:25,477
And it's not because someone is like
1056
01:07:25,477 --> 01:07:27,497
some evil person saying, I'm
1057
01:07:27,497 --> 01:07:29,376
going to pull the give me all the money lever
1058
01:07:29,376 --> 01:07:30,376
and kill everybody.
1059
01:07:31,157 --> 01:07:33,497
It's like, they're kind of
1060
01:07:33,497 --> 01:07:35,416
overmatched by the
1061
01:07:35,416 --> 01:07:36,997
difficulty and complexity of the problem.
1062
01:07:37,497 --> 01:07:39,317
And I'm not saying that's always the case,
1063
01:07:40,037 --> 01:07:41,557
but it's kind of like that saying
1064
01:07:41,557 --> 01:07:43,037
where it's like, never attribute
1065
01:07:43,037 --> 01:07:45,137
malice
1066
01:07:45,137 --> 01:07:47,697
to what you could attribute ignorance to.
1067
01:07:47,857 --> 01:07:48,517
Or incompetence.
1068
01:07:48,717 --> 01:07:49,277
Incompetence.
1069
01:07:49,377 --> 01:07:51,137
But I think there's like an even further one,
1070
01:07:51,197 --> 01:07:53,177
which is never attribute incompetence
1071
01:07:53,177 --> 01:07:55,057
to when you can attribute the difficulty
1072
01:07:55,057 --> 01:07:56,757
of the underlying problem.
1073
01:07:57,597 --> 01:08:00,797
Well, so this makes me almost more hopeful than ever
1074
01:08:00,797 --> 01:08:03,057
and a reason to be optimistic about the future
1075
01:08:03,057 --> 01:08:07,257
because if it is difficult problems
1076
01:08:07,257 --> 01:08:09,117
and building systems,
1077
01:08:09,677 --> 01:08:13,417
then we are unlocking that with AI.
1078
01:08:14,237 --> 01:08:14,597
I mean...
1079
01:08:15,137 --> 01:08:16,436
that's just a perspective.
1080
01:08:16,577 --> 01:08:17,856
Like if we were all more intelligent,
1081
01:08:18,617 --> 01:08:21,856
maybe we get way better outcomes in the future that,
1082
01:08:22,277 --> 01:08:22,417
but,
1083
01:08:22,477 --> 01:08:23,077
but again,
1084
01:08:23,177 --> 01:08:23,657
no one,
1085
01:08:23,777 --> 01:08:25,977
no one seems to be talking in this way.
1086
01:08:26,157 --> 01:08:27,396
And so they don't,
1087
01:08:28,157 --> 01:08:28,837
it's,
1088
01:08:28,936 --> 01:08:30,257
it's like you can't even,
1089
01:08:30,377 --> 01:08:32,697
you haven't even identified the problem that you should be focused on
1090
01:08:32,697 --> 01:08:33,197
solving.
1091
01:08:33,737 --> 01:08:34,337
And so,
1092
01:08:34,356 --> 01:08:37,317
and then if you don't even know the problem or that cause of the problem,
1093
01:08:37,317 --> 01:08:41,217
you can't even see what it might look like if things were different.
1094
01:08:43,257 --> 01:08:44,057
I don't know.
1095
01:08:44,157 --> 01:08:44,377
Interesting.
1096
01:08:45,137 --> 01:08:48,757
If this is what you're talking about when you talk about the centralization problem,
1097
01:08:48,757 --> 01:09:01,657
but I do wonder if simple power distribution is one of the major problems with humanity
1098
01:09:01,657 --> 01:09:10,017
and that fundamentally, if power is distributed as evenly as possible,
1099
01:09:10,017 --> 01:09:18,277
than, you know, we have some theoretical minimum.
1100
01:09:18,777 --> 01:09:19,717
I don't think it quite works
1101
01:09:19,717 --> 01:09:22,977
because I think I'm describing anarchy at the end.
1102
01:09:22,977 --> 01:09:25,817
And not that anarchy is necessarily bad,
1103
01:09:26,017 --> 01:09:33,177
but certainly there is still violence that happens in anarchy.
1104
01:09:35,057 --> 01:09:38,977
But it feels like there is...
1105
01:09:40,017 --> 01:09:54,497
Maybe the technology could be a force for even power redistribution,
1106
01:09:55,177 --> 01:10:00,837
which is fundamentally, though, at odds with any governmental structure,
1107
01:10:01,497 --> 01:10:06,017
which is, you know, fundamentally about centralized power.
1108
01:10:06,877 --> 01:10:08,417
Yeah, I think there's really something to that.
1109
01:10:08,417 --> 01:10:10,697
And, you know, there's also a spectrum, right?
1110
01:10:10,757 --> 01:10:20,297
Like, even within a government, like, let's say the U.S. government, like, having the branches of government is a way of trying to distribute power.
1111
01:10:21,017 --> 01:10:27,517
And in general, there's this problem of power, which is power corrupts, absolute power corrupts absolutely.
1112
01:10:27,517 --> 01:10:35,817
and that doesn't mean that the person who's corrupted by it is a bad person because anyone
1113
01:10:35,817 --> 01:10:43,417
who has the weight of that power is probably not going to be able to bear it the um you know i'm
1114
01:10:43,417 --> 01:10:48,557
kind of a uh lord of the rings nerd do you guys know you know gandalf and that whole thing i mean
1115
01:10:48,557 --> 01:10:54,257
i'm a fan i haven't gone too deep i won't don't quiz me on it but yeah this is just an example i
1116
01:10:54,257 --> 01:10:55,877
come back to a lot because I think it's
1117
01:10:55,877 --> 01:10:58,197
a good analogy for this, which
1118
01:10:58,197 --> 01:10:59,997
is, you know,
1119
01:11:00,197 --> 01:11:02,217
Gandalf, who is sort
1120
01:11:02,217 --> 01:11:04,277
of one of the
1121
01:11:04,277 --> 01:11:06,317
beacons of integrity in the story.
1122
01:11:06,477 --> 01:11:08,197
You know, he's like one of the good guys. He's
1123
01:11:08,197 --> 01:11:10,317
always like ethical, whatever. But
1124
01:11:10,317 --> 01:11:12,037
when he puts the ring on,
1125
01:11:12,717 --> 01:11:14,097
like in The Fellowship of the Ring,
1126
01:11:14,177 --> 01:11:16,137
that scene at the beginning or in the books where he's
1127
01:11:16,137 --> 01:11:18,257
like sitting near the fireplace, he's putting the ring on.
1128
01:11:18,337 --> 01:11:20,137
He starts to get corrupted
1129
01:11:20,137 --> 01:11:31,575
by it And he needs to get rid of it You mean Frodo You mean Bilbo Well Gandalf there a moment where he kind of like
1130
01:11:31,775 --> 01:11:33,755
they don't realize that the ring's there.
1131
01:11:34,155 --> 01:11:34,975
They don't know what it is yet.
1132
01:11:35,015 --> 01:11:36,115
They don't know what it is yet.
1133
01:11:36,275 --> 01:11:38,135
And he's kind of like in this weird daze,
1134
01:11:38,295 --> 01:11:39,575
like staring into the fire
1135
01:11:39,575 --> 01:11:40,876
and like, you know,
1136
01:11:40,876 --> 01:11:43,015
almost has to push himself, you know?
1137
01:11:43,035 --> 01:11:44,555
Or that also happens to Galadriel
1138
01:11:44,555 --> 01:11:45,955
where she like puts it on.
1139
01:11:46,035 --> 01:11:46,995
She's like, oh, you know,
1140
01:11:47,055 --> 01:11:48,435
what I would do and whatever.
1141
01:11:48,555 --> 01:11:50,195
And she has to like stop herself
1142
01:11:50,195 --> 01:11:55,035
where it's like there's almost a power where if you were given enough power,
1143
01:11:55,035 --> 01:11:59,455
no matter how good you are, it would mess you up.
1144
01:12:00,095 --> 01:12:06,155
And so I partly think of that in the context of checks and balances
1145
01:12:06,155 --> 01:12:10,315
or even something like Bitcoin where it's like the power to create money
1146
01:12:10,315 --> 01:12:14,695
is almost too much for the government to wield responsibly.
1147
01:12:14,695 --> 01:12:19,815
And we almost have to take that power away, not because the government's bad,
1148
01:12:20,195 --> 01:12:24,495
but because that restores it to its goodness by not having the corrupting power.
1149
01:12:25,355 --> 01:12:32,155
So that's another way of thinking about distributing power or decentralizing power or balance of power
1150
01:12:32,155 --> 01:12:44,275
is that if we distribute it enough, then each person can kind of be responsible for the amount of power that they're given.
1151
01:12:44,275 --> 01:12:50,895
But if we concentrate it too much, it's very difficult for that to be wielded responsibly.
1152
01:12:51,895 --> 01:13:03,395
And that's sort of the ideal of democracy, which is never really realized, right, is that each citizen is really responsible for what's going on.
1153
01:13:03,395 --> 01:13:06,735
and actually it makes a better argument
1154
01:13:06,735 --> 01:13:09,015
for the idea that everyone should vote
1155
01:13:09,015 --> 01:13:12,175
than I've heard in the past
1156
01:13:12,175 --> 01:13:13,155
because I've tended to think
1157
01:13:13,155 --> 01:13:13,975
that if you don't want to vote,
1158
01:13:14,055 --> 01:13:15,655
you probably shouldn't.
1159
01:13:16,376 --> 01:13:18,235
But if, you know,
1160
01:13:18,695 --> 01:13:19,935
the more people who vote,
1161
01:13:19,935 --> 01:13:20,715
the more,
1162
01:13:21,475 --> 01:13:24,215
the less concentrated the power is,
1163
01:13:24,335 --> 01:13:25,455
maybe, you know,
1164
01:13:25,555 --> 01:13:28,715
that's really good
1165
01:13:28,715 --> 01:13:30,215
if everybody could take responsibility.
1166
01:13:31,635 --> 01:13:32,035
And then I,
1167
01:13:32,195 --> 01:13:32,895
what I,
1168
01:13:33,395 --> 01:13:39,315
have cared a lot about voting and I, you know, I, I think not enough when people see outcomes of
1169
01:13:39,315 --> 01:13:44,615
elections or other political things where they, but they, they personally are, are not necessarily
1170
01:13:44,615 --> 01:13:49,115
happy or a group isn't happy. You know, I'm looking at like, okay, who all voted? Like,
1171
01:13:49,135 --> 01:13:52,635
let's get at the, what, you know, what was the process? And if you look at the voting numbers,
1172
01:13:52,755 --> 01:14:01,675
they're so low. And I wonder again, like, I think it can be easy to maybe blame different groups
1173
01:14:01,675 --> 01:14:03,015
or the individuals themselves.
1174
01:14:03,195 --> 01:14:07,475
But what if fundamentally it is a complexity
1175
01:14:07,475 --> 01:14:08,735
and information issue?
1176
01:14:09,175 --> 01:14:11,895
I mean, I never really felt, you know,
1177
01:14:11,995 --> 01:14:17,115
I was not in politics or even in any kind of,
1178
01:14:17,535 --> 01:14:19,035
I didn't really study that a lot in school
1179
01:14:19,035 --> 01:14:20,655
or like read a lot of that in my free time.
1180
01:14:20,655 --> 01:14:24,555
I never felt fully up to speed on making a good decision
1181
01:14:24,555 --> 01:14:27,595
or even what the, you know, looking at the evidence.
1182
01:14:27,815 --> 01:14:29,475
You can always easily hear what people say
1183
01:14:29,475 --> 01:14:30,595
if you quickly look them up,
1184
01:14:30,595 --> 01:14:32,535
but then to see, okay, well, what have they done in the past?
1185
01:14:32,615 --> 01:14:33,595
You know, so much information.
1186
01:14:33,975 --> 01:14:39,295
So that's, again, why, like, if we all level up because of AI,
1187
01:14:39,835 --> 01:14:42,395
then, and we implement mechanisms like, you know,
1188
01:14:42,855 --> 01:14:45,215
ranked choice voting looks very interesting, right,
1189
01:14:45,255 --> 01:14:50,755
as a mechanism to make it so that candidates aren't so against each other
1190
01:14:50,755 --> 01:14:54,895
because they kind of share votes at the end, the way that it works.
1191
01:14:55,315 --> 01:15:00,055
So if we all level up with more, we are literally more intelligent,
1192
01:15:00,055 --> 01:15:02,355
and we can process more information more quickly,
1193
01:15:02,975 --> 01:15:05,475
then, and we have better cryptography everywhere
1194
01:15:05,475 --> 01:15:08,555
that makes it secure and we can prove who we are
1195
01:15:08,555 --> 01:15:10,095
without maybe revealing who we are,
1196
01:15:10,555 --> 01:15:14,355
then, you know, like, why can't we digitally vote?
1197
01:15:14,995 --> 01:15:17,835
And I can chat with an AI to say, like,
1198
01:15:17,895 --> 01:15:20,376
I want the candidates that most align with my values,
1199
01:15:20,475 --> 01:15:21,675
what I personally value.
1200
01:15:22,876 --> 01:15:25,695
Like, that maybe could fix a lot of things.
1201
01:15:25,795 --> 01:15:28,955
But again, that's another potential promise
1202
01:15:28,955 --> 01:15:30,495
of everyone having AI.
1203
01:15:31,255 --> 01:15:31,455
Right.
1204
01:15:31,755 --> 01:15:34,615
And having their own AI
1205
01:15:34,615 --> 01:15:36,835
that is fiduciary,
1206
01:15:37,135 --> 01:15:38,055
that is transparent,
1207
01:15:38,235 --> 01:15:39,155
that is verifiable.
1208
01:15:39,815 --> 01:15:40,415
Because otherwise,
1209
01:15:41,255 --> 01:15:41,415
you know,
1210
01:15:41,435 --> 01:15:41,715
because like,
1211
01:15:41,876 --> 01:15:42,195
you know,
1212
01:15:42,255 --> 01:15:43,235
politicians always,
1213
01:15:43,815 --> 01:15:44,035
you know,
1214
01:15:44,075 --> 01:15:45,315
a lot of politicians lie
1215
01:15:45,315 --> 01:15:46,475
or propaganda or whatever.
1216
01:15:46,635 --> 01:15:47,555
So I would want an AI
1217
01:15:47,555 --> 01:15:48,595
that can kind of like
1218
01:15:48,595 --> 01:15:49,915
poke through the BS,
1219
01:15:50,515 --> 01:15:50,855
you know,
1220
01:15:50,855 --> 01:15:51,855
and tell me
1221
01:15:51,855 --> 01:15:53,535
what's really going on,
1222
01:15:53,595 --> 01:15:54,575
look at their voting record,
1223
01:15:54,715 --> 01:15:56,115
look at who they're getting money from,
1224
01:15:56,935 --> 01:15:57,795
compare their rhetoric.
1225
01:15:57,795 --> 01:16:00,415
let's get to the ground truth.
1226
01:16:00,935 --> 01:16:05,575
But if my AI is not allowed to tell me that
1227
01:16:05,575 --> 01:16:08,275
because for the children or whatever,
1228
01:16:08,435 --> 01:16:09,615
it can't tell me the truth,
1229
01:16:10,395 --> 01:16:11,835
then we don't get that benefit.
1230
01:16:12,555 --> 01:16:14,075
This is actually something that Dustin and I
1231
01:16:14,075 --> 01:16:15,435
are working on a bit right now.
1232
01:16:15,515 --> 01:16:18,715
There's a lens system that we proposed recently.
1233
01:16:19,175 --> 01:16:23,395
The idea being that you collect evidence
1234
01:16:23,395 --> 01:16:27,895
using LLM technology,
1235
01:16:28,035 --> 01:16:29,175
which is really good at interpreting
1236
01:16:29,175 --> 01:16:30,175
and finding all that stuff,
1237
01:16:30,235 --> 01:16:31,275
but you combine it
1238
01:16:31,275 --> 01:16:34,535
using something that's not an LLM,
1239
01:16:34,615 --> 01:16:36,075
that is something you own,
1240
01:16:36,135 --> 01:16:38,055
that's on your machine,
1241
01:16:39,135 --> 01:16:41,035
that is using, you know,
1242
01:16:41,435 --> 01:16:43,815
specialized rubrics based on your values.
1243
01:16:44,376 --> 01:16:44,775
Interesting.
1244
01:16:45,495 --> 01:16:48,135
To determine, you know,
1245
01:16:48,135 --> 01:16:48,915
what is,
1246
01:16:49,955 --> 01:16:52,195
what likely you should believe
1247
01:16:52,195 --> 01:16:53,335
about whether this is something
1248
01:16:53,335 --> 01:16:54,215
that's good for you.
1249
01:16:54,415 --> 01:16:55,695
Do you think I'm describing that well?
1250
01:16:55,795 --> 01:16:56,635
Yeah, I think absolutely.
1251
01:16:56,735 --> 01:16:59,995
But it's very much grounded in evidence ranking
1252
01:16:59,995 --> 01:17:01,615
and then how much evidence
1253
01:17:01,615 --> 01:17:03,255
and how well that evidence scores
1254
01:17:03,255 --> 01:17:04,475
against a particular rubric,
1255
01:17:04,615 --> 01:17:06,675
but that you have control over the rubric
1256
01:17:06,675 --> 01:17:08,975
or other people can make rubrics.
1257
01:17:09,815 --> 01:17:11,335
And part of the reason why that will work
1258
01:17:11,335 --> 01:17:15,315
is because while gathering information
1259
01:17:15,315 --> 01:17:19,155
is based on the LLM technology that's centralized,
1260
01:17:19,155 --> 01:17:24,255
the ability to combine the information
1261
01:17:24,255 --> 01:17:28,115
is distributed and localized to the individual.
1262
01:17:29,215 --> 01:17:33,895
Ah. So could you walk through maybe an example
1263
01:17:33,895 --> 01:17:35,795
to just kind of make it more concrete?
1264
01:17:36,435 --> 01:17:40,415
So I think of it in terms of doing something like consumer reports.
1265
01:17:41,095 --> 01:17:44,535
So say you're trying to figure out what car to buy, right?
1266
01:17:44,595 --> 01:17:45,335
Not controversial.
1267
01:17:46,155 --> 01:17:48,355
I mean, I suppose if I were in an autoconference, it would be.
1268
01:17:48,355 --> 01:18:07,655
But anyway, so you need the LLN to go out and find the best information possible about, you know, the miles per gallon, the safety record, the maintenance record of a bunch of car companies.
1269
01:18:08,435 --> 01:18:17,535
Combine the information, find out what the best thing we have is about, you know, this model line, if that model line's been around a while.
1270
01:18:17,535 --> 01:18:26,795
And it's going to try to find sources that are not marketing copy, right?
1271
01:18:27,055 --> 01:18:33,395
That are coming from people, generally, who are consumer watchdogs, stuff like that, right?
1272
01:18:33,715 --> 01:18:39,535
Going to pull that in, and the LLM is going to find and search down all that information.
1273
01:18:40,235 --> 01:18:44,615
And then, on your computer, you're going to access that archive of information.
1274
01:18:44,615 --> 01:18:46,435
and say, okay, I've got all this evidence
1275
01:18:46,435 --> 01:18:48,595
about what these things are.
1276
01:18:48,675 --> 01:18:51,515
Then I am going to apply my rubric.
1277
01:18:51,595 --> 01:18:54,655
So what I am most concerned about is energy efficiency.
1278
01:18:55,395 --> 01:18:57,115
And I'm also going to want to know
1279
01:18:57,115 --> 01:19:00,675
how long my repair times are going to be
1280
01:19:00,675 --> 01:19:01,515
or something like that.
1281
01:19:01,635 --> 01:19:05,255
And so your local evidence-based reasoner
1282
01:19:05,255 --> 01:19:07,376
is going to put together pro and con arguments
1283
01:19:07,376 --> 01:19:10,935
for each of the different vehicles you could consider,
1284
01:19:11,135 --> 01:19:14,115
saying, you know, this is the miles per gallon here.
1285
01:19:14,115 --> 01:19:20,595
This is the, you know, how, what your, your actual cost is.
1286
01:19:20,655 --> 01:19:27,735
Maybe even bring in information about, you know, what the energy, how do we say, what
1287
01:19:27,735 --> 01:19:31,995
type of energy is available to your house, you know, whether it's actually gas or solar
1288
01:19:31,995 --> 01:19:32,735
or something else.
1289
01:19:34,155 --> 01:19:42,075
And, you know, evaluate based on that, make those arguments up and tell you, you know,
1290
01:19:42,075 --> 01:19:46,235
scores based on what you want and your personal information as well as
1291
01:19:46,235 --> 01:19:49,135
information gathered from the web. These
1292
01:19:49,135 --> 01:19:53,335
arguments are fundamentally deductive.
1293
01:19:53,975 --> 01:19:54,635
They are not
1294
01:19:54,635 --> 01:20:01,895
linguistic. They're not using
1295
01:20:01,895 --> 01:20:06,195
the LLM mechanism of, you know, guess what the next
1296
01:20:06,195 --> 01:20:10,175
word is, right? They're not even structured as a
1297
01:20:10,175 --> 01:20:12,395
conversation even necessarily. Yeah, they're not necessarily
1298
01:20:12,395 --> 01:20:14,355
structured as a conversation. They're
1299
01:20:14,355 --> 01:20:16,295
structured as
1300
01:20:16,295 --> 01:20:17,915
a list of pros and cons and
1301
01:20:17,915 --> 01:20:20,155
an evaluative number, for example.
1302
01:20:20,995 --> 01:20:22,235
Then, you know,
1303
01:20:22,275 --> 01:20:24,435
you have something that's very personalized
1304
01:20:24,435 --> 01:20:26,415
that is, you know,
1305
01:20:27,655 --> 01:20:28,395
not bullshit.
1306
01:20:29,675 --> 01:20:30,255
Or at least
1307
01:20:30,255 --> 01:20:31,955
you can configure
1308
01:20:31,955 --> 01:20:34,175
what you trust.
1309
01:20:34,295 --> 01:20:36,255
So imagine you got all this evidence
1310
01:20:36,255 --> 01:20:38,115
from all these sources, but yet
1311
01:20:38,115 --> 01:20:41,715
for whatever reason, you distrust certain sources,
1312
01:20:42,115 --> 01:20:46,095
you could just simply deselect those and see how it might change.
1313
01:20:46,995 --> 01:20:50,335
And that gives you more power, too.
1314
01:20:50,595 --> 01:20:54,135
And so is there an LLM that is used at all for this part of the process?
1315
01:20:54,876 --> 01:20:57,115
No, not for the argumentation part.
1316
01:20:57,355 --> 01:21:03,535
It's only used for searching to find where the information is.
1317
01:21:03,895 --> 01:21:05,895
It'll tell you what the source of it is, right?
1318
01:21:06,875 --> 01:21:10,475
And it'll tell you, you know, what the numbers are and that kind of thing.
1319
01:21:10,495 --> 01:21:11,395
And it could hallucinate.
1320
01:21:11,575 --> 01:21:14,715
And, you know, we should, we'll do things several times to make sure.
1321
01:21:14,915 --> 01:21:23,795
But you can generally rely on the LLM to provide you, you know, concrete data more than you can to provide you an opinion, right?
1322
01:21:24,095 --> 01:21:24,455
I see.
1323
01:21:25,015 --> 01:21:30,275
And, you know, you can also, I mean, maybe you have reviewers out there that you can rely on.
1324
01:21:30,275 --> 01:21:31,715
Maybe you trust consumer reports.
1325
01:21:31,815 --> 01:21:32,135
Fine.
1326
01:21:32,595 --> 01:21:32,715
Right.
1327
01:21:33,155 --> 01:21:35,075
It's still not really individualized to you.
1328
01:21:35,075 --> 01:21:36,735
they have different priorities than you do.
1329
01:21:37,175 --> 01:21:37,275
Right?
1330
01:21:37,575 --> 01:21:37,695
Sure.
1331
01:21:38,355 --> 01:21:40,775
And so this would be a way
1332
01:21:40,775 --> 01:21:42,715
of accumulating that information
1333
01:21:42,715 --> 01:21:45,335
and developing a personalized argument.
1334
01:21:45,735 --> 01:21:47,035
It's not unbiased,
1335
01:21:47,115 --> 01:21:48,135
but it's biased toward you.
1336
01:21:48,675 --> 01:21:48,875
Right?
1337
01:21:49,335 --> 01:21:49,675
Right.
1338
01:21:49,915 --> 01:21:51,295
And the bias is transparent.
1339
01:21:51,855 --> 01:21:52,135
Sure.
1340
01:21:53,095 --> 01:21:54,255
And, yeah.
1341
01:21:54,635 --> 01:21:57,555
So the LLM gets the sources,
1342
01:21:58,295 --> 01:21:58,835
and then,
1343
01:21:59,475 --> 01:22:00,735
but there's some sort of
1344
01:22:00,735 --> 01:22:02,495
qualitative evaluation
1345
01:22:02,495 --> 01:22:05,715
of the content in the source
1346
01:22:05,715 --> 01:22:07,175
or some scoring of it?
1347
01:22:07,715 --> 01:22:10,075
Yeah, well, we want to score the source
1348
01:22:10,075 --> 01:22:13,455
in terms of a particular,
1349
01:22:13,835 --> 01:22:15,635
I mean, it's going to be domain-specific in a way.
1350
01:22:15,775 --> 01:22:18,876
So, like, if you're looking at cars,
1351
01:22:18,995 --> 01:22:21,195
you know, generally more miles per gallon
1352
01:22:21,195 --> 01:22:24,095
is better than less, right?
1353
01:22:24,815 --> 01:22:31,615
And so there is a natural ordering
1354
01:22:31,615 --> 01:22:35,635
to the data that allows you to order that if it is a value.
1355
01:22:36,215 --> 01:22:38,495
If it is, you know, language,
1356
01:22:39,015 --> 01:22:42,655
then you do need to have a way to say,
1357
01:22:43,435 --> 01:22:46,555
this says a specific, you know,
1358
01:22:46,595 --> 01:22:50,935
there is a specific logical representation of this information.
1359
01:22:51,575 --> 01:22:54,555
And I can chain these facts together
1360
01:22:54,555 --> 01:22:56,595
in a deductive reasoning chain
1361
01:22:56,595 --> 01:22:58,955
to come up with a particular conclusion.
1362
01:22:58,955 --> 01:23:04,315
You can do that without the LLM's hidden bias.
1363
01:23:05,935 --> 01:23:06,815
Yeah, that's fascinating.
1364
01:23:07,635 --> 01:23:13,995
So do you use the LLM to structure the data from the unstructured data?
1365
01:23:14,315 --> 01:23:19,095
Yeah, you do have to uncover the structure using an LLM.
1366
01:23:20,575 --> 01:23:23,895
No natural language techniques that I'm aware of
1367
01:23:23,895 --> 01:23:28,295
are strong enough to do that in general domain.
1368
01:23:28,955 --> 01:23:35,575
I see. So you're able to, you know, you take some magazine article or whatever it found.
1369
01:23:36,155 --> 01:23:39,235
And then you can extract out, let's say, hard numbers.
1370
01:23:39,675 --> 01:23:45,155
You can extract out logical arguments that are structured.
1371
01:23:45,155 --> 01:24:06,775
In fact, when individual sentences, if it says, this car has been tested by the Automobile Mechanics Association, and you have a fact that is explicitly associated with that that's being asserted.
1372
01:24:06,775 --> 01:24:18,375
So you say that according to this source, testing was performed here, and they're the people who were providing numbers.
1373
01:24:19,495 --> 01:24:25,715
So that fact, and if I want to construct an argument out of it, I don't necessarily want to use the argument from the article.
1374
01:24:25,815 --> 01:24:31,415
I want to use the facts from the article and restructure the argument.
1375
01:24:31,415 --> 01:24:33,115
because I don't know if I can trust
1376
01:24:33,115 --> 01:24:35,835
the way they've decided to put the information together.
1377
01:24:36,415 --> 01:24:36,775
Right.
1378
01:24:38,055 --> 01:24:39,555
And, you know, I want to check to see
1379
01:24:39,555 --> 01:24:41,855
whether I've got the same facts from a few sources as well.
1380
01:24:41,855 --> 01:24:45,155
And I want to check, you know, in some cases,
1381
01:24:45,275 --> 01:24:47,335
like I don't want to trust Anthropic
1382
01:24:47,335 --> 01:24:51,335
to tell me what's good about Claude.
1383
01:24:51,715 --> 01:24:53,695
What I really want is to look at the source
1384
01:24:53,695 --> 01:24:57,655
and see does it do something that it claims it does, right?
1385
01:24:57,675 --> 01:24:59,155
So I can verify directly.
1386
01:24:59,155 --> 01:25:02,835
Now, it's not open source, so I can't do that.
1387
01:25:03,155 --> 01:25:08,755
And that is part of what makes it fundamentally less trustworthy than an open source model.
1388
01:25:08,975 --> 01:25:13,915
Open source models are fundamentally more transparent, and we should give them less trust.
1389
01:25:14,495 --> 01:25:14,855
Absolutely.
1390
01:25:15,635 --> 01:25:26,155
And is there, do you have like, is there like a symbolic language you use to structure an argument in a way that you can do that kind of like deterministic analysis?
1391
01:25:26,155 --> 01:25:28,455
Yeah, I mean, you can use first order logic for it.
1392
01:25:28,455 --> 01:25:30,255
there are a lot of logics that have been invented
1393
01:25:30,255 --> 01:25:35,335
in the computer science literature,
1394
01:25:36,535 --> 01:25:39,375
one that's particularly popular right now.
1395
01:25:39,475 --> 01:25:40,955
It's called answer set programming.
1396
01:25:41,195 --> 01:25:42,155
So it's a...
1397
01:25:42,155 --> 01:25:43,215
Have you ever heard of Prolog?
1398
01:25:43,475 --> 01:25:44,415
I've heard of it, yeah.
1399
01:25:44,475 --> 01:25:47,555
So Prolog is a declarative means of programming
1400
01:25:47,555 --> 01:25:48,335
through logic.
1401
01:25:49,875 --> 01:25:53,575
Answer set programming is more powerful than it in some ways.
1402
01:25:53,675 --> 01:25:54,955
It allows abductive inference,
1403
01:25:54,955 --> 01:26:00,675
and it allows you to consider, I guess,
1404
01:26:00,775 --> 01:26:04,655
the probabilities of various facts
1405
01:26:04,655 --> 01:26:08,635
and put together an analysis.
1406
01:26:09,155 --> 01:26:11,115
And there's great tools for searching
1407
01:26:11,115 --> 01:26:16,375
for something that satisfies a number of characteristics.
1408
01:26:16,795 --> 01:26:19,075
If you're familiar with the satisfiability problem
1409
01:26:19,075 --> 01:26:21,115
in computer science,
1410
01:26:21,195 --> 01:26:23,495
there's been a lot of headway made.
1411
01:26:23,495 --> 01:26:28,275
And, you know, LLMs don't do this natively.
1412
01:26:28,515 --> 01:26:31,195
They don't solve the satisfiability problem,
1413
01:26:31,315 --> 01:26:32,215
partially because, you know,
1414
01:26:32,255 --> 01:26:35,115
it's a computationally inefficient thing to do,
1415
01:26:35,535 --> 01:26:38,595
but partially because it's not, you know,
1416
01:26:39,876 --> 01:26:45,355
it's easier in some ways to predict the next word,
1417
01:26:45,515 --> 01:26:49,975
and people don't always use logic in their arguments, right?
1418
01:26:49,975 --> 01:26:52,235
They use associations, right?
1419
01:26:52,235 --> 01:27:03,155
You know, you argue by reference to authority and stuff like that rather than putting together the logic.
1420
01:27:03,455 --> 01:27:15,035
I've spent a lot of time working with logical systems, and, you know, I see that as a fundamentally unbiased native way to put together facts, right?
1421
01:27:15,035 --> 01:27:36,675
Yeah, that's fascinating. Yeah. And there's, there's all the logical fallacies, you know, slippery slope or strawman or, you know, all of these things. And I, and I think a large percentage of arguments or supposed arguments are subject to those fallacies that people mount.
1422
01:27:36,675 --> 01:27:54,915
But that is a fascinating way of using the LLM that I had never thought about to kind of take this unstructured data, put it into structure, and then even for reasoning, use this.
1423
01:27:55,455 --> 01:28:04,435
Once it's structured in a way that the reasoning algorithm can understand, then you probably get, in some senses, a lot better reasoning.
1424
01:28:04,435 --> 01:28:12,715
because I think we've all seen LLMs make really ridiculous reasoning mistakes all the time.
1425
01:28:12,935 --> 01:28:16,015
And it'll say, oh, we're going to do this because of X.
1426
01:28:16,055 --> 01:28:18,995
And then you follow up and you're like, well, that doesn't really follow from X.
1427
01:28:19,055 --> 01:28:20,015
And it's like, oh, yeah, you're right.
1428
01:28:20,035 --> 01:28:20,935
It doesn't follow at all.
1429
01:28:21,335 --> 01:28:25,315
And it's like, you know, whereas if you have something like prologue or something like that,
1430
01:28:25,375 --> 01:28:26,355
that's not going to happen.
1431
01:28:26,355 --> 01:28:28,875
No, it can't make those kinds of mistakes.
1432
01:28:28,875 --> 01:28:37,915
It also can't generalize and just suck in information and mount something without somebody spending a lot of time on it.
1433
01:28:38,575 --> 01:28:50,355
So the generalization of LLMs is amazing, but the ability to be accurate, the logical groundedness isn't there.
1434
01:28:50,455 --> 01:28:55,135
I mean, fundamentally, they're thinking the same way people think, which is messy.
1435
01:28:55,135 --> 01:28:55,695
Right.
1436
01:28:55,695 --> 01:28:59,755
I mean, it's not really exactly the same way people think,
1437
01:28:59,876 --> 01:29:01,835
but it is messy like us.
1438
01:29:02,615 --> 01:29:05,895
And English, right, we talked before about how
1439
01:29:05,895 --> 01:29:09,355
there are ambiguous languages and unambiguous languages.
1440
01:29:09,655 --> 01:29:13,295
Constructing an argument in the language of English arguments
1441
01:29:13,295 --> 01:29:16,655
fundamentally does not have the guarantees
1442
01:29:16,655 --> 01:29:20,535
that we get from logic.
1443
01:29:21,255 --> 01:29:23,675
And I just, you know, it might feel like
1444
01:29:23,675 --> 01:29:25,155
we're talking about something...
1445
01:29:25,695 --> 01:29:28,295
like super cool part of AI you've never heard about.
1446
01:29:28,295 --> 01:29:31,815
But it's in a lot of ways analogous to writing code.
1447
01:29:32,395 --> 01:29:34,555
So, I mean, you're like,
1448
01:29:34,675 --> 01:29:36,435
why even bother writing any code?
1449
01:29:36,795 --> 01:29:39,435
Like whatever you're writing the code to do,
1450
01:29:39,675 --> 01:29:41,115
why doesn't the LLM just produce,
1451
01:29:41,155 --> 01:29:42,035
or if it's a multimodal model,
1452
01:29:42,115 --> 01:29:44,075
why doesn't it just produce that for you directly?
1453
01:29:44,555 --> 01:29:45,435
Well, the reason is,
1454
01:29:45,476 --> 01:29:49,175
is that you want to do something in a complex enough way
1455
01:29:49,175 --> 01:29:53,195
and you maybe care about certain guarantees or reliability.
1456
01:29:53,195 --> 01:29:56,255
and then you want to, you know,
1457
01:29:56,315 --> 01:29:58,075
and if you can put it into this code language
1458
01:29:58,075 --> 01:29:59,515
and then run it through a compiler,
1459
01:30:00,015 --> 01:30:02,515
well, it's kind of, you know,
1460
01:30:02,555 --> 01:30:03,955
I'm curious what you think about this,
1461
01:30:03,955 --> 01:30:06,755
but, you know, it's kind of the same.
1462
01:30:06,955 --> 01:30:08,955
It's just that instead of a compiler,
1463
01:30:09,215 --> 01:30:11,815
you have these specialized reasoning algorithms
1464
01:30:11,815 --> 01:30:13,595
or inference algorithms,
1465
01:30:13,935 --> 01:30:15,135
different kind of inference,
1466
01:30:15,295 --> 01:30:16,355
not neural network inference.
1467
01:30:18,055 --> 01:30:20,595
But then also there's a language.
1468
01:30:20,976 --> 01:30:22,955
And so what's amazing is like,
1469
01:30:22,955 --> 01:30:25,615
I used to write these by hand,
1470
01:30:25,935 --> 01:30:28,615
and oh my god, it was kind of like a painstaking,
1471
01:30:28,795 --> 01:30:30,175
it was like extra hard programming.
1472
01:30:30,595 --> 01:30:32,755
And the thing about answer set programming in general that I've heard,
1473
01:30:33,075 --> 01:30:36,035
so it runs, it basically all comes down to
1474
01:30:36,035 --> 01:30:38,235
a super optimized SAT solver.
1475
01:30:38,615 --> 01:30:42,035
So SAT is this famous computational problem
1476
01:30:42,035 --> 01:30:44,415
in, if you look at computational, like P equals N-P.
1477
01:30:44,415 --> 01:30:46,055
That's the same as the satisfiability of reference.
1478
01:30:46,175 --> 01:30:47,035
Yes, yeah, same thing.
1479
01:30:47,035 --> 01:30:48,155
In case anybody is confused.
1480
01:30:48,595 --> 01:30:48,775
SAT.
1481
01:30:49,016 --> 01:30:50,976
So there's been this ongoing work,
1482
01:30:50,976 --> 01:30:54,035
how to make more and more efficient SAT solving
1483
01:30:54,035 --> 01:30:57,095
because so many NP-hard problems
1484
01:30:57,095 --> 01:31:00,715
can be translated into a satisfiability problem.
1485
01:31:01,215 --> 01:31:03,355
And then if you can translate it,
1486
01:31:03,395 --> 01:31:05,715
then you can run it on this highly optimized SAT solver
1487
01:31:05,715 --> 01:31:06,815
to get the answer out.
1488
01:31:07,275 --> 01:31:07,395
Right.
1489
01:31:07,755 --> 01:31:10,535
So an example of this that I ran into,
1490
01:31:11,095 --> 01:31:13,275
I was working on, you know,
1491
01:31:13,295 --> 01:31:14,875
I work on key management systems
1492
01:31:14,875 --> 01:31:17,875
and I think about recovery flows.
1493
01:31:17,875 --> 01:31:20,495
Like keys for like the door.
1494
01:31:20,495 --> 01:31:22,155
Or, yeah, like getting your house.
1495
01:31:22,315 --> 01:31:25,315
No, like, you know, let's say protecting your Bitcoin
1496
01:31:25,315 --> 01:31:27,855
or protecting your AI data encryption keys.
1497
01:31:28,735 --> 01:31:29,635
Or SSHQs.
1498
01:31:29,835 --> 01:31:32,995
Or, yeah, whatever cryptographic key that's, like, high value
1499
01:31:32,995 --> 01:31:36,235
that you need to distribute it where you have, like, two of three
1500
01:31:36,235 --> 01:31:39,015
can move funds or authorize an action.
1501
01:31:39,695 --> 01:31:45,515
And I had this really hard problem where I wanted to have the system
1502
01:31:45,515 --> 01:31:48,375
be robust to all of these different scenarios,
1503
01:31:48,375 --> 01:31:58,835
like loss, theft, seizure, like every possible thing that can happen, which is really hard to kind of hit that bar.
1504
01:31:59,595 --> 01:32:05,275
And so I had these, I had this set of, I think there was like eight authentication factors.
1505
01:32:05,275 --> 01:32:10,115
Like there was like phone, hardware wallet, like a pin-based key.
1506
01:32:10,275 --> 01:32:11,515
There's like all these different keys.
1507
01:32:11,515 --> 01:32:13,275
There's like eight different factors.
1508
01:32:13,275 --> 01:32:17,815
and I wanted to create these sets
1509
01:32:17,815 --> 01:32:22,915
where you could always lose two and still spend.
1510
01:32:23,815 --> 01:32:25,995
And then I had other conditions where it was like
1511
01:32:25,995 --> 01:32:28,575
there should never be mobile-only factors
1512
01:32:28,575 --> 01:32:30,695
and all of this different stuff.
1513
01:32:30,695 --> 01:32:33,695
And I wanted to search through the entire space.
1514
01:32:34,855 --> 01:32:39,135
And it was like the power set of eight or so.
1515
01:32:39,155 --> 01:32:41,695
It was like some insanely huge number
1516
01:32:41,695 --> 01:32:45,715
number to brute force all the different ways of combining the factors
1517
01:32:45,715 --> 01:32:49,735
into these sets and then analyzing all the sets if they collected
1518
01:32:49,735 --> 01:32:52,315
them according to my rules and all this stuff.
1519
01:32:53,495 --> 01:32:57,315
And then that's when I realized, discovered, I had never heard about
1520
01:32:57,315 --> 01:33:00,976
SAT solvers. I was doing this research. It was like, oh, there's these SAT solvers
1521
01:33:00,976 --> 01:33:04,376
and you can encode the problem in this language and they can actually
1522
01:33:04,376 --> 01:33:09,595
search through this space that you could never get through with a brute force.
1523
01:33:09,595 --> 01:33:12,855
but they could search through this giant space
1524
01:33:12,855 --> 01:33:13,795
and I did that.
1525
01:33:14,015 --> 01:33:16,235
I used the, there was like a Google One
1526
01:33:16,235 --> 01:33:18,735
CP SAT or something like that.
1527
01:33:20,075 --> 01:33:22,455
I'm curious, was there many answers
1528
01:33:22,455 --> 01:33:23,555
or was there kind of one answer?
1529
01:33:23,555 --> 01:33:24,735
No, that was what was crazy.
1530
01:33:24,895 --> 01:33:28,875
It came up with like out of this huge giant space
1531
01:33:28,875 --> 01:33:31,755
there was like five different configurations
1532
01:33:31,755 --> 01:33:35,595
that actually satisfied the conditions.
1533
01:33:35,595 --> 01:33:36,035
That's wild.
1534
01:33:36,555 --> 01:33:37,555
And it felt like...
1535
01:33:37,555 --> 01:33:39,715
Talking about finding a gem or a needle in the haystack.
1536
01:33:39,715 --> 01:33:40,455
That's what it felt like.
1537
01:33:40,495 --> 01:33:42,495
It was like, this is the golden one.
1538
01:33:43,055 --> 01:33:44,735
And it was like, no wonder it's so hard.
1539
01:33:45,095 --> 01:33:47,215
You know, I'm on the whiteboard trying to think through,
1540
01:33:47,335 --> 01:33:49,355
like, what if we did this and what if we did that?
1541
01:33:49,375 --> 01:33:53,855
And I had no idea how doomed I was trying to do it that way
1542
01:33:53,855 --> 01:33:55,715
because the space was so big
1543
01:33:55,715 --> 01:33:58,215
and I was never going to find this, like, needle in the haystack.
1544
01:33:59,015 --> 01:34:01,895
But anyway, that's, like, just a whole long detour
1545
01:34:01,895 --> 01:34:05,755
just to provide an example of the power of a SAT solver
1546
01:34:05,755 --> 01:34:08,475
that you can search through this enormous space.
1547
01:34:08,555 --> 01:34:10,255
And you get strong guarantees,
1548
01:34:10,395 --> 01:34:12,575
or perfect guarantees almost, for your problem.
1549
01:34:12,875 --> 01:34:14,455
Yeah, which is hugely powerful.
1550
01:34:16,075 --> 01:34:19,155
Well, guys, should we kind of wrap up a little bit?
1551
01:34:19,215 --> 01:34:20,355
We've been talking a little while,
1552
01:34:21,195 --> 01:34:25,235
and it's been a pleasure having you here, Matt,
1553
01:34:25,235 --> 01:34:28,415
and talking about all this stuff.
1554
01:34:28,815 --> 01:34:32,235
I guess anything you guys want to...
1555
01:34:32,235 --> 01:34:34,855
Well, this has really got me thinking about
1556
01:34:34,855 --> 01:34:38,535
using LLMs in this way
1557
01:34:38,535 --> 01:34:39,995
of kind of combining them
1558
01:34:39,995 --> 01:34:41,895
with these other types of reasoning systems.
1559
01:34:42,015 --> 01:34:43,415
I think that's super fascinating.
1560
01:34:44,135 --> 01:34:46,355
I think that also could be
1561
01:34:46,355 --> 01:34:49,395
really important from a tokenomics perspective
1562
01:34:49,395 --> 01:34:50,575
where, you know,
1563
01:34:50,695 --> 01:34:52,195
who knows how expensive tokens
1564
01:34:52,195 --> 01:34:53,255
are going to get one day.
1565
01:34:53,495 --> 01:34:56,475
And if you can not use the LLM
1566
01:34:56,475 --> 01:34:57,635
for every single thing
1567
01:34:57,635 --> 01:34:58,655
that's happening in the system
1568
01:34:58,655 --> 01:35:00,715
and kind of have a more targeted use,
1569
01:35:01,535 --> 01:35:02,875
that could be interesting.
1570
01:35:02,875 --> 01:35:13,415
But it's often much more efficient as well, at least in the sense of, you know, taking less computational power to come up with an answer.
1571
01:35:13,575 --> 01:35:14,195
It doesn't necessarily...
1572
01:35:14,233 --> 01:35:22,713
necessarily, it might require you to spend more development effort
1573
01:35:22,713 --> 01:35:24,833
to get working at an LLM.
1574
01:35:25,513 --> 01:35:28,973
But for that, you generally get back some benefits
1575
01:35:28,973 --> 01:35:31,653
in not needing so many GPUs to get stuff done.
1576
01:35:32,433 --> 01:35:33,413
So good for the environment.
1577
01:35:33,933 --> 01:35:38,333
You're worried about water usage, you know, use Prologue or something like that.
1578
01:35:39,273 --> 01:35:40,813
But yeah, no, it's absolutely fascinating.
1579
01:35:40,913 --> 01:35:41,973
I'm going to be thinking about this.
1580
01:35:41,973 --> 01:35:44,293
So I really appreciate your insights about it.
1581
01:35:45,573 --> 01:35:47,373
Well, I've really enjoyed talking to you.
1582
01:35:47,833 --> 01:35:48,633
This has been fun, guys.
1583
01:35:49,193 --> 01:35:49,513
Awesome.
1584
01:35:49,893 --> 01:35:50,053
Yeah.
1585
01:35:50,893 --> 01:35:54,553
Well, until next time, please like and subscribe and leave us a comment.
1586
01:35:54,993 --> 01:35:57,073
And we'll see you next time.
1587
01:35:57,173 --> 01:35:57,373
Yeah.
1588
01:35:57,493 --> 01:35:58,053
See you next time.