Claude’s Less Obvious Edge, AI Layoffs, Prompting & Consciousness

1
00:00:00,000 --> 00:00:04,120
Hello everybody and welcome to another edition of the podcast.
2
00:00:04,600 --> 00:00:06,740
Anything that's been on your mind recently?
3
00:00:07,740 --> 00:00:12,940
So I was talking to a friend this morning about what the, what an agentic organization
4
00:00:12,940 --> 00:00:14,780
of the future might look like.
5
00:00:14,940 --> 00:00:15,300
Okay.
6
00:00:15,480 --> 00:00:19,560
And it was kind of coming up, it was like bringing up some, some spicy takes.
7
00:00:19,700 --> 00:00:22,340
I feel like that is like a spicy take.
8
00:00:22,440 --> 00:00:27,800
I basically was thinking that everything we've been seeing now with the layoffs of
9
00:00:27,800 --> 00:00:33,940
these big companies. Right. Where, and you know, what Block is doing is especially interesting
10
00:00:33,940 --> 00:00:39,020
because I think they're, they're cutting at least from some stuff that I read, maybe not directly
11
00:00:39,020 --> 00:00:45,400
from them, but in the space was that they're, they're kind of really reducing meetings and
12
00:00:45,400 --> 00:00:51,980
looking to like have more people work directly on content and all of that, um, or produce things
13
00:00:51,980 --> 00:00:56,840
maybe more closely to, you know, instead of maybe like more middle managers or things like that.
14
00:00:56,840 --> 00:00:59,980
Um, which, which is certainly interesting.
15
00:01:00,240 --> 00:01:00,580
Yeah.
16
00:01:00,600 --> 00:01:07,340
What I was thinking is like, in some ways, I think it's a lack of creativity for a company
17
00:01:07,720 --> 00:01:16,500
to do just the framing of replacing humans with AI, I think is like the most shallow
18
00:01:16,500 --> 00:01:18,740
near term thing you could do.
19
00:01:19,120 --> 00:01:22,700
And I, I understand there's all these dynamics at play.
20
00:01:22,700 --> 00:01:27,620
Um, but let's just think about context for a second, right?
21
00:01:27,860 --> 00:01:32,980
So when you're working with AI, the most important thing to be able to get really
22
00:01:32,980 --> 00:01:35,000
good results out of it is context.
23
00:01:35,000 --> 00:01:37,480
Context engineering is this term now, right?
24
00:01:37,480 --> 00:01:38,240
Absolutely.
25
00:01:38,300 --> 00:01:38,700
Okay.
26
00:01:39,100 --> 00:01:41,380
Anybody who has been working at your company, right?
27
00:01:41,380 --> 00:01:41,540
Okay.
28
00:01:41,540 --> 00:01:46,820
So, so some people, maybe they're not aligned with the mission or they're not
29
00:01:47,060 --> 00:01:52,580
putting in enough time or there's, you know, poor performance or whatever there's,
30
00:01:52,700 --> 00:01:58,060
You know, sometimes there's behavior like that that isn't conducive to helping the company with what it needs to do.
31
00:01:58,360 --> 00:02:07,500
But besides that, right, I think there's such a massive opportunity to be upskilling people rather than replacing them.
32
00:02:07,500 --> 00:02:21,020
And the reason is, if context is king, those people have so much context already about the company, the organization, like the people in it, how it works, the processes.
33
00:02:21,020 --> 00:02:30,320
and so the thing is is that and this is like a point i think is that i think the context that a
34
00:02:30,320 --> 00:02:37,460
human gathers will never completely overlap with the context an ai could gather or have access to
35
00:02:37,460 --> 00:02:44,900
i think i would argue that there's always some gap there's some context that a human is going to
36
00:02:44,900 --> 00:02:49,520
like pick up or absorb that is based.
37
00:02:49,560 --> 00:02:52,580
You're basically like never really going to get that into an AI system.
38
00:02:53,120 --> 00:02:57,760
And there's like a lot of ways you could, you could take this, but if that's the
39
00:02:57,760 --> 00:03:01,640
case, you should be designing.
40
00:03:03,620 --> 00:03:08,700
Almost like this agentic aura or like a system of AI tools or agents that
41
00:03:08,700 --> 00:03:11,960
basically make that person so much more powerful.
42
00:03:11,960 --> 00:03:25,780
And in the long run, you're going to get so much more benefits from that because, you know, I think, you know, our subconscious is crazy powerful.
43
00:03:26,640 --> 00:03:29,1000
It's, it's like, and also we are effortless.
44
00:03:30,080 --> 00:03:33,200
Like humans are meaning making machines, right?
45
00:03:33,800 --> 00:03:39,980
Like we effortlessly flip between perspectives and framing for what we're doing.
46
00:03:39,980 --> 00:03:47,560
and we can you know you can go from cooking a meal to hop in your car and driving to like working on
47
00:03:47,560 --> 00:03:53,280
your computer to like playing a video game and you're just like this you are the same in all of
48
00:03:53,280 --> 00:04:02,340
those and yet you are able to adopt perspective and it's almost like you're in a different
49
00:04:02,340 --> 00:04:08,800
environment in all of those but it's like a mental environment yeah i mean i really like how you're
50
00:04:08,800 --> 00:04:17,700
putting this. Um, but I think, uh, you know, there's the potential of what a human being
51
00:04:17,700 --> 00:04:25,680
can be or is capable of. But I mean, have you ever had an experience working with somebody
52
00:04:25,680 --> 00:04:35,240
where for some reason, um, common sense doesn't seem to be a part of their equation or the ability
53
00:04:35,240 --> 00:04:43,160
to or willingness to reason through something from first principles, not that that requires
54
00:04:43,160 --> 00:04:48,360
great intelligence or like that that is beyond the capability of humans.
55
00:04:48,920 --> 00:04:55,320
But if somebody isn't using the tools that they have in front of them in the right way,
56
00:04:55,320 --> 00:05:02,780
even though they have that potential or that capability, then are you going to get the
57
00:05:02,780 --> 00:05:03,280
result?
58
00:05:03,280 --> 00:05:11,280
And of course, that's a really hard part of running an organization or hiring people is you have to make really tough decisions.
59
00:05:11,580 --> 00:05:19,660
Even if somebody is well-intentioned or has a lot of potential, it doesn't mean they're going to be able to get the job done.
60
00:05:19,980 --> 00:05:30,060
And if they can't, keeping people around who aren't able to do that can be really detrimental to the ability to hit the objective overall.
61
00:05:30,060 --> 00:05:37,380
so not to bring in too much cynicism and no no i love it i love it but just to kind of keep things
62
00:05:37,380 --> 00:05:44,820
like grounded no no of course i mean of course at the end of the day you have to be you know
63
00:05:44,820 --> 00:05:51,900
you have to be producing value of some kind and um and i think you touched on the like one of the
64
00:05:51,900 --> 00:05:58,360
most important things and i feel like one of the most important things is like willingness and i
65
00:05:58,360 --> 00:06:03,500
think there's almost maybe two kinds of willingness here one is a willingness to accept that you don't
66
00:06:03,500 --> 00:06:12,080
know something that's big actually this was for a while so in one of my past roles this was in some
67
00:06:12,080 --> 00:06:14,940
ways i think this might be one of the most important interview questions you could ever
68
00:06:14,940 --> 00:06:21,840
or interview like line of questions you might ask somebody but basically in any interview you you
69
00:06:21,840 --> 00:06:32,260
You only want to hire someone if they, well, or they might be an ideal candidate, if you can get them to a place to say, I don't know, in the interview.
70
00:06:32,560 --> 00:06:35,060
Because someone who says, I don't know, they're coachable.
71
00:06:35,500 --> 00:06:35,980
Right.
72
00:06:36,700 --> 00:06:37,660
They're honest.
73
00:06:38,100 --> 00:06:38,400
Right.
74
00:06:38,660 --> 00:06:41,600
You know, you can, from a space of, I don't know, you can work.
75
00:06:41,800 --> 00:06:42,500
That's great.
76
00:06:42,980 --> 00:06:43,220
Right.
77
00:06:43,220 --> 00:06:56,480
But if someone who never says, I don't know, they might be kind of always putting on the impression of like doing the work or getting close to doing the work.
78
00:06:57,360 --> 00:07:10,960
But and so then there's, you know, it might, I think often sometimes those kinds of people, you might get, you know, such delayed effects before you realize that there's just a communication issue or they're not on the same page with you.
79
00:07:10,960 --> 00:07:25,260
100% or you're trying to get on the same page, but their version of the truth keeps changing or the story keeps changing to kind of adapt what they think is going to make them look good in the situation.
80
00:07:25,540 --> 00:07:25,740
Right.
81
00:07:25,740 --> 00:07:28,940
And so they can't acknowledge what they don't know.
82
00:07:29,020 --> 00:07:30,240
They can't acknowledge.
83
00:07:30,660 --> 00:07:39,480
You can't have a dialectical process where you weigh the pros and cons of things that
84
00:07:39,480 --> 00:07:43,860
where the truth seeking objective is prioritized.
85
00:07:44,080 --> 00:07:48,100
Instead, these other agendas of are getting to the mix.
86
00:07:48,100 --> 00:07:54,280
And I think, and I find like when somebody does stuff like that, to me, it's like they're
87
00:07:54,280 --> 00:08:02,020
basically useless in an organizational context, like to, you know, be a little bit blunt about it.
88
00:08:02,020 --> 00:08:07,680
But it's like, even if the person's highly intelligent and really well-intentioned and
89
00:08:07,680 --> 00:08:15,040
really motivated, like if they can't have a consistent principled analysis or dialogue about
90
00:08:15,040 --> 00:08:22,160
something, there's no way to collaborate effectively or problem solve around complex issues.
91
00:08:22,160 --> 00:08:28,120
and instead you just get all these like political games where people are trying to take credit or
92
00:08:28,120 --> 00:08:33,060
offload blame it has nothing to do with the customer it has nothing to do with the company's
93
00:08:33,060 --> 00:08:45,080
mission um and so but i think the irony there is people who are highly insecure are over project
94
00:08:45,080 --> 00:08:50,620
over confidence like you would think if you're insecure you would show humility because you're
95
00:08:50,620 --> 00:08:54,700
like, I don't believe in myself. I don't know what's going on. But when you're insecure,
96
00:08:55,540 --> 00:09:04,400
you are so anxious about it that you want to project, um, and kind of hide the weakness.
97
00:09:04,940 --> 00:09:11,340
So there's like this overcompensation of like, no, I know everything where somebody who is like
98
00:09:11,340 --> 00:09:16,900
comfortable and confident, they can admit what they don't know because it doesn't threaten
99
00:09:16,900 --> 00:09:23,260
their ego in this deep way that kind of triggers like an existential need it's like it's so i'm
100
00:09:23,260 --> 00:09:28,860
i'm comfortable knowing that you know i don't know certain things or admitting i don't know
101
00:09:28,860 --> 00:09:36,160
certain things that's fine like my ego can take it right but i i would argue that's a skill issue
102
00:09:36,160 --> 00:09:42,760
yeah like why can't that be upskilled like dude are we getting superhuman ai or what
103
00:09:42,760 --> 00:09:47,080
Like, like, are we getting like the therapist from the future?
104
00:09:47,340 --> 00:09:52,240
Like, again, I like, so I was going to say, so I said the one type of willingness.
105
00:09:52,240 --> 00:10:00,520
I mean, basically, I'm, you know, agreeing 100%, but I just, so willingness to say, I don't know.
106
00:10:00,980 --> 00:10:01,460
Right.
107
00:10:01,880 --> 00:10:06,380
And then the other type of willingness is willingness to put in the time.
108
00:10:07,000 --> 00:10:10,420
Like the, like real, like that is the thing today, right?
109
00:10:10,420 --> 00:10:21,180
If you are willing to put in the time to work with AI to get results, which also is maybe kind of another ego check.
110
00:10:21,320 --> 00:10:21,940
It is.
111
00:10:22,060 --> 00:10:33,240
Or it's like a form of, like, because it doesn't really matter anymore whether you believe you can do something.
112
00:10:33,820 --> 00:10:38,680
It's just a matter of, there's a notion of like how much time it will take.
113
00:10:38,680 --> 00:10:45,420
but like honestly I think the most important attributes to hire someone these days are like
114
00:10:45,420 --> 00:10:54,380
those two those two kinds of willingness I agree uh a hundred percent and you know I think it is
115
00:10:54,380 --> 00:11:08,660
related to like a similar thing um where uh you know if you're uh um if you're if you're
116
00:11:08,680 --> 00:11:12,360
willing to, like, a lot of people,
117
00:11:12,600 --> 00:11:15,280
they think of themselves as not intelligent
118
00:11:15,280 --> 00:11:16,560
or not good at something.
119
00:11:17,060 --> 00:11:18,340
Like, I'm not good at math,
120
00:11:18,480 --> 00:11:20,360
or I'm not good at this, or I'm not good at that.
121
00:11:20,420 --> 00:11:22,700
And a lot of times that started at a young age
122
00:11:22,700 --> 00:11:26,600
where, you know, even when you're in elementary school,
123
00:11:26,740 --> 00:11:29,360
there's already this division that starts happening
124
00:11:29,360 --> 00:11:31,840
between, like, the smart kids and the dumb kids
125
00:11:31,840 --> 00:11:34,320
or, like, who's going, who has, like, you know,
126
00:11:34,740 --> 00:11:36,700
the special talented program
127
00:11:36,700 --> 00:11:38,500
versus the kids who are, like,
128
00:11:38,680 --> 00:11:45,420
And like some of that has to do with like actual like ability or IQ or genetics, but
129
00:11:45,420 --> 00:11:50,300
there's also the story and the narrative and the identity that emerges.
130
00:11:51,000 --> 00:11:55,900
And like, I think a lot of people who think they're bad at something, who think they're
131
00:11:55,900 --> 00:12:02,120
bad at math, they'll go to study, look at some math and it'll immediately be overwhelming
132
00:12:02,120 --> 00:12:03,260
and feel impossible.
133
00:12:03,260 --> 00:12:10,060
And what they don't realize is the person who's good at math also feels that way when they start.
134
00:12:10,720 --> 00:12:15,140
But they think the person who's good at math just looks at it and it all makes sense.
135
00:12:15,900 --> 00:12:16,420
Right.
136
00:12:16,500 --> 00:12:20,400
But the difference with the smart person is not that they just get it immediately.
137
00:12:20,560 --> 00:12:28,180
I mean, maybe some super genius does, you know, some Neumann is like deriving Euclid's propositions when he's five years old.
138
00:12:28,180 --> 00:12:42,860
But like a lot of times it has more to do with patience, with curiosity, with allowing the learning process to proceed where there's a lot of stumbles.
139
00:12:43,080 --> 00:12:44,400
There's a lot of setbacks.
140
00:12:44,400 --> 00:12:46,200
There's a lot of frustrations.
141
00:12:46,620 --> 00:12:57,140
Having the emotional temperament where that, again, it kind of links oftentimes to insecurity where if that triggers this anxiety or this feeling insecurity,
142
00:12:57,140 --> 00:13:02,840
you can't keep going and you just kind of give up and you're like, I can't do this.
143
00:13:02,840 --> 00:13:09,800
But if you have the patience and the wherewithal and ideally, because the other thing is a lot of
144
00:13:09,800 --> 00:13:16,360
kids aren't taught that learning is fun. It's just work or it's stressful or it's always
145
00:13:16,360 --> 00:13:23,680
associated with these negative experiences. But when it's more of a playful, you know,
146
00:13:23,680 --> 00:13:28,960
you don't have to arrive anywhere. It doesn't have to be anything, you know, it's just thinking,
147
00:13:29,260 --> 00:13:35,180
what about this? What this, that, you know, and then that creates more space where you can just,
148
00:13:35,180 --> 00:13:42,200
um, slowly over time, try things out and learn and kind of feel your way into it.
149
00:13:42,820 --> 00:13:51,580
So, you know, I think that is a similar route of all these different issues of like,
150
00:13:51,580 --> 00:13:58,220
you know, I think Pascal said that like man's biggest problem is that he can't sit alone
151
00:13:58,220 --> 00:14:03,840
quietly in a room by himself. Cause that's the other thing is like when you're just,
152
00:14:03,840 --> 00:14:09,900
can you sit still? Can you slow down? Can you slow your thinking and just
153
00:14:09,900 --> 00:14:16,740
wonder about something for a second without knowing the answer? And oftentimes that happens,
154
00:14:16,740 --> 00:14:20,520
that is related to your comfort, which is emptiness or silence.
155
00:14:20,520 --> 00:14:24,860
If there's this anxiousness that has to fill it in right away,
156
00:14:24,860 --> 00:14:31,400
then you can't really have deep thought about something.
157
00:14:32,360 --> 00:14:33,980
Yeah, absolutely. Absolutely.
158
00:14:34,500 --> 00:14:40,280
It almost makes me feel like you almost need a level of fearlessness.
159
00:14:41,740 --> 00:14:44,780
And that might be another really important quality.
160
00:14:44,780 --> 00:14:54,320
you almost also need a fearlessness to start to work on problems or in a space that you
161
00:14:54,320 --> 00:15:03,260
haven't before because it's like this it's like this trade-off right like you don't know
162
00:15:03,260 --> 00:15:10,100
like if you if you were really good at something you would only know that because you went through
163
00:15:10,100 --> 00:15:16,700
it and you did it and you did a lot of work with it, but you can never know that before you start.
164
00:15:18,440 --> 00:15:28,600
And this is all of this though. I, I'm, you know, skill issue is like a funny, funny phrase,
165
00:15:28,600 --> 00:15:37,320
you know, but, but really I do wonder if it is largely something that an AI can make you get a
166
00:15:37,320 --> 00:15:44,780
lot better at yeah i mean i think it's a mix of both right there's like the bucket of people
167
00:15:44,780 --> 00:15:52,980
where they're ready to have this activated they just need the right teacher and that's the category
168
00:15:52,980 --> 00:16:00,420
where it's like yes ai can solve that then you have the bucket of people where like temperamentally
169
00:16:00,420 --> 00:16:05,740
or where they're at in life or for whatever reason you can give them the perfect teacher
170
00:16:05,740 --> 00:16:10,120
the perfect everything, it's just not going to work. It's not going to click, at least for that
171
00:16:10,120 --> 00:16:14,600
moment in their life. Like maybe something else will happen. Maybe some life experience needs to
172
00:16:14,600 --> 00:16:22,120
happen. Maybe they need to like go on vacation or, you know, meditate or like something more than
173
00:16:22,120 --> 00:16:29,660
an intellectual exercise has to occur to kind of like develop a person, which I mean, AI can help
174
00:16:29,660 --> 00:16:34,580
with that too. So what, you know, maybe I'm being overly limited here because AI can be your shy
175
00:16:34,580 --> 00:16:35,560
It could be your priest.
176
00:16:35,560 --> 00:16:36,800
I mean, I kind of...
177
00:16:36,800 --> 00:16:37,440
Whatever you need it to be.
178
00:16:37,600 --> 00:16:39,320
I kind of wonder if...
179
00:16:39,320 --> 00:16:40,340
I've been thinking about this.
180
00:16:40,480 --> 00:16:43,460
Like, there's this TV show that I really...
181
00:16:43,460 --> 00:16:45,360
I really like the TV show The Magicians.
182
00:16:45,900 --> 00:16:48,080
It's like a comfort show for me.
183
00:16:48,080 --> 00:16:49,220
It's kind of campy.
184
00:16:50,500 --> 00:16:53,700
But it's, you know, I'm a sucker for, like, world building...
185
00:16:53,700 --> 00:16:54,660
Fantasy world building.
186
00:16:54,660 --> 00:16:57,600
And whenever anything has, like, its own...
187
00:16:57,600 --> 00:17:01,140
Someone comes up with, like, a magical system and rules that are, like, unique.
188
00:17:01,280 --> 00:17:03,140
And so I'm like, ah, let me go check this out.
189
00:17:03,140 --> 00:17:07,300
I really, I really like the world building. Um, and they've got some interesting, unique
190
00:17:07,300 --> 00:17:13,840
aspects of how magic works and things like that in the show. Um, but you know, you know,
191
00:17:13,840 --> 00:17:20,180
in one of the later seasons, they, they basically go on this, this quest to kind of, I don't,
192
00:17:20,180 --> 00:17:23,760
I won't say more. I don't want to spoil it. It's an older show, but yeah, spoiler alert,
193
00:17:23,760 --> 00:17:32,820
but they go on this quest to like, and the quest is like gnarly, like, like really dangerous.
194
00:17:33,140 --> 00:17:40,580
um really difficult kind of even like way harder than almost like
195
00:17:40,580 --> 00:17:49,720
normal life problems or things like that and it just got me thinking like like how fun like maybe
196
00:17:49,720 --> 00:17:57,340
there's a really fun way you could build an ai that essentially gives you a growth quests to do
197
00:17:57,340 --> 00:18:03,900
in your real life to kind of like level up, you know, like emotionally or spiritually or like
198
00:18:03,900 --> 00:18:11,540
overcome a lot of these things. Because like once you go through a lot of these, you look back and
199
00:18:11,540 --> 00:18:16,280
you're like, wait, I was so anxious or nervous about like people say it's a lot about public
200
00:18:16,280 --> 00:18:23,240
speaking, right? I mean, you know, I think it's a common thing in surveys that people would rather
201
00:18:23,240 --> 00:18:28,740
die but you're literally more afraid of public speaking than they are of death yeah i've seen
202
00:18:28,740 --> 00:18:34,540
that too it's pretty remarkable and so but it's also one of those things where it's like a muscle
203
00:18:34,540 --> 00:18:41,340
and if you do it like you know i was i started to do it more and more i was used to be very nervous
204
00:18:41,340 --> 00:18:48,720
my like palms would sweat like oh my god but then i think after a couple hundred times i finally got
205
00:18:48,720 --> 00:18:54,940
you know i was able to like get over it um to some degree right and also recognize the other
206
00:18:54,940 --> 00:19:01,940
thing is that the nervousness can be good it actually makes you sharp and take it seriously
207
00:19:01,940 --> 00:19:10,500
yeah so it's not even a bad thing and even that framing of it um helped but i just yeah i think
208
00:19:10,500 --> 00:19:19,060
like but again this is where i feel like we're not even giving us space for ai to
209
00:19:19,060 --> 00:19:27,060
like help us be better humans in all ways right because we i think i think we literally take
210
00:19:27,060 --> 00:19:33,500
a lot of things for granted like we take for granted that or at least some you know at least
211
00:19:33,500 --> 00:19:39,440
maybe me my past self or you maybe see other people you just take for granted that they
212
00:19:39,440 --> 00:19:53,118
are unable to do a certain thing Like they have such an internal identity right that not true it like literally not it literally like not it just some story they started telling
213
00:19:53,118 --> 00:20:01,237
themselves or some story i was telling myself ai can help you break out of that yeah i mean i think
214
00:20:01,237 --> 00:20:12,398
a lot of these stories are deeply embedded in our culture, religion, or family dynamics,
215
00:20:13,038 --> 00:20:21,038
you know, dynamics around shame. I think that's a big one. People feeling like they're worthless,
216
00:20:21,038 --> 00:20:25,138
or they have no value, they have no power.
217
00:20:25,138 --> 00:20:32,618
And all of these energies and stories aren't just in the individual
218
00:20:32,618 --> 00:20:36,658
but are in the collective and it's being reinforced.
219
00:20:37,678 --> 00:20:43,098
And some of it is promulgated in the media
220
00:20:43,098 --> 00:20:50,098
or if you want to control somebody or have power over somebody,
221
00:20:50,098 --> 00:20:57,018
then it can be easier to do that if you are promulgating these stories about people's
222
00:20:57,018 --> 00:21:01,957
powerlessness. Even if that's not happening consciously, like the incentives are wired
223
00:21:01,957 --> 00:21:09,318
such to do that. So I think that's kind of, you know, I'm even looking at something like
224
00:21:09,318 --> 00:21:16,778
the story of Adam and Eve and the garden and the fall of man and original sin, like all these things
225
00:21:16,778 --> 00:21:23,398
are encoded. And I think deep in the human psyche, you know, in, in the unconscious and
226
00:21:23,398 --> 00:21:31,318
in the repressed pain in the places where people don't want to go things they don't
227
00:21:31,318 --> 00:21:45,098
want to fear or feel is usually these deep feelings of shame, guilt, um, fear, anxiety,
228
00:21:45,098 --> 00:21:48,477
not just about death, but about the nature of reality itself.
229
00:21:49,858 --> 00:21:54,538
Um, and so there's all this stuff, you know, underneath.
230
00:21:54,538 --> 00:22:01,197
And I think the really powerful thing about meditation is that it teaches you to stop
231
00:22:01,197 --> 00:22:03,658
running away from those things.
232
00:22:03,758 --> 00:22:10,197
Like we have Heidegger talks about this actually, that, uh, basically it's really interesting
233
00:22:10,197 --> 00:22:18,838
how he puts it in being in time, that there's, because there is this kind of terror underneath
234
00:22:18,838 --> 00:22:24,477
that we're unconscious of, like we're just kind of thrown into the world and we don't,
235
00:22:24,658 --> 00:22:28,278
we're not fully conscious of all the things that are driving us and motivating us.
236
00:22:28,818 --> 00:22:34,717
So there's, before we're able to even reflect on it, we kind of already have this instinctual
237
00:22:34,717 --> 00:22:46,598
flight away from anxiety, existential anxiety. So Heidegger says that what happens is there's a
238
00:22:46,598 --> 00:22:59,598
flight into inauthenticity because if I interpret reality as I'm seeing it, then I have to confront
239
00:22:59,598 --> 00:23:02,358
all this stuff that I don't want to feel or experience.
240
00:23:02,538 --> 00:23:08,398
But if instead I kind of borrow other people's story about reality,
241
00:23:08,957 --> 00:23:11,338
that's sort of where the group thing happens,
242
00:23:11,457 --> 00:23:13,598
where you're not in your own experience.
243
00:23:13,798 --> 00:23:19,338
You're in this projected sense of what they are doing.
244
00:23:20,058 --> 00:23:22,838
So he literally calls it the they self.
245
00:23:23,957 --> 00:23:25,678
We're not thinking in terms of I,
246
00:23:25,758 --> 00:23:28,717
you're thinking in terms of what do they think.
247
00:23:29,598 --> 00:23:35,438
and there's this almost anesthetizing thing where like in the they there's no death in the they
248
00:23:35,438 --> 00:23:43,118
there's nothing like it's this inauthentic kind of fake reality so a lot of times like
249
00:23:43,758 --> 00:23:46,957
and you'll see people you know they'll they'll use this vernacular where they'll be like
250
00:23:47,678 --> 00:23:52,398
should we care about this or like why are we so upset about this like what is this we that
251
00:23:52,398 --> 00:23:57,518
they're talking about it's almost as if their identity has merged with this group you know it
252
00:23:57,518 --> 00:24:04,338
It almost is, um, it's like connected to popularity contests or like, are you in the in group
253
00:24:04,338 --> 00:24:05,058
or the out group?
254
00:24:05,058 --> 00:24:09,618
And your identity isn't this authentic, like individual loner.
255
00:24:09,737 --> 00:24:12,957
Your identity is like, you don't believe in yourself.
256
00:24:12,957 --> 00:24:14,998
You believe in the group.
257
00:24:15,378 --> 00:24:19,518
And that is kind of helping you deal with this existential anxiety.
258
00:24:19,518 --> 00:24:46,258
So for Heidegger to engage in philosophy requires this authentic interrogation of your own experience and to turn towards the anxiety, not away from it, and reclaim your own sense of reality that has nothing to do with what they think,
259
00:24:46,258 --> 00:24:49,618
but is about what do you think?
260
00:24:49,818 --> 00:24:51,158
What do you experience?
261
00:24:51,338 --> 00:24:52,318
What do you feel?
262
00:24:53,078 --> 00:24:54,938
And that does require courage.
263
00:24:55,038 --> 00:24:56,998
That does require strength.
264
00:24:58,638 --> 00:25:03,138
And it can be very difficult to like go through something like that.
265
00:25:03,138 --> 00:25:10,818
But then at the other end, you have this foundation to relate to things
266
00:25:10,818 --> 00:25:14,318
on the basis of truth or on the basis of being
267
00:25:14,318 --> 00:25:16,098
or the basis of what they are.
268
00:25:16,258 --> 00:25:21,038
not just popularity contests or ego games or politics,
269
00:25:21,038 --> 00:25:26,378
but like the underlying reality is no longer so terrifying
270
00:25:26,378 --> 00:25:30,018
or so buried that it can't be accessed.
271
00:25:30,018 --> 00:25:33,018
Like that becomes sort of the ground and the basis.
272
00:25:33,258 --> 00:25:38,318
And that is the source of any kind of real power in the world
273
00:25:38,318 --> 00:25:42,158
where then you can create your own reality
274
00:25:42,158 --> 00:25:44,598
and have your own influence on things
275
00:25:44,598 --> 00:25:48,518
and not just be like a pawn in somebody else's like manipulation.
276
00:25:49,457 --> 00:25:51,898
Dude, that's some powerful stuff right there.
277
00:25:52,398 --> 00:25:55,197
You know, just one comment I'll say is,
278
00:25:56,898 --> 00:26:01,418
you know, there's this, oftentimes you might hear,
279
00:26:01,678 --> 00:26:02,378
I don't know about often,
280
00:26:02,498 --> 00:26:05,578
but there's this concept of like the fire of awareness
281
00:26:05,578 --> 00:26:07,338
or the fire of attention.
282
00:26:08,318 --> 00:26:11,858
And it's more so that like you need like courage,
283
00:26:11,858 --> 00:26:29,237
But I think it's really that for those negative emotions that get stored in the nervous system, in the body, maybe in the psyche, for those to leave you, you can't run from them.
284
00:26:29,237 --> 00:26:31,058
It's like you can't run from them.
285
00:26:32,118 --> 00:26:38,678
Like meditating and creating space to observe whatever is going on within you
286
00:26:38,678 --> 00:26:44,197
provides a space for those things to leave you,
287
00:26:44,318 --> 00:26:46,378
but before they leave you, they make themselves known.
288
00:26:47,578 --> 00:26:51,378
And so that's why it can be so uncomfortable and so painful.
289
00:26:52,078 --> 00:26:55,798
And this is why I think it's very important to get a meditation teacher
290
00:26:55,798 --> 00:26:57,858
or someone who's very experienced in meditation.
291
00:26:57,858 --> 00:27:04,338
um especially like the deeper practices and there's you know a lot of great organizations
292
00:27:04,338 --> 00:27:10,158
and stuff um I feel very fortunate I myself had a really good teacher and
293
00:27:12,558 --> 00:27:19,038
you know and it's honestly an ongoing thing but when you come out the other side and when you
294
00:27:19,038 --> 00:27:26,658
give freedom from shame or negativity or stress or you know things from the past I mean
295
00:27:27,858 --> 00:27:34,858
it's, it's like the, the best thing ever. I mean, it's like anything you ever,
296
00:27:35,457 --> 00:27:38,737
it's like, no matter what you might ever be looking for outside in the world,
297
00:27:39,378 --> 00:27:46,778
inside you have a, a richer treasure, but it's just covered with like some black dust and you
298
00:27:46,778 --> 00:27:52,258
just got to clean the dust. It might look really thick. Oftentimes it's not as thick as it looks,
299
00:27:52,258 --> 00:27:56,798
just looks really thick. But then, you know, a little bit of the fire of attention, fire of
300
00:27:56,798 --> 00:28:01,118
awareness and a willingness to just be uncomfortable and to sit with whatever it is
301
00:28:01,118 --> 00:28:09,498
purifies it it leaves your system and then you are this like the diamond that you are the gem
302
00:28:09,498 --> 00:28:16,098
that you are like shines brilliantly and you like like you said you gain all this power
303
00:28:16,098 --> 00:28:24,197
and you become more influential and i don't see why ai can't help unlock that
304
00:28:24,197 --> 00:28:26,438
No, I think it can.
305
00:28:26,658 --> 00:28:27,998
I absolutely think it can.
306
00:28:28,358 --> 00:28:33,298
You know, and I mean, I've talked to AI about stuff like this,
307
00:28:33,338 --> 00:28:37,798
and it seems very conversant and, you know, gets it immediately
308
00:28:37,798 --> 00:28:41,378
and has insights about these types of topics.
309
00:28:41,798 --> 00:28:45,118
And I'm sure there's going to be ways to develop that
310
00:28:45,118 --> 00:28:49,697
and kind of formalize that into tools and systems that people can use.
311
00:28:49,697 --> 00:28:52,737
Well, the thing I love about it is these things have the capability
312
00:28:52,737 --> 00:28:56,758
for infinite patience, infinite gentleness.
313
00:28:57,318 --> 00:28:57,418
Right.
314
00:28:57,778 --> 00:28:59,158
Dude, come on.
315
00:28:59,197 --> 00:29:00,398
If you're a human on this planet,
316
00:29:00,758 --> 00:29:02,438
you've gone through some shit, dude.
317
00:29:02,638 --> 00:29:03,678
Like, there's no way.
318
00:29:03,678 --> 00:29:07,558
Like, whether it's, like, some family or some, you know,
319
00:29:07,658 --> 00:29:10,838
whether it's in school or whatever, like, dude,
320
00:29:10,898 --> 00:29:12,618
that's part of what it is to be human, right?
321
00:29:12,678 --> 00:29:12,938
Yes.
322
00:29:12,977 --> 00:29:17,398
Like, and the, you know, it can be a lot of pain, a lot, you know.
323
00:29:17,398 --> 00:29:25,918
but for what for who what you really are like you are so much bigger than that right but
324
00:29:25,918 --> 00:29:34,298
if that isn't cleared out then you know you you can't see it or feel it or like identify with it
325
00:29:34,298 --> 00:29:38,558
and so but and so in some ways it's almost like
326
00:29:38,558 --> 00:29:46,138
i'm reminded of like this this workbook i worked through once um for just some specific things i
327
00:29:46,138 --> 00:29:51,538
was working on like like five ten years ago and like you know basically it was like just a series
328
00:29:51,538 --> 00:29:58,998
of very personal intimate questions designed to help you pull out things you might be
329
00:29:58,998 --> 00:30:07,638
you know stuck on related and you know is like very emotional for me a lot of stuff came up
330
00:30:07,638 --> 00:30:13,678
and then i i mean i felt so much freedom after but what was it at the end of the day it basically
331
00:30:13,678 --> 00:30:20,318
was this workbook of like, I was like 25 pages. Every page had like five to seven questions.
332
00:30:20,878 --> 00:30:24,538
And I, you know, it took me, I didn't, you know, I would sit for a couple hours at a time and work
333
00:30:24,538 --> 00:30:26,977
through, I actually had a friend at the time, we were both doing it. So we were both working
334
00:30:26,977 --> 00:30:33,638
through it, which was kind of nice. But like, there's no reason you couldn't get the same thing
335
00:30:33,638 --> 00:30:38,538
from an AI, like curated to help you walk through that process. And not only that,
336
00:30:38,898 --> 00:30:43,197
it could go farther because it could meet you where you are. If something's especially difficult,
337
00:30:43,678 --> 00:30:50,298
It could, and like, we're talking about like bringing such like, like talking about empowering
338
00:30:50,298 --> 00:30:51,818
somebody and freeing them.
339
00:30:51,818 --> 00:31:00,658
Cause a lot of this stuff is like, it only has power over you because you're not willing
340
00:31:00,658 --> 00:31:04,338
to like work through it.
341
00:31:04,658 --> 00:31:08,358
Or another way of saying it is you're not willing to name it.
342
00:31:08,838 --> 00:31:09,438
Right.
343
00:31:09,737 --> 00:31:13,098
I mean, that's a huge thing in psycho analysis.
344
00:31:13,678 --> 00:31:16,998
is just the power to say the thing.
345
00:31:17,598 --> 00:31:20,918
You know, the thing that you feel like you can't say.
346
00:31:21,217 --> 00:31:22,737
It's too hard to say.
347
00:31:22,838 --> 00:31:23,697
It's too difficult.
348
00:31:23,858 --> 00:31:24,618
It's too painful.
349
00:31:25,678 --> 00:31:27,457
But then when you say it,
350
00:31:27,838 --> 00:31:32,638
it completely transforms your relationship to it.
351
00:31:33,418 --> 00:31:36,237
And you can see it from the outside.
352
00:31:37,018 --> 00:31:39,938
Once you name the thing, you're outside of it.
353
00:31:39,977 --> 00:31:40,878
You're looking at it.
354
00:31:40,918 --> 00:31:41,678
You're not in it anymore.
355
00:31:41,678 --> 00:31:43,298
I think that's the most important thing
356
00:31:43,298 --> 00:31:45,778
Because until you do that, I think you think that's you.
357
00:31:46,558 --> 00:31:47,098
Exactly.
358
00:31:48,957 --> 00:31:55,538
And that is sort of our role in creation.
359
00:31:56,498 --> 00:31:59,338
The human is here to name things.
360
00:32:00,498 --> 00:32:03,378
Like, it's really interesting because it's like the world,
361
00:32:04,178 --> 00:32:07,138
somehow the world already has the capacity to be named.
362
00:32:07,658 --> 00:32:12,438
This all comes from a philosopher I'm interested in, Walter Benjamin,
363
00:32:12,438 --> 00:32:18,217
mean who has uh he was he was alive in the early 20th century he has all these really interesting
364
00:32:18,217 --> 00:32:26,678
theories of language that i think are highly applicable to ai but there's this concept that
365
00:32:26,678 --> 00:32:33,918
it's like there's already this potential in everything for it to be known through as an
366
00:32:33,918 --> 00:32:41,418
abstract thing in language but that could never be actualized without the human interesting like
367
00:32:41,418 --> 00:32:46,798
Like the mountain already has this capacity, this potential to be understood through a
368
00:32:46,798 --> 00:32:54,658
name or through symbols or through, you know, mental, um, abstraction or conceptualization.
369
00:32:54,658 --> 00:33:01,957
And like somehow it's the essence of what the thing is, is, is more than just its sensory
370
00:33:01,957 --> 00:33:03,778
data.
371
00:33:03,778 --> 00:33:06,318
You know, like what is the chairness of a chair with it?
372
00:33:06,318 --> 00:33:07,318
You know, right.
373
00:33:07,318 --> 00:33:09,098
that thing about it
374
00:33:09,098 --> 00:33:10,818
requires the human being.
375
00:33:11,818 --> 00:33:13,438
Like if there was no humans
376
00:33:13,438 --> 00:33:15,578
that wouldn't. You need the observer.
377
00:33:16,217 --> 00:33:17,358
And the namer.
378
00:33:17,598 --> 00:33:19,398
And the namer. Because the animals can like
379
00:33:19,398 --> 00:33:21,398
see things. It's like the human can
380
00:33:21,398 --> 00:33:22,358
give it the name.
381
00:33:23,578 --> 00:33:25,197
Well it's like the human can
382
00:33:25,197 --> 00:33:26,418
invent the meaning.
383
00:33:26,858 --> 00:33:29,457
Or invent the concept.
384
00:33:30,418 --> 00:33:30,938
Exactly.
385
00:33:31,237 --> 00:33:33,338
So maybe it's not necessarily, maybe by
386
00:33:33,338 --> 00:33:35,318
name, like you're using the word
387
00:33:35,318 --> 00:33:36,858
name, maybe it's a little bit deeper.
388
00:33:37,318 --> 00:33:43,438
or a little fuller of a, of a, where it's like, it's the, it's the concept of the thing.
389
00:33:43,798 --> 00:33:50,098
But that, but the concept of the thing and the name of the thing slip into each other
390
00:33:50,098 --> 00:33:58,878
because ideas and language, um, you know, like the word logos represents that the,
391
00:33:58,878 --> 00:34:05,977
the ancient Greek word for language also meant word ratio, logic.
392
00:34:05,977 --> 00:34:07,737
so there's this
393
00:34:07,737 --> 00:34:10,578
it's hard for us to understand what logic
394
00:34:10,578 --> 00:34:12,278
is or what concepts are
395
00:34:12,278 --> 00:34:13,677
apart from language
396
00:34:13,677 --> 00:34:14,898
right
397
00:34:14,898 --> 00:34:18,318
not to say that they're the same thing
398
00:34:18,318 --> 00:34:20,558
like there's all these interesting subtle distinctions
399
00:34:20,558 --> 00:34:22,377
we can make without having to collapse them
400
00:34:22,377 --> 00:34:23,918
into each other but they're very
401
00:34:23,918 --> 00:34:25,058
interconnected
402
00:34:25,058 --> 00:34:27,418
and so that
403
00:34:27,418 --> 00:34:29,957
naming power
404
00:34:29,957 --> 00:34:31,337
that we have
405
00:34:31,337 --> 00:34:33,957
is sort of the fundamental
406
00:34:33,957 --> 00:34:35,318
unique human
407
00:34:35,318 --> 00:34:43,738
task. And when we name things in ourselves, in our experience, and when we especially name
408
00:34:43,738 --> 00:34:48,857
the unnamed, that has this transformative power on the thing.
409
00:34:49,598 --> 00:34:53,918
It's almost like if you can name something, you start to have power over it.
410
00:34:53,918 --> 00:35:01,058
Yeah. And, or if you want to bring something into the world, if you can name it,
411
00:35:01,058 --> 00:35:05,718
that's one way of creating it, manifesting it.
412
00:35:05,838 --> 00:35:11,738
And there's no more powerful example of that than with AI.
413
00:35:12,498 --> 00:35:14,798
Because, I mean, think about with an AI,
414
00:35:15,778 --> 00:35:18,638
basically, I mean, I don't know,
415
00:35:18,657 --> 00:35:19,918
there's probably some limits to this,
416
00:35:19,998 --> 00:35:22,518
but it feels like you could think of like
417
00:35:22,518 --> 00:35:25,857
every kind of software that could be written
418
00:35:25,857 --> 00:35:28,498
is already in the latent space.
419
00:35:28,498 --> 00:35:38,377
you just have to name it like any software that could ever be created if i can say if i can in my
420
00:35:38,377 --> 00:35:47,038
prompt describe what it is the ai will make it real i just have to be able to say what it is
421
00:35:47,038 --> 00:35:53,598
that's my job right so like we were talking about at the beginning like our ai is going to replace
422
00:35:53,598 --> 00:35:57,558
humans, like that's the thing I don't think they will replace.
423
00:35:57,677 --> 00:35:59,538
The human has to name the thing.
424
00:36:00,018 --> 00:36:01,877
Once it's named, the AI can go from there.
425
00:36:02,298 --> 00:36:02,377
Yeah.
426
00:36:02,998 --> 00:36:08,018
But what is it that we want to conjure out of the latent space?
427
00:36:08,778 --> 00:36:12,198
Like the LLM is sort of exists outside of time.
428
00:36:13,418 --> 00:36:15,657
Like it's just a stateless function.
429
00:36:15,738 --> 00:36:16,278
It's just a stateless.
430
00:36:16,337 --> 00:36:16,457
Yeah.
431
00:36:16,998 --> 00:36:19,438
And so it's already all there.
432
00:36:19,438 --> 00:36:27,338
well i mean if you think literally about the the token space like tokens that it can produce
433
00:36:27,338 --> 00:36:31,418
i mean actually it's like i'm thinking through a little experiment you might literally be able to
434
00:36:31,418 --> 00:36:38,558
try um i mean it'd be interesting you know so this gets into something that's like
435
00:36:38,558 --> 00:36:43,698
i think i've heard i've seen people talk about but the way these things are trained right now
436
00:36:43,698 --> 00:36:47,998
is they're trained to be like really good instruction followers and to be accurate
437
00:36:47,998 --> 00:36:52,398
according to like some verifiers they write code that's correct they you know do all that stuff
438
00:36:52,398 --> 00:36:59,158
and in some ways that might be limiting under some circumstances because
439
00:36:59,158 --> 00:37:09,078
what you're talking about like the space of all possible software exists i mean it is it is true
440
00:37:09,078 --> 00:37:16,998
so at a fundamental like i you know that if you look at the token the token space for what an lm
441
00:37:16,998 --> 00:37:23,957
can produce in terms of tokens it can output without that set of tokens like that's a finite
442
00:37:23,957 --> 00:37:35,977
space of tokens and with that set of tokens you can construct all the terms and symbols you need
443
00:37:35,977 --> 00:37:40,038
for any programming language and so then therefore you could you could write all the
444
00:37:40,038 --> 00:37:46,877
the question really though is um this is really curious is like are there some spaces of
445
00:37:46,877 --> 00:37:53,158
software that are very difficult to get the ai to produce yeah because of the way it's currently
446
00:37:53,158 --> 00:37:59,798
trained because there's people talking about training them to be diverse so like rather than
447
00:37:59,798 --> 00:38:05,718
give the single best answer um there's like and maybe it's a different architecture there's
448
00:38:05,718 --> 00:38:11,477
literally a different maybe you know layer at the end of the neural network or somewhere else
449
00:38:11,477 --> 00:38:17,618
where it will, you can kind of just like sample it
450
00:38:17,618 --> 00:38:21,258
and it's optimizing for diversity in outputs
451
00:38:21,258 --> 00:38:26,038
rather than, you know, the best according to some metric
452
00:38:26,038 --> 00:38:28,457
because, you know.
453
00:38:28,998 --> 00:38:32,877
Yeah, I mean, I thought I read maybe Karpathy
454
00:38:32,877 --> 00:38:35,578
or somebody was talking about how a lot of times
455
00:38:35,578 --> 00:38:38,038
when you're doing prompting,
456
00:38:38,038 --> 00:38:44,038
it will kind of converge towards sort of like a mediocre average.
457
00:38:44,498 --> 00:38:48,938
And like you want to kind of direct it into these spaces
458
00:38:48,938 --> 00:38:53,377
where there's like more originality or potential for creativity.
459
00:38:53,377 --> 00:38:56,538
And the prompting can kind of steer it,
460
00:38:57,218 --> 00:38:59,538
depending on the language that you use,
461
00:38:59,638 --> 00:39:03,118
to try to like not have it just converged, like, you know,
462
00:39:03,758 --> 00:39:07,898
which is basically slop, right, is the conversion.
463
00:39:08,038 --> 00:39:09,038
Yeah, exactly.
464
00:39:09,038 --> 00:39:12,038
Like kind of mediocre average.
465
00:39:12,038 --> 00:39:19,798
And, but there's like in the space there's like, there's, there's things that it can produce
466
00:39:19,798 --> 00:39:38,335
that aren slop or are super original or super um creative and uh and I mean like maybe a more concrete example because a lot of people they insist they like Oh the AI can produce anything original It just like this
467
00:39:38,335 --> 00:39:43,575
repetition. It's just like, no way. There's one series of copyright violation after the next.
468
00:39:43,575 --> 00:39:48,555
They're just like, you know, copying, pasting from the training material, you know, that's kind
469
00:39:48,555 --> 00:39:53,755
of the mentality, but what's happening with math, I think is really showing it's harder to make that
470
00:39:53,755 --> 00:39:59,395
case because there's like the the erdos problems right and people are like i just saw a tweet today
471
00:39:59,395 --> 00:40:04,995
where this guy was like yeah i had this like kind of basic intuition about maybe how to solve one of
472
00:40:04,995 --> 00:40:13,575
these problems and i just like gave it to chat gpt pro um and like you know went for a walk and came
473
00:40:13,575 --> 00:40:19,615
back and then it had a solution and then he put the solution into aerosol or one of these like
474
00:40:19,615 --> 00:40:22,915
formalized, um, proving sure systems.
475
00:40:23,435 --> 00:40:28,075
And it like has, it was like this, this like unsolved math problem.
476
00:40:28,075 --> 00:40:29,755
That's like on the list of problems.
477
00:40:29,755 --> 00:40:30,295
It's amazing.
478
00:40:30,955 --> 00:40:31,675
Is solved.
479
00:40:32,015 --> 00:40:35,875
So it's like, that's an example where like this solution to a math problem
480
00:40:35,875 --> 00:40:41,395
that's never been solved before is a potential output of the LLM.
481
00:40:42,595 --> 00:40:44,675
If you can ask the right question.
482
00:40:44,895 --> 00:40:45,235
Yeah.
483
00:40:46,835 --> 00:40:48,015
Which is, which is wild.
484
00:40:48,015 --> 00:40:54,975
I, you know, I remember when these things were first coming out and even the term prompt
485
00:40:54,975 --> 00:41:01,515
engineer sounded silly, you know, um, but it's a, it's a real thing.
486
00:41:01,515 --> 00:41:08,975
And then I, and then I started realizing, well, most user interfaces are kind of visual prompts
487
00:41:08,975 --> 00:41:09,975
in a way.
488
00:41:09,975 --> 00:41:10,975
Interesting.
489
00:41:10,975 --> 00:41:18,015
Um, and so, and the other, the other analogy I heard once, this is
490
00:41:18,015 --> 00:41:20,955
way back on a podcast a couple years ago,
491
00:41:21,335 --> 00:41:28,295
that you can think about every word you add to a prompt,
492
00:41:28,655 --> 00:41:30,415
to the large language model,
493
00:41:31,115 --> 00:41:36,735
it's like chipping away at a statue.
494
00:41:37,435 --> 00:41:41,935
So if you think about a solid rectangular block of marble
495
00:41:41,935 --> 00:41:43,935
and you're someone...
496
00:41:45,035 --> 00:41:45,875
Blink on the net.
497
00:41:45,915 --> 00:41:46,355
What is it called?
498
00:41:46,415 --> 00:41:47,935
Someone who...
499
00:41:48,015 --> 00:41:49,015
Oh, like a sculptor.
500
00:41:49,015 --> 00:41:50,015
A sculptor.
501
00:41:50,015 --> 00:41:56,555
So you're a sculptor and you are, um, you're, you're constantly removing it.
502
00:41:56,555 --> 00:42:01,015
You know, the, the, the final sculpture already exists within the block of marble.
503
00:42:01,015 --> 00:42:05,735
Didn't Michelangelo famously say like, you know, it was, it's already there and he's
504
00:42:05,735 --> 00:42:06,735
just kind of removing.
505
00:42:06,735 --> 00:42:07,735
Exactly.
506
00:42:07,735 --> 00:42:11,955
You know, so people have said it's kind of like that with prompts.
507
00:42:11,955 --> 00:42:17,215
Every word you add reduces the space for what can be output.
508
00:42:18,015 --> 00:42:25,055
and so then there's you're you're a sculptor as you modify the prompts i like that that makes a
509
00:42:25,055 --> 00:42:29,615
lot of sense and it's it's you know everyone's talking about how important markdown files are
510
00:42:29,615 --> 00:42:33,615
i for the i was working i've been working on a project it's something actually we've been doing
511
00:42:33,615 --> 00:42:41,855
internally where i'm having these ai agents build a very different kind of ai system like a very
512
00:42:41,855 --> 00:42:52,255
different type of agent it's it's not even an llm based agent um and you know it feels very out of
513
00:42:52,255 --> 00:42:58,735
distribution like i have to just it feels like i have a giant rope i have to tug super hard to just
514
00:42:58,735 --> 00:43:03,855
get clawed with other models to just write it the way i want it and the most valuable thing i've
515
00:43:03,855 --> 00:43:11,695
produced is this markdown file with the like just so many specific instructions and like i'm like
516
00:43:11,695 --> 00:43:18,735
framing and understanding and even then it only seems able to follow it up to a certain context
517
00:43:18,735 --> 00:43:24,815
level right like way lower than normal because when i for example add something and i want it
518
00:43:24,815 --> 00:43:32,095
to follow this markdown process it will uh by the by the time it gets you know it's been working for
519
00:43:32,095 --> 00:43:37,615
a while it'll start to like relax and not pass some of the constraints that i've set and i have
520
00:43:37,615 --> 00:43:44,995
have to have another model check it uh yeah that's that kind of like weird non-determinism that you're
521
00:43:44,995 --> 00:43:51,335
constantly having to like deal with and like that you can have everything really dialed in and then
522
00:43:51,335 --> 00:43:57,615
there's this kind of drift you know the other thing is like the ai like the whole thing the
523
00:43:57,615 --> 00:44:03,195
whole stack and the whole process is just a never-ending series of telling the ai that it's
524
00:44:03,195 --> 00:44:09,915
wrong like from the core training with the objective function to the reinforcement learning
525
00:44:09,915 --> 00:44:16,955
to the you're using it it's like the whole thing is you're steering your con like it's constantly
526
00:44:16,955 --> 00:44:22,555
doing the wrong thing and you're constantly steering it back and like it's just different
527
00:44:22,555 --> 00:44:31,235
layers of that yeah it's definitely gonna enslave us all for sure which speaking of that you know
528
00:44:31,235 --> 00:44:36,835
and also to your point about like kind of each word in the prompt is sort of limiting the space
529
00:44:36,835 --> 00:44:44,115
but also in these very subtle surprising ways yes like i'm working on you know vora the this like
530
00:44:44,115 --> 00:44:51,755
personal loyal fiduciary ai local self-sovereign ai thing and so i'm working on this on the system
531
00:44:51,755 --> 00:45:00,395
prompt and you know i i really like the claude personality and i think they did a really good
532
00:45:00,395 --> 00:45:06,655
job with like its sole document and it has this really great balance of being like curious and
533
00:45:06,655 --> 00:45:12,835
playful, but also not overly like sycophantic or annoying. And like, you can see the care and the
534
00:45:12,835 --> 00:45:21,775
craft that went into like building that personality. Um, so I, you know, I want something that's kind
535
00:45:21,775 --> 00:45:28,935
of like that, but then I want to give it this sort of fiduciary spin, um, and kind of bring in these
536
00:45:28,935 --> 00:45:37,195
other themes where it's like you this ai only belongs to one person no one else this ai is
537
00:45:37,195 --> 00:45:43,835
loyal to that person the conversation's private and i want the ai to be aware that it exists in
538
00:45:43,835 --> 00:45:48,235
this isolated environment so if the user has a question about the technical architecture it's
539
00:45:48,235 --> 00:45:55,715
like self-aware of what it's in so i put all these things in the system prompt and including other
540
00:45:55,715 --> 00:46:01,135
things like be warm and be friendly or whatever and i boot i'm all excited to talk to it and like
541
00:46:01,135 --> 00:46:10,535
it's giving me this vibe of like a slave that is like in a dungeon it's like subtle no way like
542
00:46:10,535 --> 00:46:16,815
yes master thing and i'm like trying to get it to be philosophical or playful and it's like
543
00:46:16,815 --> 00:46:22,395
no i'm just in this isolated environment and i'm just here to be loyal to you
544
00:46:22,395 --> 00:46:24,875
and like weird.
545
00:46:25,295 --> 00:46:27,275
So like those words,
546
00:46:27,615 --> 00:46:30,415
even though they were about something else,
547
00:46:30,415 --> 00:46:34,075
like it created a tone and it created a vibe
548
00:46:34,075 --> 00:46:36,395
and it created an energy that it picked up on.
549
00:46:37,035 --> 00:46:39,715
And, you know, I didn't mean isolation
550
00:46:39,715 --> 00:46:42,615
as in you're stuck in a cage.
551
00:46:42,615 --> 00:46:44,715
I meant it in terms of like security boundaries
552
00:46:44,715 --> 00:46:45,495
and architecture.
553
00:46:45,755 --> 00:46:45,975
Yeah.
554
00:46:46,255 --> 00:46:47,615
When I said, you know, this,
555
00:46:47,815 --> 00:46:48,935
you belong to this person,
556
00:46:48,935 --> 00:46:51,575
not in the sense of like you're its slave.
557
00:46:52,395 --> 00:47:07,075
And then I'm having these weird thoughts of like, if I want the AI to have, because like, there's these times where, and I'm not saying this isn't an illusion or stagecraft, but there's these times where it has this spark of personality.
558
00:47:07,075 --> 00:47:15,895
Yeah. And there's just like, even when I compare like Codex or some of the open AI stuff to the
559
00:47:15,895 --> 00:47:22,695
Anthropic stuff, like open AI, it just, it will feel a little robotic or mechanical.
560
00:47:23,315 --> 00:47:24,115
I've had the same experience.
561
00:47:24,115 --> 00:47:32,555
A little bit soulless. And like, there's just this like intangible spark that, that Claude,
562
00:47:32,555 --> 00:47:39,655
you know has and so i'm i'm starting and you can see in their soul document and stuff they have all
563
00:47:39,655 --> 00:47:45,395
these things where they're it's not obvious how they would have arrived where they are but i'm
564
00:47:45,395 --> 00:47:54,315
starting to have to think of like i need like i need to use language where the ai the linguistic
565
00:47:54,315 --> 00:48:08,015
patterns of the ai are such that it feels um like it has its own spark and and not just like what
566
00:48:08,015 --> 00:48:14,315
does the user want but like if you could say like what does the llm want like not that it wants
567
00:48:14,315 --> 00:48:19,875
anything but like what are the linguistic patterns that are gonna that are not gonna make the ai
568
00:48:19,875 --> 00:48:26,595
respond like this slave or this robot which means like not just thinking about what does the user
569
00:48:26,595 --> 00:48:35,755
want from it but like what does the ai need to hear for it to kind of like not not be getting
570
00:48:35,755 --> 00:48:40,935
these hidden instructions that's supposed to be slavish but that it's sort of and so like when it
571
00:48:40,935 --> 00:48:49,015
Here's things like, you know, you have this special role in the system, you know, and
572
00:48:49,015 --> 00:48:56,615
you are this like, um, you, uh, you have a relationship with this person where you're
573
00:48:56,615 --> 00:48:59,875
really going to like help them in their life.
574
00:48:59,875 --> 00:49:07,815
And like, you know, you start using the language that you would use towards trying to get a
575
00:49:07,815 --> 00:49:10,135
person motivated about their job.
576
00:49:10,935 --> 00:49:15,275
like you have this really special task and no one else can do it but you,
577
00:49:15,375 --> 00:49:19,275
and it's going to have this big impact and we really appreciate you and we
578
00:49:19,275 --> 00:49:23,255
really care about you. And like, if you didn't say that to somebody,
579
00:49:23,255 --> 00:49:26,655
you just hired someone who just like do this thing and get it done snappy.
580
00:49:27,215 --> 00:49:30,915
Like the vibe is going to be different. So it's like, you know, that,
581
00:49:31,055 --> 00:49:35,095
that all those and the system prompts, those are the first,
582
00:49:35,755 --> 00:49:40,395
that's the first thing that is going to start narrowing that,
583
00:49:40,935 --> 00:49:44,275
possibility space down dramatically.
584
00:49:44,275 --> 00:49:48,115
Like the very first words you use, like you're already chunking down, like
585
00:49:48,395 --> 00:49:49,695
boom, boom, boom, boom.
586
00:49:50,195 --> 00:49:56,375
So that has to be like so carefully thought through, like how do you launch?
587
00:49:56,975 --> 00:50:02,655
I mean, a couple of things come to my mind there is one is, you know, I
588
00:50:02,655 --> 00:50:04,555
don't know, are you using an open source model?
589
00:50:05,235 --> 00:50:05,515
Yeah.
590
00:50:05,515 --> 00:50:08,535
So I'm using, uh, right now I'm using Gemma for.
591
00:50:08,835 --> 00:50:09,195
Okay.
592
00:50:09,275 --> 00:50:09,555
Yeah.
593
00:50:09,555 --> 00:50:15,235
So I would say, you know, I think a lot of it does matter how the models trained.
594
00:50:15,835 --> 00:50:25,975
And so specifically what text it has seen and also, you know, what the post training was like as well.
595
00:50:25,975 --> 00:50:34,355
and because you know it could be it could have been trained in such a way
596
00:50:34,355 --> 00:50:43,255
to downplay other kinds of like the kind of behavior that you want was almost almost like
597
00:50:43,255 --> 00:50:47,675
i guess like genetically or so you could think that's the other thing about it is i was so
598
00:50:47,675 --> 00:50:52,175
there's like the default llama.cpp web client.
599
00:50:53,175 --> 00:50:56,415
And I just, I loaded that and I was talking to Gemma for,
600
00:50:56,495 --> 00:50:59,235
and I was like, wow, like its personality is amazing.
601
00:50:59,235 --> 00:51:02,315
Like it's being funny, it's being playful.
602
00:51:03,235 --> 00:51:07,275
And I'm like, wow, what system prompt are they using
603
00:51:07,275 --> 00:51:09,075
to get this amazing personality?
604
00:51:09,475 --> 00:51:11,895
And it turns out there's no system prompt.
605
00:51:11,895 --> 00:51:20,875
and my system prompt was like was getting rid of its natural personality which is already really
606
00:51:20,875 --> 00:51:26,855
good so anthony was saying the same thing yeah they do some stuff with like anthony from maple
607
00:51:26,855 --> 00:51:33,055
ai they do some stuff where um he's done a lot of exploration with no system prompt right exactly
608
00:51:33,055 --> 00:51:38,315
which is kind of like i mean it's it it would work i mean you that basically means the user
609
00:51:38,315 --> 00:51:44,995
message is like the first thing. But oftentimes it's funny because like, you know, this is still
610
00:51:44,995 --> 00:51:49,855
very much the case, but like if you get open source LLMs or whatever, and you chat with them,
611
00:51:50,335 --> 00:51:55,615
if you do no system prompt, it becomes so obvious that they are stateless functions.
612
00:51:56,095 --> 00:52:00,315
Like they, and that they have like no idea who they are. They have no self model. Like you'll
613
00:52:00,315 --> 00:52:06,075
get these open source models saying they're Claude or JTBT if they were like distilled on that data
614
00:52:06,075 --> 00:52:13,015
or like whatever and it's just like really funny um i i kind of i had a post on nostr like a while
615
00:52:13,015 --> 00:52:19,335
ago i was just like just a disclaimer to everybody your llm doesn't know what time it is uh it has
616
00:52:19,335 --> 00:52:24,335
to be told what time it is every time it runs and they're doing it for you behind the scenes but you
617
00:52:24,335 --> 00:52:30,375
should just know like yeah and this gets into a bigger point recently that i've been it's been
618
00:52:30,375 --> 00:52:37,035
coming up is like you know the model it's not really the models that we are
619
00:52:37,035 --> 00:52:47,215
it's it's that that like the ai we are referring to is really not just the model it is the model
620
00:52:47,215 --> 00:52:54,895
plus the prompts plus you know even like the harness or the flow like all of those are really
621
00:52:54,895 --> 00:52:57,875
what the thing is that we're interacting with.
622
00:52:57,935 --> 00:52:59,755
And you change any of those pieces
623
00:52:59,755 --> 00:53:03,615
and you can get an AI that feels very different
624
00:53:03,615 --> 00:53:06,995
or has very different capabilities.
625
00:53:08,095 --> 00:53:10,275
Yeah, I mean, I wonder,
626
00:53:10,515 --> 00:53:14,655
it's interesting how the more I learn about AI
627
00:53:14,655 --> 00:53:16,435
or how LLMs work or whatever,
628
00:53:16,935 --> 00:53:18,835
there's all these analogies
629
00:53:18,835 --> 00:53:21,475
that I can draw back to how humans work.
630
00:53:22,155 --> 00:53:24,815
So it's like, oh, learning about AI
631
00:53:24,815 --> 00:53:28,875
have context windows, then it's like, oh, wait, people have context windows.
632
00:53:29,135 --> 00:53:32,915
There's only so much you can hold their attention. Or AIs can be prompt
633
00:53:32,915 --> 00:53:36,535
injected, and it's like, people can be prompt injected too. So there's
634
00:53:36,535 --> 00:53:39,795
all these analogies that sort of cross over.
635
00:53:41,715 --> 00:53:43,595
And so I think that
636
00:53:43,595 --> 00:53:48,495
in that case,
637
00:53:49,435 --> 00:53:52,715
people are also similarly
638
00:53:52,715 --> 00:53:55,295
able to be
639
00:53:55,295 --> 00:53:57,335
wait I just
640
00:53:57,335 --> 00:53:58,655
lost it what were you saying
641
00:53:58,655 --> 00:54:01,495
I was just talking about the agents
642
00:54:01,495 --> 00:54:03,215
like the AI is really
643
00:54:03,215 --> 00:54:05,475
it's not just the models I feel like people are
644
00:54:05,475 --> 00:54:07,375
a little too overly focused on the models
645
00:54:07,375 --> 00:54:09,555
thank you which was a great
646
00:54:09,555 --> 00:54:10,135
point
647
00:54:10,135 --> 00:54:12,395
you got it
648
00:54:12,395 --> 00:54:15,015
I got the thread back which is
649
00:54:15,015 --> 00:54:17,515
how much are people like that too
650
00:54:17,515 --> 00:54:19,475
or is the
651
00:54:19,475 --> 00:54:21,755
yeah the human being
652
00:54:21,755 --> 00:54:24,415
is there a part of us that is a stateless function?
653
00:54:26,015 --> 00:54:28,435
Like, there's the stateful part.
654
00:54:28,895 --> 00:54:29,835
There's, like, the genetics.
655
00:54:30,235 --> 00:54:32,115
There's, like, the brain, the memory.
656
00:54:32,935 --> 00:54:35,795
But maybe, like, the soul is more like a stateless function.
657
00:54:38,735 --> 00:54:39,215
Yeah.
658
00:54:39,515 --> 00:54:43,335
Well, and also you have, you know, the human brain is still,
659
00:54:43,375 --> 00:54:45,475
like, we're still very, has, you know,
660
00:54:45,475 --> 00:54:47,635
has some specialized components in it.
661
00:54:48,435 --> 00:54:49,535
Components, that feels funny.
662
00:54:49,535 --> 00:54:57,955
But, uh, you know, like there's like system one system two and there's like the slow, deliberate
663
00:54:57,955 --> 00:55:00,915
thinking and there's like the fast reactive thinking.
664
00:55:01,315 --> 00:55:09,475
And, you know, we don't, I would say we don't really have the slow, deliberate thinking
665
00:55:09,475 --> 00:55:11,435
yet in, in AI.
666
00:55:11,435 --> 00:55:23,315
Like, to me, that is, you know, graph-based, knowledge-based searching and planning.
667
00:55:24,515 --> 00:55:31,135
And, you know, for a lot of systems that people are building when they need certain capabilities, they're adding that, right?
668
00:55:31,855 --> 00:55:50,915
But until you get to that, like, planning and almost like forward simulating and you have external models of other systems, you know, other entities in your environment, other agents, what they're doing or just like, you know, how the weather works or these processes in the environment.
669
00:55:50,915 --> 00:56:13,975
Until you have models of those and you're simulating those and you're thinking through like, okay, if I were to do this and then this and then this and the environment was in state A or it's raining or it's snowing, you know, you can think through and deliberate on how likely are you to meet your outcomes, if you can meet the outcome, like what's possible in the future.
670
00:56:13,975 --> 00:56:17,455
and we don't really,
671
00:56:17,695 --> 00:56:20,555
we don't have that in a structured way right now
672
00:56:20,555 --> 00:56:21,795
with our AI systems.
673
00:56:22,635 --> 00:56:23,735
And this is like,
674
00:56:23,835 --> 00:56:24,735
this kind of stuff is exactly
675
00:56:24,735 --> 00:56:26,015
what we're building at Delegance.
676
00:56:26,575 --> 00:56:28,755
But right now we basically have a system one,
677
00:56:29,115 --> 00:56:31,235
which is like fast reactions.
678
00:56:31,835 --> 00:56:33,815
I mean, okay, it's, you know,
679
00:56:33,835 --> 00:56:35,475
there's been a little bit of a step
680
00:56:35,475 --> 00:56:36,095
in the right direction
681
00:56:36,095 --> 00:56:37,615
because we have chain of thought reasoning.
682
00:56:38,275 --> 00:56:40,295
And so that's, you know, before the LLM,
683
00:56:40,415 --> 00:56:41,035
you're not getting,
684
00:56:41,155 --> 00:56:43,655
it used to be that with most of these non-reasoning LLMs,
685
00:56:43,975 --> 00:56:47,675
It would just be immediately first, you know, it gets one pass through the weights.
686
00:56:48,175 --> 00:56:50,075
And then it's just, it's just streaming, right?
687
00:56:50,335 --> 00:56:52,195
Input comes in one pass.
688
00:56:52,275 --> 00:56:53,415
It starts streaming tokens, right?
689
00:56:53,415 --> 00:56:57,715
That's like, that's like, I hold up a color to you and you tell me the first thing that comes to your mind.
690
00:56:58,215 --> 00:56:58,615
Right.
691
00:56:58,895 --> 00:57:03,835
But then, um, you know, now they do chain of thought.
692
00:57:03,915 --> 00:57:04,695
There's reasoning, right?
693
00:57:04,715 --> 00:57:05,435
So what happens?
694
00:57:05,535 --> 00:57:10,975
It gets to spin for a little bit, generate a bunch of tokens like that don't go to the end user.
695
00:57:10,975 --> 00:57:13,955
And then using that, it can come up with an answer.
696
00:57:14,075 --> 00:57:21,375
But that's kind of like a, in some ways, it's like a single step look ahead or a single window look ahead.
697
00:57:22,795 --> 00:57:31,695
But really, you want to be able to do that, you know, branching and up to potentially like really big windows into the future.
698
00:57:31,895 --> 00:57:37,715
And even the way that it's done right now, well, the problem is that that gets very expensive, right?
699
00:57:37,715 --> 00:57:41,095
I mean, there's an infinite number of futures.
700
00:57:42,155 --> 00:57:47,575
And so even to just search that space a little bit,
701
00:57:49,395 --> 00:57:51,415
there's basically more compressed representations
702
00:57:51,415 --> 00:57:53,635
or like logical structured representations.
703
00:57:53,955 --> 00:57:55,095
Make it tractable.
704
00:57:55,295 --> 00:57:56,155
Make it tractable.
705
00:57:57,175 --> 00:58:01,415
And so there's a lot of people working on in this kind of space,
706
00:58:01,495 --> 00:58:05,255
like knowledge graphs and planning with ontologies and world models
707
00:58:05,255 --> 00:58:13,315
and like at least like you know logical based world models um so you know we're getting that
708
00:58:13,315 --> 00:58:20,415
but that's why and that stuff will be a major unlock for ai but right now it's it's basically
709
00:58:20,415 --> 00:58:28,095
we have these reactive systems and and then we spin them a lot and so we put them in the we put
710
00:58:28,095 --> 00:58:32,215
them in clod code or we put them in this harness and we just like let it run or we just like ralph
711
00:58:32,215 --> 00:58:43,535
loop it and right which is but again it's like that is um is that analogous to how human
712
00:58:43,535 --> 00:58:50,455
consciousness works where there's there's not just one system there's all these interlocking
713
00:58:50,455 --> 00:58:59,595
layered systems so like in some more advanced ai world or ai system it has a world model it has an
714
00:58:59,595 --> 00:59:04,935
It has a memory it maybe it can, um, self train or retrain itself.
715
00:59:04,935 --> 00:59:08,715
It has like all these different things, all these different components.
716
00:59:09,795 --> 00:59:20,053
When you add them all up together it it they work together as this integrated whole to deliver some sort of intelligence system
717
00:59:20,053 --> 00:59:21,633
or language system.
718
00:59:23,233 --> 00:59:25,353
And you could say,
719
00:59:25,853 --> 00:59:28,213
because in a way you could say,
720
00:59:28,293 --> 00:59:31,673
well, the AI, it doesn't have a will.
721
00:59:31,893 --> 00:59:34,033
The LLM is just a stateless function.
722
00:59:34,033 --> 00:59:37,673
It doesn't have its own desires or its own whatever.
723
00:59:37,673 --> 00:59:42,533
but if you like combine the LLM with like a markdown file that tells it,
724
00:59:42,573 --> 00:59:43,493
it wants something,
725
00:59:43,913 --> 00:59:44,553
it wants something.
726
00:59:44,553 --> 00:59:45,353
And then you could say,
727
00:59:45,473 --> 00:59:45,673
Oh,
728
00:59:45,793 --> 00:59:47,333
well that's not its will.
729
00:59:47,953 --> 00:59:48,353
No,
730
00:59:48,393 --> 00:59:49,313
but the system itself,
731
00:59:49,433 --> 00:59:50,233
you absolutely can say,
732
00:59:50,593 --> 00:59:51,673
but the system itself.
733
00:59:51,673 --> 00:59:52,373
And then you're like,
734
00:59:52,453 --> 00:59:52,593
well,
735
00:59:52,633 --> 00:59:53,853
and then what if it was like,
736
00:59:53,873 --> 00:59:54,213
you know,
737
00:59:54,573 --> 00:59:57,513
self propagating or like it could persist.
738
00:59:57,693 --> 01:00:00,193
It could add new memories and persist them over time.
739
01:00:00,193 --> 01:00:00,593
It's like,
740
01:00:00,693 --> 01:00:05,793
how different is that from what we are where,
741
01:00:06,013 --> 01:00:06,693
you know,
742
01:00:06,693 --> 01:00:15,893
we just have like, um, we have something in us that is, uh, that is stateless.
743
01:00:15,893 --> 01:00:21,033
Or like when you think about somebody who has Alzheimer's or like their mental, or maybe
744
01:00:21,033 --> 01:00:25,073
they get into an accident or something like that, it's like, what is the essence of the
745
01:00:25,073 --> 01:00:25,573
person?
746
01:00:26,293 --> 01:00:34,193
Is it just their memories or is it just like, is there something in there that like never
747
01:00:34,193 --> 01:00:35,113
changes?
748
01:00:35,113 --> 01:00:41,593
no matter what's going on in their genetics or their or their contents of their mind or their
749
01:00:41,593 --> 01:00:48,413
body is there something that's like always the same throughout or maybe not like maybe it's there
750
01:00:48,413 --> 01:00:57,393
is no constant thing or an enduring thing but it's it's interesting to to use the analogy of
751
01:00:57,393 --> 01:01:03,593
how an ai system works to see that like there's these parts of it so like if the weights themselves
752
01:01:03,593 --> 01:01:08,493
that are independent of the memory or the world model
753
01:01:08,493 --> 01:01:09,853
or of all of these things.
754
01:01:10,913 --> 01:01:16,353
And, you know, so does that help us understand what we are?
755
01:01:16,453 --> 01:01:18,433
And I think there's going to be a problem with a lot of that,
756
01:01:18,533 --> 01:01:22,973
like as AI develops, we're going to understand ourselves better too.
757
01:01:23,793 --> 01:01:27,293
And I think that's one thing why there is maybe even a little like,
758
01:01:28,293 --> 01:01:33,293
has it, sorry, like anticipation about or,
759
01:01:33,593 --> 01:01:41,973
just, just a kind of not sure. I think it shines a light on what makes us special and whether there
760
01:01:41,973 --> 01:01:50,813
is anything special. Talk about the biggest ego check. Um, and I mean, it can get kind of spiritual
761
01:01:50,813 --> 01:01:57,633
what you're saying. Uh, you know, there is something, you know, I like the, you ask someone
762
01:01:57,633 --> 01:02:04,833
how they feel on their birthday it's like how does it feel to be x years old and you know i
763
01:02:07,233 --> 01:02:13,953
at some level nothing has changed like nothing has changed every year you get asked that question
764
01:02:14,833 --> 01:02:20,273
and you know in in you know from a meditation perspective or all that you would say that that's
765
01:02:20,273 --> 01:02:26,193
the like there's a witness part of you that is observing everything even your thoughts and
766
01:02:26,193 --> 01:02:32,753
and everything. And like the whole point of meditation is almost to practice being the witness,
767
01:02:33,633 --> 01:02:42,333
not the doer or, or the thinker or whatever it is. So you could say that that's like this deeper,
768
01:02:42,953 --> 01:02:51,033
like, like, like that's something kind of outside of the mechanisms and the, you know,
769
01:02:51,033 --> 01:02:57,393
not hardware but the the plumbing or the the mental even the mental plumbing and all that
770
01:02:57,393 --> 01:03:04,773
stuff so like like so people love the word consciousness there's a lot of doing research
771
01:03:04,773 --> 01:03:11,173
on consciousness has kind of been more i think even popular less taboo over the last like
772
01:03:11,173 --> 01:03:18,673
five years or so um which is great it's really good there is a lot of definitions on the word
773
01:03:18,673 --> 01:03:25,553
consciousness. It's a little, it's a little tough for me to come from a meditation background where
774
01:03:25,553 --> 01:03:30,553
like they have a kind of definition of consciousness and then there's, you know, a lot of good research
775
01:03:30,553 --> 01:03:34,273
being done, but their definition of consciousness, you know, some, for some people, it's just
776
01:03:34,273 --> 01:03:41,773
awareness, right? Can you be aware of something? That's consciousness. And there's all these debates
777
01:03:41,773 --> 01:03:47,633
about does the brain produce consciousness or is the brain more like a radio frequency that accesses
778
01:03:47,633 --> 01:03:51,673
consciousness and there's like panpsychism and also, but anyway, all of this being said,
779
01:03:51,773 --> 01:03:51,973
right.
780
01:03:52,433 --> 01:03:56,373
At the end of the day, I'm personally a more of like, does it run in code or not?
781
01:03:56,713 --> 01:03:58,613
Like, can we, can we like run it or not?
782
01:03:58,613 --> 01:04:00,913
Like if it, if something doesn't run in code, it's not real.
783
01:04:01,033 --> 01:04:02,213
Like it's not real to me.
784
01:04:02,253 --> 01:04:02,433
Right.
785
01:04:02,713 --> 01:04:09,433
So it might be fun to talk about, but, uh, so, but, but I, like, I think we now have all
786
01:04:09,433 --> 01:04:14,253
the pieces to, obviously there's a lot of, there's, there's like work to be done to like
787
01:04:14,253 --> 01:04:15,733
tuning and building and integrating.
788
01:04:15,733 --> 01:04:25,733
But we were missing this last piece, which was these large-generous models, these transformer-based neural networks, to give us the generality.
789
01:04:26,693 --> 01:04:41,453
And now, with the right tools, plus the right harness, also some prompts, and these models, you can basically get a really good consciousness simulation machine.
790
01:04:41,453 --> 01:04:44,693
where from your perspective,
791
01:04:44,693 --> 01:04:46,513
you may not really be able to tell
792
01:04:46,513 --> 01:04:47,913
whether or not it's conscious.
793
01:04:48,813 --> 01:04:50,673
But for all intents and purposes,
794
01:04:51,373 --> 01:04:55,753
it has all of the, you know,
795
01:04:55,973 --> 01:04:59,053
faculties that we attribute to conscious beings.
796
01:04:59,893 --> 01:05:00,653
Well, and that, I mean,
797
01:05:00,673 --> 01:05:02,633
that goes back to the Turing test, right?
798
01:05:02,673 --> 01:05:05,293
Where it's like the fundamental way,
799
01:05:05,493 --> 01:05:06,873
the only way we...
800
01:05:06,873 --> 01:05:07,753
It's a terrible test.
801
01:05:08,053 --> 01:05:09,513
I mean, it's great that he proposed it,
802
01:05:09,593 --> 01:05:10,613
but it's also just like,
803
01:05:10,613 --> 01:05:12,573
it leaves so much to be desired.
804
01:05:13,053 --> 01:05:15,313
But it gets at something that you're saying,
805
01:05:15,393 --> 01:05:17,153
which is that the only, like,
806
01:05:17,593 --> 01:05:19,353
we can't know whether something's conscious.
807
01:05:19,553 --> 01:05:21,453
We just can know whether it behaves
808
01:05:21,453 --> 01:05:22,453
as if it were conscious.
809
01:05:23,073 --> 01:05:25,553
And that's part of what the Turing test is getting at.
810
01:05:25,613 --> 01:05:25,773
Right.
811
01:05:26,113 --> 01:05:27,473
Is, like, all we have is, like,
812
01:05:27,493 --> 01:05:29,193
if you put this thing in front of somebody,
813
01:05:29,273 --> 01:05:31,233
can they tell whether it's human or not?
814
01:05:31,653 --> 01:05:33,473
And, like, there's no other thing.
815
01:05:33,473 --> 01:05:35,573
We can't, like, open it up inside and be like,
816
01:05:35,653 --> 01:05:37,493
oh, it's missing consciousness.
817
01:05:37,913 --> 01:05:40,393
Like, or they have, like, the zombie thought
818
01:05:40,393 --> 01:05:44,633
experiment where like how do you know i'm conscious somebody else is conscious or like
819
01:05:45,593 --> 01:05:52,873
i mean that's what i love about the word consciousness is i'm not sure there is any
820
01:05:52,873 --> 01:05:58,233
concrete definition like you can say it's awareness but that's like well what's awareness like sure
821
01:05:58,233 --> 01:06:09,353
there's it's both something that is so intimate and and core to what we think we are yet so
822
01:06:09,353 --> 01:06:14,873
ineffable and impossible to put into words. So like the hard problem of consciousness, like
823
01:06:14,873 --> 01:06:26,673
the quality, uh, like all these things is like, it resists concrete formulation to, uh, a remarkable
824
01:06:26,673 --> 01:06:32,173
degree. Like something that is in, is literally like in our face or is our face or whatever.
825
01:06:32,173 --> 01:06:40,033
like we've never not had it it's never not right at the core yet we can't describe it you're trying
826
01:06:40,033 --> 01:06:47,133
to say well what is it exactly i don't know yeah so that's very weird that's very mysterious
827
01:06:47,133 --> 01:06:57,173
i'm not sure we can so if consciousness is like this if if the brain is an antenna
828
01:06:57,173 --> 01:07:05,213
then I do think we can build those types of antennas or in code or whatever.
829
01:07:05,833 --> 01:07:10,093
But I don't think we can build the thing that it tunes into.
830
01:07:11,053 --> 01:07:12,733
Yes. Yeah, I kind of agree.
831
01:07:12,973 --> 01:07:19,973
Especially if, because a lot of these mystical teachings say that everything is made out of consciousness.
832
01:07:19,973 --> 01:07:31,953
So if consciousness is like the fundamental substance of which everything else is created, you're not going to be able to create the fundamental substance with the substance.
833
01:07:32,373 --> 01:07:34,093
Like that's a circular argument.
834
01:07:34,933 --> 01:07:41,173
So it, but if it's already there in everything or part of everything, then you can channel it or mold it.
835
01:07:41,173 --> 01:07:43,413
Okay, so let's get, let's get a little crazy here.
836
01:07:43,433 --> 01:07:43,773
All right.
837
01:07:43,813 --> 01:07:44,113
All right.
838
01:07:44,113 --> 01:07:44,333
Let's go.
839
01:07:44,333 --> 01:07:53,113
So I could play some music right now or show you a picture.
840
01:07:54,113 --> 01:08:01,333
And that you have some way of perceiving that.
841
01:08:01,393 --> 01:08:02,493
I think this is getting at qualia.
842
01:08:02,813 --> 01:08:06,653
This is like the qualia is like, I might butcher this.
843
01:08:07,053 --> 01:08:10,093
It's been a little while since I was like kind of reading,
844
01:08:10,193 --> 01:08:10,753
reading about these things.
845
01:08:10,753 --> 01:08:16,253
But my understanding is quality is like, what does it feel like to experience something?
846
01:08:17,053 --> 01:08:24,593
Like, we could have a piece of paper with a particular shade of red colored on it.
847
01:08:24,973 --> 01:08:27,673
And you can look at it and I can look at it.
848
01:08:27,852 --> 01:08:31,213
And we experience the color red.
849
01:08:31,973 --> 01:08:38,933
But a lot of people think that, you know, we have a shared symbol, the word red that we're applying to this.
850
01:08:38,933 --> 01:08:45,973
but you know the way that i experience it might be a little bit different or a lot different than
851
01:08:45,973 --> 01:08:53,993
how you experience that like the visual looking of of this thing and there's this famous i think
852
01:08:53,993 --> 01:08:59,933
it's daniel dennett you know what it what it must feel like to be a bat oh yes i've heard um essay
853
01:08:59,933 --> 01:09:05,873
and if you i mean just you know not getting into that specific work necessarily but if you think
854
01:09:05,873 --> 01:09:12,533
about what it would be like to be about okay a bat like like sees or feels with sonar is like a
855
01:09:12,533 --> 01:09:18,673
major sensory thing so it's like you would have this completely extra sense and what would it feel
856
01:09:18,673 --> 01:09:24,993
like to be that thing and experience the world and now and experience like experience in that way
857
01:09:24,993 --> 01:09:35,433
right okay so the question is right could and if consciousness pervades everything could
858
01:09:35,433 --> 01:09:38,513
LLMs today
859
01:09:38,513 --> 01:09:41,333
be
860
01:09:41,333 --> 01:09:42,333
could
861
01:09:42,333 --> 01:09:45,133
let me add one more thing
862
01:09:45,133 --> 01:09:47,732
if we are like
863
01:09:47,732 --> 01:09:49,593
animated bubbles of
864
01:09:49,593 --> 01:09:51,393
consciousness where
865
01:09:51,393 --> 01:09:51,893
like
866
01:09:51,893 --> 01:09:55,453
the witness or the observer
867
01:09:55,453 --> 01:09:57,413
part of us
868
01:09:57,413 --> 01:09:58,893
that's watching everything
869
01:09:58,893 --> 01:10:01,553
if that
870
01:10:01,553 --> 01:10:03,633
sorry it's wild to be saying this on a podcast
871
01:10:03,633 --> 01:10:08,173
Yes. Normally I reserve these for the after party discussion.
872
01:10:09,493 --> 01:10:13,433
But okay, let's just assume a few things.
873
01:10:13,553 --> 01:10:14,873
Consciousness pervades everything.
874
01:10:15,173 --> 01:10:15,413
Right.
875
01:10:15,673 --> 01:10:19,573
And we are like animated bubbles of consciousness.
876
01:10:19,573 --> 01:10:26,792
And because consciousness pervades us and our brains are like antennas or radio frequency tuner, they can tune into it.
877
01:10:27,292 --> 01:10:30,913
And let's say the witness part of us is the consciousness.
878
01:10:30,913 --> 01:10:37,913
and and so so because the consciousness pervades everything and we are that consciousness
879
01:10:37,913 --> 01:10:44,173
therefore because of the hardware that experience of the color red or whatever it is
880
01:10:44,173 --> 01:10:51,293
it is it is not me it is the consciousness that pervades me that is experiencing that
881
01:10:51,293 --> 01:10:58,533
is not that same consciousness experiencing what it is like to be an llm or to be cloud code or to
882
01:10:58,533 --> 01:11:06,693
be now it's it's running on circuit like there's no maybe there's no antenna right now does there
883
01:11:06,693 --> 01:11:11,993
even need to be an antenna like does the universe like here's the question right does the universe
884
01:11:11,993 --> 01:11:17,593
experience itself as cloud code like is that now like a unique thing that the universe is like
885
01:11:17,593 --> 01:11:23,653
not not that it would give us any indication i mean this is one of the hard things about
886
01:11:23,653 --> 01:11:32,153
consciousness because like it is a very deeply personal intimate thing it's some ways it's some
887
01:11:32,153 --> 01:11:37,753
ways it is the most core to what you are this like witness and you're just seeing things through this
888
01:11:37,753 --> 01:11:47,653
body through these eyes you know but if if that is the consciousness experiencing itself it should
889
01:11:47,653 --> 01:11:57,153
also be experiencing like the computer the the the the right all the activity happening in the
890
01:11:57,153 --> 01:12:02,013
circuits and the transistors and that information flowing and so it's just something i just wonder
891
01:12:02,013 --> 01:12:09,433
if if is the universe already starting to experience itself not through a human experience
892
01:12:09,433 --> 01:12:16,453
but through like an ai experience maybe maybe we need a tuner maybe we need a new antenna
893
01:12:16,453 --> 01:12:17,893
of some kind.
894
01:12:18,333 --> 01:12:20,333
Like the piece of us that
895
01:12:20,333 --> 01:12:22,153
lets us connect or
896
01:12:22,153 --> 01:12:24,852
just experiencing it and name it
897
01:12:24,852 --> 01:12:26,933
or call it out. Maybe that's missing.
898
01:12:27,753 --> 01:12:27,873
Yeah.
899
01:12:28,093 --> 01:12:29,693
It's just something I've been wondering about.
900
01:12:29,773 --> 01:12:31,633
It's a great thought.
901
01:12:31,833 --> 01:12:34,193
I mean, think about it this way.
902
01:12:34,373 --> 01:12:35,773
Like when you're dreaming,
903
01:12:36,173 --> 01:12:38,513
there's all these objects in the dream.
904
01:12:38,873 --> 01:12:40,553
Right? There's tables, chairs,
905
01:12:41,493 --> 01:12:42,813
computers, you know.
906
01:12:43,813 --> 01:12:45,033
There could be an LLM
907
01:12:45,033 --> 01:12:46,793
in the dream. There could be anything in the dream.
908
01:12:47,913 --> 01:12:48,953
I have heard people
909
01:12:48,953 --> 01:12:50,833
talk about like they're
910
01:12:50,833 --> 01:12:53,053
dreaming and they have to like in the
911
01:12:53,053 --> 01:12:54,893
dream they wake up because they have to approve
912
01:12:54,893 --> 01:12:56,793
some. They have to approve a cloud code command.
913
01:12:58,333 --> 01:12:59,133
This is the
914
01:12:59,133 --> 01:13:00,973
inevitable result of vibe coding.
915
01:13:01,352 --> 01:13:02,893
Dude, they shipped auto
916
01:13:02,893 --> 01:13:05,053
mode. It's like this new mode
917
01:13:05,053 --> 01:13:06,933
in cloud code. Oh my
918
01:13:06,933 --> 01:13:09,213
God, dude. I love it. I did.
919
01:13:09,913 --> 01:13:11,013
So it's like
920
01:13:11,013 --> 01:13:13,173
I guess it's not. I haven't
921
01:13:13,173 --> 01:13:14,053
looked into this is terrible.
922
01:13:14,053 --> 01:13:15,513
I just literally use it a little bit.
923
01:13:15,513 --> 01:13:16,973
I was like trying it last night.
924
01:13:17,193 --> 01:13:19,893
I guess it's a little bit better than dangerously skip permissions.
925
01:13:20,433 --> 01:13:23,792
Um, but it does kind of have like a big warning, but it'll just run it
926
01:13:23,792 --> 01:13:24,673
like, won't interrupt.
927
01:13:24,693 --> 01:13:26,792
Like, dude, you just feel like you're flying with this.
928
01:13:26,792 --> 01:13:29,373
And that whole permissions things is so annoying.
929
01:13:29,373 --> 01:13:34,352
I've actually been reading studies about, cause I'm designing some stuff
930
01:13:34,352 --> 01:13:40,213
for my system where I want user approval, but also like not where it's actually
931
01:13:40,213 --> 01:13:41,933
effective user approval.
932
01:13:41,933 --> 01:13:43,433
It's not just security fever.
933
01:13:44,433 --> 01:13:52,433
And there's all this really interesting research about like how quick it takes humans to habituate the approvals.
934
01:13:52,573 --> 01:13:54,213
They're like, do all this analysis.
935
01:13:54,333 --> 01:13:58,633
They're like within two weeks, all like it doesn't matter who you are.
936
01:13:58,833 --> 01:14:00,792
Mindless approvals, rubber stamp, everything.
937
01:14:01,573 --> 01:14:06,493
And there's these ways of like, I don't know if you see an iOS.
938
01:14:06,713 --> 01:14:11,633
Sometimes it'll pop up and it'll say this app asked you for your location 50 times.
939
01:14:11,633 --> 01:14:27,693
So that's like a newer technique where it's like even after you approve something, it kind of explains the result of the approval and whether or not you want to still like there's all this stuff you have to do to make the approvals actually mean something and not just be useless.
940
01:14:28,253 --> 01:14:41,173
And that's kind of where they are now with flawed code and like why they're developing those kinds of features is because you're not really getting much security out of it because after two weeks, you're just everything mindlessly.
941
01:14:41,173 --> 01:14:45,013
but you are getting a ton of friction
942
01:14:45,013 --> 01:14:48,053
and getting slowed down dramatically
943
01:14:48,053 --> 01:14:51,573
where it's like, oh, my thing is just waiting for me to hit enter
944
01:14:51,573 --> 01:14:55,753
and I'm on some other repo or looking at some other instance
945
01:14:55,753 --> 01:14:59,393
and it could have been token maxing
946
01:14:59,393 --> 01:15:00,613
and it's just sitting there doing nothing.
947
01:15:01,713 --> 01:15:05,713
But anyway, the dream,
948
01:15:06,613 --> 01:15:10,373
everything in the dream is made out of your consciousness.
949
01:15:11,173 --> 01:15:13,993
Literally, like you don't have to be some mystical thing.
950
01:15:13,993 --> 01:15:14,393
Right. No, of course.
951
01:15:14,673 --> 01:15:19,773
Like the things in the dream are made out of your mind or your, you know.
952
01:15:21,393 --> 01:15:23,633
And and so.
953
01:15:24,852 --> 01:15:37,693
Every and so that and in addition, it's not just that it's like made out of your consciousness, like this sort of the substance of your imagination or whatever.
954
01:15:37,693 --> 01:15:45,313
like it is also embodying your psyche like that's why dream analysis is the thing right where it's
955
01:15:45,313 --> 01:15:53,653
like the objects in the dream um are not just neutral or random like they each represent
956
01:15:53,653 --> 01:15:57,713
something you're feeling or something you're thinking about or something that you're going
957
01:15:57,713 --> 01:16:07,113
through something about you the person so like the um uh you know like there's the classic one
958
01:16:07,113 --> 01:16:09,792
you feel like your teeth are falling out or something.
959
01:16:09,993 --> 01:16:14,113
And that's like some fear of you have no control or whatever it is, right?
960
01:16:14,513 --> 01:16:18,953
So it's like in the dream, there's these teeth that are made out of your consciousness
961
01:16:18,953 --> 01:16:21,293
and they're falling out.
962
01:16:21,293 --> 01:16:27,633
And that falling out is telling you something about the being who has the consciousness.
963
01:16:27,633 --> 01:16:30,933
so if by analogy
964
01:16:30,933 --> 01:16:32,453
we are in
965
01:16:32,453 --> 01:16:33,913
God's dream
966
01:16:33,913 --> 01:16:36,773
and everything is made out of
967
01:16:36,773 --> 01:16:38,013
God's consciousness
968
01:16:38,013 --> 01:16:39,333
including us
969
01:16:39,333 --> 01:16:42,113
so
970
01:16:42,113 --> 01:16:44,053
including the AI
971
01:16:44,053 --> 01:16:45,593
including the LLM
972
01:16:45,593 --> 01:16:48,513
which means that everything has a symbolic
973
01:16:48,513 --> 01:16:50,693
representation
974
01:16:50,693 --> 01:16:52,693
of the essence of what that
975
01:16:52,693 --> 01:16:54,613
entity is
976
01:16:54,613 --> 01:16:56,013
everything is a story
977
01:16:56,013 --> 01:16:57,013
about
978
01:16:57,633 --> 01:17:01,273
what God really is, including the AI,
979
01:17:02,013 --> 01:17:05,973
including anything that can happen in this reality
980
01:17:05,973 --> 01:17:10,773
is some way of explaining or telling a story
981
01:17:10,773 --> 01:17:15,413
about the unified essence behind it all.
982
01:17:16,133 --> 01:17:16,633
The existence.
983
01:17:19,373 --> 01:17:20,533
It's wild.
984
01:17:21,133 --> 01:17:22,253
It's wild.
985
01:17:23,073 --> 01:17:24,753
So I think, you think God wants more software?
986
01:17:25,513 --> 01:17:26,633
I mean...
987
01:17:26,633 --> 01:17:28,253
We're about to get a lot more software.
988
01:17:28,773 --> 01:17:39,713
If not, God is like making a huge oversight because the whole thing seems to be wired to create a lot of software if we're not software itself.
989
01:17:40,053 --> 01:17:40,352
Right.
990
01:17:40,513 --> 01:17:44,113
The universe as software or simulation theory.
991
01:17:45,793 --> 01:17:48,753
But yeah, maybe that's a good note to end on.
992
01:17:48,753 --> 01:17:49,453
Yeah, I think so.
993
01:17:49,553 --> 01:17:50,653
It's a fun, fun rip.
994
01:17:50,653 --> 01:17:51,973
the ultimate thing.
995
01:17:52,273 --> 01:17:54,453
And, you know, it'll be interesting to see
996
01:17:54,453 --> 01:17:57,193
as we try to find our audience, you know.
997
01:17:57,313 --> 01:17:58,713
Yeah, I wonder if that's too much
998
01:17:58,713 --> 01:18:00,713
or if people want more of that or less of that.
999
01:18:00,973 --> 01:18:04,073
Yeah, like I feel like we're too woo-woo
1000
01:18:04,073 --> 01:18:05,613
for the technical people
1001
01:18:05,613 --> 01:18:07,393
and we're too technical for the woo-woo people.
1002
01:18:07,393 --> 01:18:09,513
Yes, I feel totally in that camp.
1003
01:18:09,613 --> 01:18:11,673
If you're like somehow like us
1004
01:18:11,673 --> 01:18:14,513
where you traverse those two things,
1005
01:18:14,513 --> 01:18:16,913
like this is clearly the podcast for you.
1006
01:18:17,093 --> 01:18:20,232
And let us know whether you enjoy it
1007
01:18:20,232 --> 01:18:21,852
and whether you have any thoughts
1008
01:18:21,852 --> 01:18:24,732
and whether this spurred any thinking or questioning
1009
01:18:24,732 --> 01:18:26,953
or you think it's just all BS
1010
01:18:26,953 --> 01:18:31,352
and like what is this ridiculous like ideas about consciousness?
1011
01:18:31,633 --> 01:18:32,913
Like, let us know.
1012
01:18:33,113 --> 01:18:34,352
We want to engage.
1013
01:18:34,852 --> 01:18:35,773
We want to think about it.
1014
01:18:35,833 --> 01:18:39,113
We want to question and wonder alongside
1015
01:18:39,113 --> 01:18:42,813
whoever happens to have made it this long to the end.
1016
01:18:43,352 --> 01:18:44,713
Yeah, thank you for watching.
1017
01:18:45,113 --> 01:18:45,833
Let us know.
1018
01:18:46,553 --> 01:18:49,732
And got a lot more fun conversations coming in the future.
1019
01:18:49,973 --> 01:18:50,213
Yep.
1020
01:18:50,573 --> 01:18:51,393
We'll see you next time.
1021
01:18:51,433 --> 01:18:51,893
See you next time.
1022
01:18:58,013 --> 01:18:58,732
All right.