Tech Acceleration vs Deceleration: e/acc vs. d/acc debate | Erik Torenberg & Haseeb Qureshi

WEBVTT
1
00:00:00.060 --> 00:00:02.660
If you look at the companies that are actually accelerating,
2
00:00:03.643 --> 00:00:04.780
they're mostly EAs, right?
3
00:00:04.820 --> 00:00:06.380
Which like kind of fucks up the narrative,
4
00:00:06.501 --> 00:00:08.780
which is like, oh, the EAs are ruining everything.
5
00:00:08.900 --> 00:00:09.819
They're slowing everyone down.
6
00:00:09.920 --> 00:00:11.660
And like, it's the accelerationists who want to go fast.
7
00:00:12.080 --> 00:00:13.720
The accelerationists are mostly VCs.
8
00:00:14.141 --> 00:00:15.260
They're mostly not building anything.
9
00:00:15.380 --> 00:00:17.260
They're just like on the sidelines cheerleading
10
00:00:17.381 --> 00:00:20.020
for this imaginary team that they think is in the driver's seat
11
00:00:20.060 --> 00:00:22.120
when in reality, it's fucking EA people
12
00:00:22.181 --> 00:00:23.500
who are even thinking that this was possible
13
00:00:23.561 --> 00:00:24.440
in the first place, right?
14
00:00:24.621 --> 00:00:27.720
It was all the VCs who were not investing in any of this stuff
15
00:00:27.963 --> 00:00:29.240
because they thought it was so impractical
16
00:00:29.340 --> 00:00:29.783
and it was worth it.
17
00:00:29.984 --> 00:00:32.760
It's the EAs who believed that AI risk was a real thing because AI was
18
00:00:32.820 --> 00:00:36.240
going to become really powerful really quickly who created all this shit.
19
00:00:40.301 --> 00:00:44.680
Welcome to Bankless, where today we're exploring the frontier of the tech acceleration versus
20
00:00:44.921 --> 00:00:46.607
deceleration debate.
21
00:00:46.929 --> 00:00:47.832
This is Ryan Sean Adams.
22
00:00:47.873 --> 00:00:49.599
I'm here with David Hoffman, and we're here to
23
00:00:49.740 --> 00:00:51.768
help you become more bankless.
24
00:00:52.390 --> 00:00:54.960
The question on today's episode, I think a question that is
25
00:00:55.221 --> 00:00:55.984
everywhere right now.
26
00:00:56.204 --> 00:00:59.396
Should we accelerate or decelerate our tech progress?
27
00:00:59.958 --> 00:01:00.460
And how about when
28
00:01:00.500 --> 00:01:02.729
it comes to something as powerful as AI?
29
00:01:03.211 --> 00:01:05.340
This isn't just a debate in tech circles, at least not
30
00:01:05.380 --> 00:01:05.882
any longer.
31
00:01:06.364 --> 00:01:07.729
This is now a political debate.
32
00:01:08.211 --> 00:01:10.540
I think it's poised to define the next decade.
33
00:01:11.180 --> 00:01:14.355
We've had leaders from both sides of the debate on the podcast.
34
00:01:14.395 --> 00:01:15.580
You might remember our episode
35
00:01:15.660 --> 00:01:21.780
with Eliezer Yudkowsky, who argues very much from an AI safety perspective, and Beth Jesus, which is
36
00:01:22.142 --> 00:01:23.772
almost the complete opposite of that.
37
00:01:23.793 --> 00:01:24.900
So you've already heard what they think.
38
00:01:25.500 --> 00:01:28.371
Today's episode is more of a synthesis type episode.
39
00:01:28.611 --> 00:01:31.000
This is a debate on these two opposing
40
00:01:31.100 --> 00:01:34.214
opinions to help you think through where you might stand on the issue.
41
00:01:34.455 --> 00:01:35.460
We have Hasib Qureshi,
42
00:01:35.660 --> 00:01:39.560
who tends towards the EA side of the conversation, which favors more caution,
43
00:01:40.080 --> 00:01:42.188
more regulatory intervention around AI.
44
00:01:42.690 --> 00:01:43.995
And we have Eric Tornenberg.
45
00:01:44.316 --> 00:01:45.340
He tends more towards the
46
00:01:45.480 --> 00:01:50.840
EAC side of the issue, which is a bit more tech forward, faster progress, and a lighter touch
47
00:01:51.261 --> 00:01:52.105
from art regulators.
48
00:01:52.567 --> 00:01:55.780
We talk about the history of the debate, the cultures in Silicon Valley.
49
00:01:56.121 --> 00:01:57.247
We talk about Sam Altman.
50
00:01:57.267 --> 00:01:59.640
We talk about the regulators, the big bad US government,
51
00:02:00.120 --> 00:02:06.060
SBF, and what they think is the most sensible position on the acceleration versus deceleration
52
00:02:06.381 --> 00:02:06.703
debate.
53
00:02:07.084 --> 00:02:11.020
Before we get into some of the takeaways, our friends and sponsors over at SAFE wants to
54
00:02:11.080 --> 00:02:13.231
let you know that SAFE, this is the multisig.
55
00:02:13.311 --> 00:02:14.176
Bankless uses this.
56
00:02:14.336 --> 00:02:15.040
I've used this personally.
57
00:02:15.100 --> 00:02:16.285
I have my own personal multisig.
58
00:02:16.345 --> 00:02:18.754
A lot of people use the SAFE multisig.
59
00:02:18.774 --> 00:02:20.360
And now SAFE is launching
60
00:02:20.440 --> 00:02:26.420
an activity rewards program so as you use your multi-sig you can get rewarded with points so
61
00:02:26.601 --> 00:02:31.240
weekly users or just for pushing volume through the safe or having a high number of transactions
62
00:02:31.340 --> 00:02:36.440
or you store a lot of assets with safe everything that's an activity around safe gets you now safe
63
00:02:36.480 --> 00:02:40.340
points and they also want you to stay on the lookout for further activities that will offer
64
00:02:40.460 --> 00:02:44.600
additional points as they unlock if you want to learn more about safe points and what you can do
65
00:02:44.640 --> 00:02:48.440
with them click the link in the show notes to getting started with safe points today one of
66
00:02:48.440 --> 00:02:52.240
One of the reasons, Ryan, why I was really motivated to get this episode out here is
67
00:02:52.380 --> 00:02:57.860
because I think right now this conversation is relevant to both crypto and AI is about
68
00:02:57.921 --> 00:02:58.897
like accelerationism.
69
00:02:59.261 --> 00:03:00.800
And are we pushing the gas?
70
00:03:00.961 --> 00:03:02.079
Do we want to push the gas further?
71
00:03:03.342 --> 00:03:05.480
It's now just relevant to these two industries.
72
00:03:06.121 --> 00:03:11.280
I think there's more industries that is going to be relevant for beyond just what the scope
73
00:03:11.866 --> 00:03:13.100
of what this conversation is.
74
00:03:13.261 --> 00:03:15.440
I think this is like a sign of things to come.
75
00:03:15.560 --> 00:03:24.640
There are other industries that are also going to present some very hard conversations to humanity and therefore government.
76
00:03:25.320 --> 00:03:27.560
And our tech sector is going to be coming out of the tech sector.
77
00:03:28.060 --> 00:03:37.780
When we talk about accelerationism, the world is going to look so much more different in 20 years than any other 20-year increment that we've ever had before.
78
00:03:37.920 --> 00:03:39.220
And that's only going to increase.
79
00:03:39.320 --> 00:03:42.220
So we have biotech and gene editing on the horizon.
80
00:03:42.424 --> 00:03:43.057
We have longevity.
81
00:03:43.381 --> 00:03:46.800
People like Brian Johnson literally trying to become immortal.
82
00:03:47.660 --> 00:03:52.140
And I think AI and crypto are like banded together as kind of these accelerationist movements.
83
00:03:52.260 --> 00:03:57.560
But the thing is, I think these are the first of many industries to show up on the accelerationist side.
84
00:03:58.381 --> 00:04:04.420
And this conversation about this like caution versus gas, do we press on the gas pedal or do we hit the brakes?
85
00:04:04.940 --> 00:04:11.620
That is going to be a more, it's going to define society more and more and more, I think, as we move forward into the future.
86
00:04:11.741 --> 00:04:15.540
So that's kind of the frame of thought that I hope listeners go into in this conversation.
87
00:04:15.921 --> 00:04:16.560
I completely agree.
88
00:04:16.640 --> 00:04:20.320
This is the political debate of the decade, but it's also probably the political debate
89
00:04:20.340 --> 00:04:21.760
of the entire next century.
90
00:04:22.963 --> 00:04:26.300
So a lot of podcast episode fodder for us in the future.
91
00:04:26.780 --> 00:04:28.540
All right, guys, we're getting right to the conversation.
92
00:04:28.760 --> 00:04:33.680
But before we do, we want to thank the sponsors that made this episode possible, including
93
00:04:34.303 --> 00:04:38.180
our number one recommended exchange, the place you can accelerate your crypto buys.
94
00:04:38.845 --> 00:04:39.620
That is Kraken.
95
00:04:40.022 --> 00:04:40.700
Go create an account.
96
00:04:40.820 --> 00:04:50.120
If you want a crypto trading experience backed by world-class security and award-winning support teams, then head over to Kraken, one of the longest-standing and most secure crypto platforms in the world.
97
00:04:50.220 --> 00:04:57.680
Kraken is on a journey to build a more accessible, inclusive, and fair financial system, making it simple and secure for everyone, everywhere, to trade crypto.
98
00:04:57.800 --> 00:05:04.900
Kraken's intuitive trading tools are designed to grow with you, empowering you to make your first or your hundredth trade in just a few clicks.
99
00:05:05.000 --> 00:05:09.440
And there's an award-winning client support team available 24-7 to help you along the way,
100
00:05:09.640 --> 00:05:12.760
along with a whole range of educational guides, articles, and videos.
101
00:05:12.980 --> 00:05:16.340
With products and features like Kraken Pro and Kraken NFT Marketplace
102
00:05:16.441 --> 00:05:18.140
and a seamless app to bring it all together,
103
00:05:18.460 --> 00:05:21.860
it's really the perfect place to get your complete crypto experience.
104
00:05:22.040 --> 00:05:25.440
So check out the simple, secure, and powerful way for everyone to trade crypto,
105
00:05:25.580 --> 00:05:28.040
whether you're a complete beginner or a seasoned pro.
106
00:05:28.220 --> 00:05:30.900
Go to kraken.com slash banklist to see what crypto can be.
107
00:05:31.001 --> 00:05:31.500
Not investment advice.
108
00:05:31.600 --> 00:05:32.700
Crypto trading involves risk of loss.
109
00:05:32.760 --> 00:05:33.860
Have you heard about SAFE?
110
00:05:33.920 --> 00:05:39.200
They're not just pioneers in crypto custody, they're also leading the transition to the world of smart accounts.
111
00:05:39.440 --> 00:05:42.440
With SAFe, managing your crypto has never been smarter or safer.
112
00:05:42.560 --> 00:05:49.600
They've recently passed $100 billion in total value secured and are now deployed everywhere with over 15 supported networks.
113
00:05:49.840 --> 00:05:50.422
But that's not all.
114
00:05:50.703 --> 00:05:54.055
SAFe is revolutionizing the game with their full compatibility with ERC-4337,
115
00:05:55.560 --> 00:05:59.320
making embedding wallets super easy to integrate into your daily crypto activities.
116
00:05:59.561 --> 00:06:02.660
And for the crypto enthusiasts out there, SAFe has something special.
117
00:06:03.160 --> 00:06:05.400
the SafePass Activity Rewards Program.
118
00:06:05.560 --> 00:06:08.100
It's a fantastic way to engage and earn rewards,
119
00:06:08.240 --> 00:06:11.960
launching just after Safedow voted in favor of safe token transferability.
120
00:06:12.080 --> 00:06:12.899
So what are you waiting for?
121
00:06:13.260 --> 00:06:17.320
Head over to safe.global, sign up for the SafePass Activity Rewards Program,
122
00:06:17.562 --> 00:06:20.500
and don't forget to check out SafeCon during Berlin Blockchain Week,
123
00:06:20.580 --> 00:06:23.680
where the future of smart accounts is unfolding in real life.
124
00:06:23.800 --> 00:06:25.760
Click the link in the show notes to secure your spot today
125
00:06:25.881 --> 00:06:28.960
and be a part of the transition that's setting a new standard in the world of custody.
126
00:06:29.060 --> 00:06:33.320
Taking self-custody of your crypto is one of the most important things you can do on your bankless journey.
127
00:06:33.440 --> 00:06:37.040
It's also one of the hardest things to get right, with huge consequences if you don't.
128
00:06:37.080 --> 00:06:38.880
If you want help going bankless, talk to Kasa.
129
00:06:39.100 --> 00:06:43.820
Kasa helps you take custody of your crypto assets so you don't have to wonder whether you're doing it right.
130
00:06:43.920 --> 00:06:47.340
Kasa is a one-stop shop for doing self-custody the right way.
131
00:06:47.460 --> 00:06:54.220
With Kasa Vaults, you can hold Ether, Bitcoin, stablecoins, all with one simple app and multiple keys for the ultimate peace of mind,
132
00:06:54.560 --> 00:06:56.900
with a support team to help you every step of the way.
133
00:06:56.940 --> 00:07:01.960
But it doesn't stop at self-custody, because even though crypto is forever, you are not.
134
00:07:02.200 --> 00:07:06.340
We all plan on making life-changing wealth in crypto, but with Kasa's inheritance product,
135
00:07:06.480 --> 00:07:08.900
life-changing wealth can elevate to generational wealth.
136
00:07:08.960 --> 00:07:11.620
For your kids and your loved ones who don't know anything about crypto.
137
00:07:11.780 --> 00:07:15.880
With Kasa, you won't lose your private keys, and you won't accidentally take them to the grave either.
138
00:07:16.120 --> 00:07:19.520
Click the link in the description to get started securing your generational wealth.
139
00:07:19.760 --> 00:07:24.920
Bankless Nation, I'm excited to introduce you to Eric Torenberg, an investor, technologist, and fellow podcaster.
140
00:07:25.061 --> 00:07:29.480
He's got a podcast network called Turpentine Media, lots of excellent podcasts in that
141
00:07:29.540 --> 00:07:34.040
network at the intersection of technology, culture, philosophy, social commentary.
142
00:07:34.300 --> 00:07:35.880
Eric, welcome to the podcast, my man.
143
00:07:36.100 --> 00:07:36.660
Thanks for having me.
144
00:07:36.840 --> 00:07:38.040
Longtime Bankless fan.
145
00:07:38.400 --> 00:07:38.939
Just stoked to be on.
146
00:07:40.622 --> 00:07:43.499
You know, actually, before we introduce your fellow guests here, I actually remember a
147
00:07:43.720 --> 00:07:47.120
very early podcast episode that you had Ryan on as a guest, actually.
148
00:07:47.520 --> 00:07:52.600
Ryan and Cyrus Unessi, I believe, which is when some of the formative years around like
149
00:07:52.742 --> 00:07:53.680
ETH as money as a topic.
150
00:07:54.245 --> 00:07:54.479
Amazing.
151
00:07:55.152 --> 00:07:55.320
Yeah.
152
00:07:55.620 --> 00:07:59.580
Joining Eric here on the podcast today, Bankless listeners probably know Haseeb Qureshi, another
153
00:07:59.741 --> 00:08:04.680
investor, technologist, podcaster, who we've had on the Bankless many, many times before.
154
00:08:05.360 --> 00:08:11.460
GP over at Dragonfly and is also in the effective altruist camp well before anyone in crypto
155
00:08:11.622 --> 00:08:12.900
even knew what that meant.
156
00:08:13.020 --> 00:08:14.100
Haseeb, welcome back to Bankless.
157
00:08:14.761 --> 00:08:15.580
Thanks for having me, guys.
158
00:08:15.863 --> 00:08:16.519
Always fun to be here.
159
00:08:17.742 --> 00:08:22.020
So just to lay some foundation for this episode, I think what this episode is going to be is
160
00:08:22.040 --> 00:08:23.845
kind of is going to evolve on its own.
161
00:08:23.925 --> 00:08:29.220
But really, we are watching two camps out evolve in the tech
162
00:08:29.280 --> 00:08:29.621
space.
163
00:08:29.682 --> 00:08:31.830
And it's also migrated into government as well.
164
00:08:32.593 --> 00:08:33.818
But there's also some nuances.
165
00:08:33.858 --> 00:08:34.320
This is the
166
00:08:34.380 --> 00:08:36.889
accelerationism versus decelerationism camp.
167
00:08:37.451 --> 00:08:39.820
There are some other tribes inside of these tribes.
168
00:08:39.880 --> 00:08:40.965
There are some adjacent tribes.
169
00:08:41.005 --> 00:08:44.460
But overall, there's identities forming in the world of tech
170
00:08:45.042 --> 00:08:48.757
and governance and where we are going as a society in the future.
171
00:08:48.838 --> 00:08:49.400
And it's causing
172
00:08:49.902 --> 00:08:51.888
arguments, is causing debates.
173
00:08:51.948 --> 00:08:55.299
The accelerationists want to just hit on the gas pedal and innovate,
174
00:08:55.600 --> 00:08:57.228
innovate, innovate as fast as possible.
175
00:08:57.650 --> 00:08:59.680
And the decelerationists are like, ooh, scary.
176
00:09:00.501 --> 00:09:01.183
Let's regulate.
177
00:09:01.304 --> 00:09:02.267
Let's regulate these things.
178
00:09:02.668 --> 00:09:06.100
Now, I don't think we've totally seen perfect
179
00:09:06.281 --> 00:09:09.092
identification or identities form around these tribes.
180
00:09:09.132 --> 00:09:11.080
I think this is all new for Silicon Valley
181
00:09:11.221 --> 00:09:14.820
as a whole, which is why we want to bring you guys here on the podcast today to kind of like
182
00:09:14.960 --> 00:09:19.880
us out and like figure out this landscape of these growing alignments across tech across
183
00:09:20.061 --> 00:09:26.319
halocon valley across politics i think we can do that here on the show today let's do it perfect
184
00:09:26.861 --> 00:09:32.719
so i kind of just like laid down a whole bunch of stuff uh different identities different values
185
00:09:33.241 --> 00:09:38.120
um maybe hisib you can kind of get us started here because uh you were you were are an effective
186
00:09:38.281 --> 00:09:42.560
altruist well before anyone in crypto knew or cared what that meant and that was one of the
187
00:09:42.741 --> 00:09:47.880
like earliest call it tribes both in silicon valley in the venture space and and also just
188
00:09:47.960 --> 00:09:52.900
broadly um maybe you can kind of identify about like the formations of that whole camp and how
189
00:09:53.020 --> 00:09:58.739
it's uh innovated or just been a part of the whole like silicon valley tech sector sure so i think
190
00:09:59.101 --> 00:10:04.420
effective altruism for most people it came into the popular imagination with spf which now it's
191
00:10:04.640 --> 00:10:11.300
kind of a irrevocable taint on the movement and on the kind of the ideas behind ea but ea came
192
00:10:11.460 --> 00:10:16.960
originally in around like 2011 2012 mostly around oxford and cambridge from a small number of
193
00:10:17.120 --> 00:10:23.180
philosophers who were very interested in this idea of trying to apply a more rigorous form of ethics
194
00:10:23.340 --> 00:10:27.620
and morality to philanthropy it originally started with philanthropy and it broadened into a number
195
00:10:27.821 --> 00:10:33.280
of other ideas um so it starts with folks like peter singer um and uh you know a small a small
196
00:10:33.480 --> 00:10:40.980
number of um uh what what's what's who's the guy um who was affiliated with sbf um will mccaskill
197
00:10:43.021 --> 00:10:47.300
that's right will mccaskill toby ord uh a few guys who all basically came up around that time
198
00:10:47.561 --> 00:10:52.340
who started writing about a lot of these ideas in public and the origination of the ea movement is
199
00:10:52.440 --> 00:10:57.320
very close to the rationalist movement which was also founded by eliezer yukowski who also has a
200
00:10:57.400 --> 00:11:02.379
lot of ties to ai doomerism um and so the ea though is a little bit different from rationalism
201
00:11:03.402 --> 00:11:07.859
Rationalism is the cohort that really came out of Silicon Valley, whereas actually EAA
202
00:11:08.260 --> 00:11:11.677
mostly came outside of the tech sector, mostly came from universities.
203
00:11:12.180 --> 00:11:17.960
So now EAA, the core ideas behind effective altruism very simply are that most people,
204
00:11:18.000 --> 00:11:21.200
when they think about doing charity or doing good for the world, they kind of turn their
205
00:11:21.321 --> 00:11:21.688
brains off.
206
00:11:22.380 --> 00:11:24.259
They're basically guided by their emotions.
207
00:11:24.340 --> 00:11:28.380
They kind of donate based on mood affiliations or what makes people look good or what sounds
208
00:11:28.462 --> 00:11:28.666
good.
209
00:11:29.541 --> 00:11:34.539
think as rigorously as they do when they are for example doing science or technology or politics or
210
00:11:34.820 --> 00:11:38.020
i don't know maybe not politics but when they're building businesses for example right so when
211
00:11:38.100 --> 00:11:42.340
you're building a business you a b test you look for statistical significance you're like very
212
00:11:42.501 --> 00:11:46.100
careful in making sure does this thing actually improve my product does this thing actually get
213
00:11:46.261 --> 00:11:51.040
the thing that i want and if not you are very very careful to measure the difference and allocate
214
00:11:51.241 --> 00:11:55.179
resources accordingly but when most people do charity they're just kind of like well you know
215
00:11:55.561 --> 00:11:59.540
here for the teaching violin in the inner cities that sounds good sure let's like go you know start
216
00:11:59.661 --> 00:12:04.020
a museum uh that'll that'll be good for kids learning how to read or whatever these kinds
217
00:12:04.161 --> 00:12:08.717
of things that like they sound really good but are they actually good and how good are they and so uh
218
00:12:09.701 --> 00:12:14.500
ea kind of came up as this very scientific engineering heavy approach to thinking about
219
00:12:14.620 --> 00:12:21.780
how to do ethical um ethical actions in the most rigorous way and there's a whole framework around
220
00:12:21.780 --> 00:12:24.880
it i don't think we need to go into the details but one of the things that came out of ea which
221
00:12:24.900 --> 00:12:29.760
which is now very, very relevant in the world of Silicon Valley, is for a long time, EAs have
222
00:12:30.202 --> 00:12:34.880
argued about what is the most important cause area that can move the most impact in the world
223
00:12:35.081 --> 00:12:37.292
over a short amount of time with a small amount of resources.
224
00:12:37.935 --> 00:12:38.960
And, you know, many people talk
225
00:12:38.980 --> 00:12:43.600
about global poverty, animal welfare, but one of the most popular cause areas for a very long time,
226
00:12:43.821 --> 00:12:46.670
almost 10 years now, has been AI risk.
227
00:12:47.412 --> 00:12:49.860
The idea being that as people develop more and more
228
00:12:50.101 --> 00:12:55.620
powerful AIs, AIs might be so powerful that they could end up disrupting society or even causing,
229
00:12:56.021 --> 00:12:57.767
you know, mass human extinction.
230
00:12:58.510 --> 00:13:01.480
And this was before ChatGPT, this was before any of the modern,
231
00:13:01.801 --> 00:13:05.420
you know, kind of acceleration in AI advancement that we've seen over the last 10 years.
232
00:13:06.061 --> 00:13:09.180
And so people have been working on this, you know, a lot of small groups like Miri and
233
00:13:09.742 --> 00:13:14.980
Chai and these, you know, very weird kind of, you know, people in the back room doing stuff that
234
00:13:15.142 --> 00:13:16.660
It was very disconnected from Silicon Valley.
235
00:13:17.480 --> 00:13:33.020
Only now, with the rise of large language models and the sense that everybody has of like, oh, shit, this is actually happening and this is real, has the discourse around AI risk started to really intersect with society, with tech, with economics, with politics.
236
00:13:33.442 --> 00:13:34.760
All these things suddenly matter now.
237
00:13:34.820 --> 00:13:37.520
That's kind of been happening in a back room for about 10 years.
238
00:13:38.301 --> 00:13:44.800
That part of EA, so I was aligned with these ideas basically like 10 years ago before any of this stuff was really on a global stage.
239
00:13:45.101 --> 00:13:46.600
of thinking, hey, you read these arguments.
240
00:13:47.221 --> 00:13:48.620
There's a very famous book by Nick Bostrom
241
00:13:48.681 --> 00:13:49.480
called Superintelligence,
242
00:13:49.560 --> 00:13:51.940
where he sketches out a lot of the in-principle arguments
243
00:13:52.365 --> 00:13:53.600
of how rapid AI advancement
244
00:13:53.802 --> 00:13:55.700
might cause some degree of existential risk.
245
00:13:57.026 --> 00:13:57.700
And you read the arguments,
246
00:13:57.942 --> 00:13:59.380
you're like, yeah, this kind of makes sense.
247
00:14:00.202 --> 00:14:01.180
It's not a slam dunk.
248
00:14:01.220 --> 00:14:02.360
It's like obviously correct.
249
00:14:02.440 --> 00:14:04.080
It's not that clearly this will happen,
250
00:14:04.423 --> 00:14:05.660
but there's at least a risk that it will happen.
251
00:14:05.700 --> 00:14:07.179
And nobody seems to be thinking about that risk
252
00:14:07.542 --> 00:14:10.040
besides these small cohort of weirdos
253
00:14:10.101 --> 00:14:11.437
in the East Bay or in Oxford.
254
00:14:12.382 --> 00:14:13.320
And so people thought at that time,
255
00:14:13.380 --> 00:14:14.840
oh yeah, you should donate some money to these people
256
00:14:14.941 --> 00:14:18.200
because this seems like a very underexplored risk
257
00:14:18.301 --> 00:14:19.440
that right now society is ignoring.
258
00:14:20.161 --> 00:14:21.220
Now today, it's quite different.
259
00:14:21.300 --> 00:14:22.340
Society is not ignoring that risk.
260
00:14:22.440 --> 00:14:22.959
Now it's like a,
261
00:14:23.461 --> 00:14:24.700
it's something that the White House is thinking about,
262
00:14:24.760 --> 00:14:25.580
which is kind of crazy, right?
263
00:14:25.640 --> 00:14:27.060
Very recently, Paul Cristiano,
264
00:14:27.842 --> 00:14:29.800
who was the head of, he was a very, very early EA.
265
00:14:30.222 --> 00:14:31.399
He was the head of safety at OpenAI.
266
00:14:32.025 --> 00:14:32.619
He left OpenAI.
267
00:14:33.160 --> 00:14:35.220
Now he is at NIST.
268
00:14:35.680 --> 00:14:38.540
NIST recently hired him as like the head of AI safety at NIST.
269
00:14:39.062 --> 00:14:40.140
So now at this point, you know,
270
00:14:40.340 --> 00:14:44.060
it's very firmly entrenched in at least the US government
271
00:14:44.382 --> 00:14:47.020
is this concern that, hey, AI safety is a real thing.
272
00:14:47.480 --> 00:14:48.279
And that's not just,
273
00:14:49.987 --> 00:14:53.000
it's not just sort of AI bias or AI replacing jobs.
274
00:14:53.081 --> 00:14:54.520
It's also catastrophic risks
275
00:14:54.903 --> 00:14:57.160
that might arise from the rapid advancement in AI.
276
00:14:58.285 --> 00:14:59.540
Yeah, I think probably listeners
277
00:14:59.661 --> 00:15:01.456
are generally familiar with GiveWell,
278
00:15:02.020 --> 00:15:03.640
which is this charitable organization
279
00:15:03.821 --> 00:15:06.260
that does a lot of research for understanding
280
00:15:06.583 --> 00:15:08.278
if you give this one charity $1,
281
00:15:09.081 --> 00:15:11.360
how effective will that $1 actually go
282
00:15:11.440 --> 00:15:13.060
into like improving people's lives?
283
00:15:13.783 --> 00:15:17.460
and then like this is where famously like they came down to like malaria nets if you actually
284
00:15:17.540 --> 00:15:22.340
want to like put your one dollar and save the most number of human lives you would just buy
285
00:15:22.501 --> 00:15:26.360
malaria nets for like kids in africa to prevent them from getting malaria but this would also
286
00:15:26.440 --> 00:15:30.859
like this is a very effective and rigorous process of understanding how to out capital allocate in
287
00:15:31.200 --> 00:15:35.520
the world of charitable giving but i think as we are working on defining the landscape of all these
288
00:15:35.540 --> 00:15:40.640
different like tribes and identities about like the future arises here i think has seen what you're
289
00:15:40.760 --> 00:15:45.800
also alluding to is that, well, there's like this like risk off low risk research and pragmatic
290
00:15:45.921 --> 00:15:47.812
capital allocation strategy in the world of charity.
291
00:15:48.113 --> 00:15:49.220
But then there's also like, okay,
292
00:15:49.280 --> 00:15:53.736
well, what about some like venture bet future high impact things as well?
293
00:15:53.756 --> 00:15:54.760
So not just about like,
294
00:15:54.860 --> 00:15:59.360
hey, current kids could be impacted by this, but like, we could also have future generations.
295
00:16:00.142 --> 00:16:03.820
And like, not just like next future generation, but literally every single future generation.
296
00:16:04.320 --> 00:16:08.400
So that's, this is how like the world of effective altruism got interested in like,
297
00:16:08.982 --> 00:16:11.191
AI risk because like that's literally the whole thing.
298
00:16:11.231 --> 00:16:13.039
It's like all of human society.
299
00:16:13.079 --> 00:16:13.300
And so
300
00:16:13.902 --> 00:16:18.100
effective altruism is kind of taking the like, not just like a pragmatic approach to charitable
301
00:16:18.200 --> 00:16:23.200
giving today, but like also taking like venture bets in charity or venture bets in like effective
302
00:16:23.321 --> 00:16:23.803
altruism.
303
00:16:24.526 --> 00:16:27.400
How do you, how do you, how would you feel about just that out on right there?
304
00:16:27.540 --> 00:16:28.445
Reflect on that for a second.
305
00:16:28.786 --> 00:16:29.932
No, I think that's exactly right.
306
00:16:30.394 --> 00:16:31.520
Most people, when they think about
307
00:16:31.982 --> 00:16:36.359
EA or they hear these recommendations among EA, they're immediately turned off.
308
00:16:36.379 --> 00:16:36.740
There's a few
309
00:16:36.780 --> 00:16:38.128
reasons why people immediately get turned off.
310
00:16:38.330 --> 00:16:40.040
One is that like, well, you know, who are you to say
311
00:16:40.161 --> 00:16:43.180
what's the most effective and what's not, you know, it seems like you guys are kind of these
312
00:16:43.240 --> 00:16:44.609
big central planning type people.
313
00:16:44.729 --> 00:16:46.500
And, you know, naturally people have an aversion to that.
314
00:16:47.681 --> 00:16:52.180
And, you know, one thing to understand about EA is that EA is making recommendations at the margin,
315
00:16:52.940 --> 00:16:53.081
right?
316
00:16:53.503 --> 00:16:56.599
EA is sort of not assuming that most people will follow EA because they don't.
317
00:16:56.680 --> 00:16:56.740
And
318
00:16:56.740 --> 00:16:59.174
that would be stupid to assume that everyone will do what you tell them to do.
319
00:16:59.536 --> 00:17:00.160
So you always want
320
00:17:00.160 --> 00:17:03.052
to be thinking, what should more people be doing at the margin?
321
00:17:03.674 --> 00:17:04.738
And you'll probably influence 1%,
322
00:17:05.261 --> 00:17:06.519
maybe 2% of people at the most.
323
00:17:07.704 --> 00:17:08.800
If you're influencing everybody,
324
00:17:08.960 --> 00:17:09.820
like let's say you are the government,
325
00:17:09.900 --> 00:17:11.160
then okay, the way you allocate capital
326
00:17:11.160 --> 00:17:12.140
and the way you make recommendations
327
00:17:12.547 --> 00:17:13.280
should be very different.
328
00:17:13.681 --> 00:17:14.800
But if you are saying, okay,
329
00:17:14.941 --> 00:17:17.238
I can basically influence about 1% to 2% of people,
330
00:17:17.901 --> 00:17:19.956
what should 1% to 2% of more people do
331
00:17:21.125 --> 00:17:22.960
given where dollars are already being allocated?
332
00:17:23.943 --> 00:17:25.540
And the answer usually is, well, okay,
333
00:17:25.862 --> 00:17:27.640
there's some obvious things like we should have schools,
334
00:17:28.201 --> 00:17:29.720
we should have sanitation, we should have...
335
00:17:29.781 --> 00:17:30.620
And so somebody might say, well,
336
00:17:30.741 --> 00:17:32.400
are you arguing that we should not have schools
337
00:17:32.503 --> 00:17:33.220
or not have sanitation?
338
00:17:33.402 --> 00:17:34.860
Instead, we should all donate to malaria.
339
00:17:34.940 --> 00:17:35.161
bed nets?
340
00:17:35.181 --> 00:17:36.346
The answer is no, obviously not.
341
00:17:37.451 --> 00:17:39.500
Given where the majority of dollars are being deployed
342
00:17:39.580 --> 00:17:41.448
today, where should the next dollar go?
343
00:17:41.589 --> 00:17:44.220
That's the real question that EA is tasked with trying
344
00:17:44.300 --> 00:17:44.642
to answer.
345
00:17:45.164 --> 00:17:48.700
So given that backdrop, now you might ask the question of like, well, you know, if you
346
00:17:48.740 --> 00:17:51.274
donate to an anti-malaria bed net, okay, I guess that makes sense.
347
00:17:51.294 --> 00:17:52.280
You save somebody's life from
348
00:17:52.601 --> 00:17:53.344
dying of malaria.
349
00:17:53.364 --> 00:17:55.472
That's a very cheap way to save somebody's life.
350
00:17:55.874 --> 00:17:57.480
But, you know, donating to AI
351
00:17:57.580 --> 00:17:59.609
risk, maybe AI is not a risk, right?
352
00:17:59.629 --> 00:18:02.040
Maybe there's only a 10% chance that AI is a risk, in which case
353
00:18:02.281 --> 00:18:03.144
you'd be wasting that money.
354
00:18:04.128 --> 00:18:06.939
And again, you know, the way that EA frames this is that,
355
00:18:07.300 --> 00:18:12.220
you know, you're thinking more about a portfolio of good as opposed to thinking, well, definitely
356
00:18:12.743 --> 00:18:14.572
this good will do this much and this good will do that much.
357
00:18:15.196 --> 00:18:16.020
Most people, when they're thinking
358
00:18:16.100 --> 00:18:18.934
about, okay, I want to do one thing, you know, I want to eat lunch.
359
00:18:18.955 --> 00:18:20.060
And if you're going to eat
360
00:18:20.100 --> 00:18:23.820
lunch, are you going to take a risk on like, you know, maybe this random restaurant is like an
361
00:18:23.840 --> 00:18:25.328
amazing restaurant and I should try it out.
362
00:18:25.609 --> 00:18:27.620
Or you're just like, look, I know that if I go to,
363
00:18:28.140 --> 00:18:30.777
you know, I don't know, Denny's at least have a decent meal, right?
364
00:18:30.978 --> 00:18:31.260
I don't know,
365
00:18:31.441 --> 00:18:32.325
depending on how you feel about dinies.
366
00:18:33.530 --> 00:18:35.600
And most people are naturally very risk averse when it
367
00:18:35.620 --> 00:18:37.872
comes to their own lives or the things that they do with their time.
368
00:18:38.234 --> 00:18:39.320
But the idea is that when you
369
00:18:39.380 --> 00:18:43.660
have a large portfolio of giving, of all the giving in the world, you actually want there to
370
00:18:43.680 --> 00:18:47.120
be some people who are doing high risk giving because, you know, sort of high risk, high reward.
371
00:18:47.381 --> 00:18:51.040
If nobody's doing high risk giving and nobody's doing stuff that looks kind of weird or that
372
00:18:51.321 --> 00:18:55.220
may not pay off or that may not even matter, but in the times when it does matter, it really,
373
00:18:55.340 --> 00:18:56.144
really matters a lot.
374
00:18:56.967 --> 00:18:59.720
So the analogy that a lot of EA people give is that, you know, imagine
375
00:19:00.262 --> 00:19:03.697
that this is something like nuclear research.
376
00:19:05.223 --> 00:19:06.292
If you were in the 1930s or the 1920s
377
00:19:07.282 --> 00:19:08.480
and you were thinking about nuclear research,
378
00:19:09.763 --> 00:19:11.840
at the time it was kind of this relatively boring
379
00:19:12.183 --> 00:19:13.840
kind of scientific thing that nobody really,
380
00:19:14.021 --> 00:19:15.240
nobody really understood why this mattered
381
00:19:15.342 --> 00:19:16.300
until they saw the atom bomb
382
00:19:16.320 --> 00:19:18.580
and they saw the destructive power of nuclear energy.
383
00:19:18.862 --> 00:19:21.160
And there was a time that this was considered
384
00:19:21.241 --> 00:19:22.820
to be very speculative,
385
00:19:23.200 --> 00:19:25.020
that this was even possible at this kind of scale
386
00:19:25.080 --> 00:19:27.480
to have this much destructive damage from this technology.
387
00:19:28.124 --> 00:19:30.380
and on the off chance that it turned out to be right,
388
00:19:31.141 --> 00:19:32.860
that the order of magnitude estimates of this stuff
389
00:19:32.963 --> 00:19:33.520
was correct,
390
00:19:33.842 --> 00:19:35.500
then you would really want there to be
391
00:19:35.621 --> 00:19:36.880
a lot of research into the safety
392
00:19:37.726 --> 00:19:39.880
of how to prevent nuclear proliferation,
393
00:19:40.000 --> 00:19:42.900
how to prevent mass nuclear scale wars.
394
00:19:44.282 --> 00:19:46.440
But people may be under-invested in that.
395
00:19:46.520 --> 00:19:48.715
And as a result, people in the 50s, 60s, 70s
396
00:19:49.822 --> 00:19:51.920
kind of lived under this specter
397
00:19:51.940 --> 00:19:53.900
that maybe there would be
398
00:19:53.981 --> 00:19:55.700
maybe one or two more decades of human existence
399
00:19:55.822 --> 00:19:57.000
and then the world would just wipe itself out.
400
00:19:57.741 --> 00:20:01.000
And for all we know, maybe we weren't that far from that happening.
401
00:20:01.941 --> 00:20:11.200
Maybe we were kind of various times on the brink of massive amounts of human damage and catastrophic risk from this proliferation of nukes.
402
00:20:11.740 --> 00:20:14.239
And so the question is, is AI like this?
403
00:20:15.202 --> 00:20:16.340
Is AI a similar technology?
404
00:20:16.480 --> 00:20:18.900
And the answer might be, well, probably not.
405
00:20:19.280 --> 00:20:23.360
Maybe AI only has like a 20 percent chance of that, 10 percent chance of that.
406
00:20:23.781 --> 00:20:33.680
But even if it's 20% or 10% chance, what an EA would likely say is that, well, that is high enough that you probably want some people, maybe 1% to 2% of people spending their time on this.
407
00:20:33.740 --> 00:20:40.500
Because 1% to 2% of people spending their time on something that might be 10% as bad as nukes, that's a pretty good trade.
408
00:20:40.580 --> 00:20:44.980
That's a pretty good investment if you're thinking about the overall portfolio of how people are spending their time and spending their resources.
409
00:20:45.520 --> 00:20:57.960
So if you think about EA as here is what I decree as an EA, as this East Bay rationalist guy that everybody in the world should be doing, then I think you'll come away with the sense that EA or EAs are a bunch of selfish pricks and they think they should run the world.
410
00:20:58.740 --> 00:21:14.820
But if you come away with the sense that, OK, EAs are advocating about at the margin how to alter the portfolio of all the different things that human beings are working on to ameliorate risk and increase the likelihood of human flourishing, then you might say, OK, that sounds like a reasonable investment strategy.
411
00:21:15.422 --> 00:21:18.640
Maybe I don't agree with it completely, but it's a lot more reasonable if you think about it that way.
412
00:21:19.981 --> 00:21:25.160
So Haseeb, I want to bring Eric into this conversation, just kind of reflect on what you said and add.
413
00:21:25.482 --> 00:21:27.200
But I have just one question that kind of plagues me.
414
00:21:27.340 --> 00:21:31.240
So we started this conversation talking about accelerationism versus tolerationism.
415
00:21:31.360 --> 00:21:39.700
The way we started this conversation was talking about EA and the rationale behind how EA came to support and fund a lot of the AI safety initiatives.
416
00:21:39.860 --> 00:21:46.180
And it's really, when I think of AI, it's really like utilitarianism or consequentialism as a philosophy kind of applied to giving.
417
00:21:46.823 --> 00:21:48.380
That part to me sort of makes sense.
418
00:21:48.841 --> 00:21:54.260
But it doesn't necessarily follow that AI must lead to investments and funding of AI safety.
419
00:21:54.740 --> 00:22:06.260
Because if you are in the EA community and you believe in this sort of giving, you could just as well make the case that in order to benefit the most amount of, like the greatest number of future humans, say trillions of humans,
420
00:22:06.741 --> 00:22:19.540
We want to actually invest in AI technology because we believe AI technology will build a utopia and cure diseases, cure illnesses, solve the problem of death and old age and all of these things.
421
00:22:19.960 --> 00:22:26.979
And so you could easily make the case in the EA community under that philosophy that, no, we shouldn't be funding necessarily anything that would slow down AI.
422
00:22:27.662 --> 00:22:29.060
We should actually be funding startups.
423
00:22:29.260 --> 00:22:30.500
We should be funding open AI.
424
00:22:30.660 --> 00:22:33.798
We should be funding all the tech that goes into AI.
425
00:22:34.601 --> 00:22:42.460
But it seems like the community just forked mainly in one direction such that EA has become somewhat associated with AI safety.
426
00:22:42.860 --> 00:22:52.040
I think, by the way, recently, maybe some of the founders of EA, Will McCaskill and others, have kind of walked back from that and be like, hey, you know, that's not – EA is not necessarily AI safety.
427
00:22:52.180 --> 00:22:52.940
They're not one and the same.
428
00:22:53.040 --> 00:22:53.840
There's kind of some differences.
429
00:22:54.280 --> 00:22:56.320
Anyway, I just throw that idea out there.
430
00:22:56.520 --> 00:22:59.640
And, Eric, what are your reflections on this conversation so far?
431
00:23:00.041 --> 00:23:06.300
Well, first, I just want to add that not all long-termism is in the venture bet category,
432
00:23:06.720 --> 00:23:06.900
right?
433
00:23:07.061 --> 00:23:10.440
It's also things like pandemic prevention or reducing nuclear proliferation, things
434
00:23:10.460 --> 00:23:11.840
that already exist, already happened.
435
00:23:11.920 --> 00:23:14.400
And they were talking about pandemic prevention before COVID, right?
436
00:23:14.922 --> 00:23:16.519
So they're really onto some things.
437
00:23:17.261 --> 00:23:24.720
The leap that long-termism can be broadly explained by this idea that EA is about a
438
00:23:24.840 --> 00:23:26.320
kind of arbitrage, right?
439
00:23:26.822 --> 00:23:31.400
uh dave you described how if one dollar can make a better difference uh bigger difference to people
440
00:23:31.440 --> 00:23:36.180
in africa than it can to people in san francisco well maybe we should put our dollar to to people
441
00:23:36.300 --> 00:23:42.860
in africa and and ea says okay um because a person in africa's life is worth as much as a person in
442
00:23:42.920 --> 00:23:47.280
san francisco's life what about a person that doesn't exist yet what about a future person
443
00:23:47.762 --> 00:23:52.920
that person's life is also worth uh as much as a current person's life and guess what there's
444
00:23:52.940 --> 00:23:57.014
going to be trillions of future people, way more than exist today.
445
00:23:57.476 --> 00:23:58.760
And thus, we should really care
446
00:23:58.941 --> 00:23:59.443
about the future.
447
00:23:59.824 --> 00:24:02.535
And there's a lot of people who got on board with that.
448
00:24:03.037 --> 00:24:03.660
There's a lot of people
449
00:24:04.062 --> 00:24:05.267
who didn't get on board with that.
450
00:24:05.428 --> 00:24:07.437
And this was a schism in the EA community.
451
00:24:07.457 --> 00:24:08.000
Some people are like,
452
00:24:08.040 --> 00:24:12.419
hey, I'm all about effective charity, but this long-termism stuff, that's a bit too out there
453
00:24:13.643 --> 00:24:13.924
for me.
454
00:24:14.165 --> 00:24:15.650
And there's some people who go all the way.
455
00:24:16.673 --> 00:24:18.820
And now, even among the people who go all
456
00:24:18.840 --> 00:24:21.311
the way, there's another schism that Ryan just outlined.
457
00:24:21.331 --> 00:24:23.540
Some people say, hey, it really depends
458
00:24:23.580 --> 00:24:25.168
on what you think about the future of AI.
459
00:24:25.730 --> 00:24:27.780
Some people say that the idea that you just said,
460
00:24:28.080 --> 00:24:31.600
Ryan, which is if you're concerned about helping the most future people, you want to accelerate
461
00:24:32.483 --> 00:24:32.583
AI.
462
00:24:33.125 --> 00:24:36.696
And these are people who are effective accelerationists.
463
00:24:36.716 --> 00:24:37.880
This is why there is an
464
00:24:37.940 --> 00:24:39.024
additional schism.
465
00:24:39.064 --> 00:24:42.777
It's people who are not on board with the AI safety agenda.
466
00:24:43.218 --> 00:24:43.700
And one thing
467
00:24:43.740 --> 00:24:48.495
that's worth noting, you know, EA and EAC are having the strongest conflict.
468
00:24:48.857 --> 00:24:49.920
But what did Henry
469
00:24:49.980 --> 00:24:50.703
Kissinger say?
470
00:24:50.763 --> 00:24:56.140
Something like the tyranny of small differences, like AI safety people and EAC people
471
00:24:56.401 --> 00:25:00.171
agree on 99% of things in many cases.
472
00:25:00.773 --> 00:25:03.180
They both believe that the people who are worried about
473
00:25:03.541 --> 00:25:11.098
misinformation or DEI or AI saying racist things, they're not focused.
474
00:25:11.218 --> 00:25:12.000
They're not at the adult
475
00:25:12.060 --> 00:25:14.549
table of the important conversation.
476
00:25:14.830 --> 00:25:15.834
Those people are misguided.
477
00:25:16.095 --> 00:25:17.600
They're not disagreeing with that.
478
00:25:18.080 --> 00:25:21.833
And in many ways, AI safety people and EA people are tech accelerationists.
479
00:25:22.275 --> 00:25:23.680
They agree that nuclear
480
00:25:23.800 --> 00:25:25.586
regulation has been really bad.
481
00:25:25.927 --> 00:25:28.977
They want to accelerate things like self-driving cars or
482
00:25:30.823 --> 00:25:34.694
longevity or other things that EA people get on board with.
483
00:25:35.236 --> 00:25:36.740
The one place in which they disagree
484
00:25:37.346 --> 00:25:38.539
is this field of AI.
485
00:25:38.941 --> 00:25:41.340
They say AI, this time it's different.
486
00:25:41.762 --> 00:25:43.500
We're not just introducing a new technology.
487
00:25:44.081 --> 00:25:46.540
We're almost introducing like a new form of evolution
488
00:25:46.862 --> 00:25:49.479
or a new species that one day is going to be
489
00:25:49.822 --> 00:25:52.080
cognitively way smarter than us.
490
00:25:52.401 --> 00:25:55.660
And thus is a threat, not just to fake news.
491
00:25:56.381 --> 00:25:58.500
It's not like social media, not just to jobs,
492
00:26:00.408 --> 00:26:02.300
but to actual humanity,
493
00:26:02.741 --> 00:26:05.700
to actually the chance of us being extinct.
494
00:26:06.044 --> 00:26:07.400
and thus we should treat it differently.
495
00:26:07.841 --> 00:26:10.220
And that is the core of the argument
496
00:26:10.402 --> 00:26:12.499
between effective accelerationists
497
00:26:12.882 --> 00:26:15.180
and effective altruists as it relates to AI safety.
498
00:26:15.460 --> 00:26:15.919
Okay, okay.
499
00:26:16.081 --> 00:26:17.040
So we're dealing with like,
500
00:26:17.080 --> 00:26:18.920
this is like an argument between like Catholics
501
00:26:19.063 --> 00:26:20.020
and Protestants basically.
502
00:26:20.200 --> 00:26:22.080
So like all in the same religion,
503
00:26:22.321 --> 00:26:25.060
but like you guys are arguing over these specifics here.
504
00:26:25.320 --> 00:26:26.300
I'm going to dispute that.
505
00:26:26.402 --> 00:26:27.400
I'm going to dispute that a little bit.
506
00:26:27.620 --> 00:26:30.699
I think, I think I'd say a lot of EACs.
507
00:26:31.444 --> 00:26:32.600
So first of all, I think it's probably worth,
508
00:26:33.566 --> 00:26:35.260
worth actually defining what EAC is,
509
00:26:35.421 --> 00:26:39.920
what it means as a movement because i think it's also been in a sense co-opted by what people want
510
00:26:40.000 --> 00:26:44.960
it to mean which is like accelerationism right this idea that oh all technology is good and all
511
00:26:44.980 --> 00:26:48.640
technology should be faster we should be super libertarian we should you know kind of let all
512
00:26:48.700 --> 00:26:53.320
the guardrails off and just let it vroom vroom vroom right that's not actually what yak says
513
00:26:53.601 --> 00:26:57.880
right if you read the yak manifesto from from beth jayzos who's the guy who invented it is like this
514
00:26:57.900 --> 00:27:01.880
google engineer worst quantified it's actually pretty fucking crazy like what he actually
515
00:27:02.021 --> 00:27:02.563
literally says.
516
00:27:03.647 --> 00:27:07.280
Like, I don't know, Eric, I, I, I'm not, it's been a while since I've actually
517
00:27:07.581 --> 00:27:12.280
looked at that manifesto, but there's stuff in there about how like, yes, if, if AI ends up,
518
00:27:12.320 --> 00:27:14.370
you know, totally conquering humanity, like that's good.
519
00:27:14.913 --> 00:27:16.240
We actually want to feed entropy.
520
00:27:16.421 --> 00:27:17.527
We actually worship entropy.
521
00:27:17.848 --> 00:27:20.120
Entropy is the thing, you know, the, the, the core value from
522
00:27:20.140 --> 00:27:21.064
which all good flows.
523
00:27:21.686 --> 00:27:25.160
It's like actually very weirdly cult-like and kind of anti-humanist.
524
00:27:25.621 --> 00:27:29.440
I think most people don't, haven't actually read it or they, they just kind of take it as a vibe
525
00:27:29.862 --> 00:27:33.640
and they're like well the vibe that i want it to be is like you know basically mark andreessen i
526
00:27:33.861 --> 00:27:38.880
think that that is what i imagine eac is but if you actually read the eac manifesto it's like some
527
00:27:38.960 --> 00:27:43.260
it's a bunch of other weird shit that's adjacent to like the mark andreessen philosophy but it's
528
00:27:43.320 --> 00:27:48.639
not quite the same thing so i think it's sort of been typecast into this is like silicon valley uh
529
00:27:49.321 --> 00:27:53.080
you know pro-tech libertarian energy but that's not that's not literally what eac says
530
00:27:54.545 --> 00:27:59.340
i don't know eric if you if you would agree or disagree with that i i i do agree with it and i
531
00:27:59.480 --> 00:28:05.340
should say that that is my um my tribe in some sense is is people like mike solana people like
532
00:28:05.400 --> 00:28:12.320
balaji people like mark andreessen um and so that is where that is where i come from and so i've
533
00:28:12.320 --> 00:28:20.680
been very sympathetic to this eac movement right which um to to some degree says hey um over the
534
00:28:20.700 --> 00:28:27.980
last decade, there have been a lot of anti-tech movements and sentiments and people trying to
535
00:28:28.722 --> 00:28:33.740
use arguments to thus regulate tech in ways that have been very bad from whether it's energy or
536
00:28:33.920 --> 00:28:36.671
social media or a wide variety of things.
537
00:28:36.711 --> 00:28:37.253
How about crypto?
538
00:28:37.314 --> 00:28:38.860
Is that bad to the West?
539
00:28:39.761 --> 00:28:40.925
Yes, exactly.
540
00:28:41.427 --> 00:28:43.816
And we're not going to let that happen here.
541
00:28:43.976 --> 00:28:45.040
And we're finally going to stand
542
00:28:45.100 --> 00:28:50.120
up for ourselves because you saw Zuckerberg and others apologize, and that didn't really get them
543
00:28:50.200 --> 00:28:50.481
anywhere.
544
00:28:51.203 --> 00:28:57.940
It only conceded their sort of lack of moral sort of credibility and that they were
545
00:28:58.000 --> 00:28:59.285
wrong and blah, blah, blah.
546
00:28:59.305 --> 00:29:00.490
And that's all these bad things.
547
00:29:00.931 --> 00:29:03.160
And so this EAC is coinciding with
548
00:29:03.220 --> 00:29:08.057
a desire for tech to stand up for itself, to say, hey, we are actually good.
549
00:29:08.178 --> 00:29:08.820
We are actually going
550
00:29:08.900 --> 00:29:12.570
to push back and we are going to defend ourselves.
551
00:29:13.092 --> 00:29:16.020
Now, I think it brings up a really interesting
552
00:29:16.100 --> 00:29:22.500
point because yeah, Mike Solana, Balaji and Marc Andreessen are not saying that, you know,
553
00:29:22.560 --> 00:29:26.255
we're going to be pro-human or sorry, post-human in some sense.
554
00:29:26.275 --> 00:29:27.500
They're saying that technology is
555
00:29:27.540 --> 00:29:31.150
going to make human life way better for decades to come.
556
00:29:31.692 --> 00:29:34.640
And I think what Beth was doing,
557
00:29:35.021 --> 00:29:38.433
and you know, in some ways, EAC is critiqued as being about vibes.
558
00:29:38.975 --> 00:29:40.340
And the benefit of being about
559
00:29:40.400 --> 00:29:42.309
vibes is you can't really be pinned down.
560
00:29:42.591 --> 00:29:44.560
You can't really be critiqued on your philosophy.
561
00:29:45.201 --> 00:29:56.160
And I think that's partially because of what Hasib just identified, which is if you take it to its logical conclusion, it's not something that a lot of people want to get on board with.
562
00:29:56.460 --> 00:30:03.980
If you take sort of advancing technology to its logical conclusion, like what is the role for humans once we have super intelligence?
563
00:30:05.121 --> 00:30:09.540
What is like when we're like cyborgs and stuff and then different speciation as a result?
564
00:30:09.680 --> 00:30:14.840
Like, so it is very weird, but also I think in some ways it's intellectually honest of
565
00:30:14.940 --> 00:30:16.700
like, that's where things are going.
566
00:30:16.800 --> 00:30:22.060
And I couldn't pin down Solana or Bology or Mark on like, what do they think about that?
567
00:30:22.140 --> 00:30:24.459
But, but I think there is some truth to that.
568
00:30:25.321 --> 00:30:32.460
But I also think there's some, like, I do think it's hard to be a pro tech on a long
569
00:30:32.500 --> 00:30:37.880
enough timescale and not think that we're going to like be like, think that humans are
570
00:30:37.920 --> 00:30:42.259
the last sort of uh you know form of intelligence that is going to going to rule the world so i i
571
00:30:42.561 --> 00:30:46.980
think it's intellectually honest i also think it is very weird so has he well what's this what's
572
00:30:47.161 --> 00:30:51.820
what's the problem with like yak plus some humanism right it's like a more you know like
573
00:30:51.981 --> 00:30:55.560
moderated to be clear there's yeah there's there's nothing there's nothing wrong with the yak plus
574
00:30:55.640 --> 00:30:59.740
some humanism i think it's much preferable to like raw yak as it it was actually articulated
575
00:30:59.901 --> 00:31:04.279
in its genesis i think it's evolved that way the thing that's weird yeah well i don't know
576
00:31:04.620 --> 00:31:06.280
I don't know that it's evolved so much it's been co-opted.
577
00:31:06.642 --> 00:31:08.060
I think it's more like it's been co-opted, right?
578
00:31:08.080 --> 00:31:08.720
It's been co-opted.
579
00:31:09.720 --> 00:31:20.540
The reality is that any time you get a culture war, you basically get the superimposition of what was once like a pretty esoteric argument about like, well, will AIs take over the world?
580
00:31:20.600 --> 00:31:21.000
Is it risky?
581
00:31:21.100 --> 00:31:21.559
Is it not risky?
582
00:31:21.700 --> 00:31:22.360
Is it good to have more?
583
00:31:22.460 --> 00:31:23.480
Is it good to slow down?
584
00:31:24.281 --> 00:31:31.620
This is like a pretty conceptual theoretical argument that's been going on in very obscure circles for a very long time.
585
00:31:32.263 --> 00:31:35.960
And then basically in the span of about a year and a half, it has now become a culture war.
586
00:31:36.020 --> 00:31:39.140
And once it becomes a culture war, then all of a sudden people get typecast, right?
587
00:31:39.200 --> 00:31:41.720
So like, okay, if you're an EA, you're like the decelerationist.
588
00:31:41.800 --> 00:31:43.060
That means that you hate technology.
589
00:31:43.381 --> 00:31:48.160
And if you're an EAC, you love technology and you love Silicon Valley and you love Marc Andreessen and you love Peter Thiel.
590
00:31:48.500 --> 00:31:50.320
And it's like, well, you know, not necessarily.
591
00:31:50.521 --> 00:31:52.072
Like in reality, there's a panoply of views.
592
00:31:53.901 --> 00:31:56.660
There's people who arrive at different parts of the spectrum from many different places.
593
00:31:57.020 --> 00:32:00.738
And there's people who say, well, look, I think the probability of AI is going badly is 5%.
594
00:32:01.521 --> 00:32:04.397
but the likelihood that actually it's better if it's open source is like 25%.
595
00:32:05.604 --> 00:32:07.398
Or some people say like, well, look, I think it's like 10%,
596
00:32:07.821 --> 00:32:10.440
but I actually think that decelerationism is a bad idea
597
00:32:10.541 --> 00:32:12.140
and it ends up just politically backfiring,
598
00:32:12.240 --> 00:32:13.760
and so we kind of have to live with the acceleration.
599
00:32:14.922 --> 00:32:17.320
In reality, there are EAs and there are EACs, quote-unquote,
600
00:32:17.460 --> 00:32:19.580
who have views anywhere along the spectrum,
601
00:32:20.044 --> 00:32:21.960
but once you become a culture war, all that shit goes away.
602
00:32:22.380 --> 00:32:24.300
And it's just like, oh, you're EAC, oh, you're Decel,
603
00:32:24.420 --> 00:32:25.580
or you're EA, or whatever the fuck.
604
00:32:26.101 --> 00:32:29.059
And people ultimately want a one-dimensional
605
00:32:29.220 --> 00:32:33.940
or sort of one-bit answer to how you feel about what is probably one of the most important
606
00:32:34.020 --> 00:32:38.560
questions right now facing technologists or, you know, technology regulation in America.
607
00:32:39.101 --> 00:32:43.760
So, you know, if you ask EAs, what do you think we should do, you know, regulatorily
608
00:32:43.901 --> 00:32:46.860
about AIs, you'll get a lot of different answers.
609
00:32:47.040 --> 00:32:49.900
You won't get, well, we should shut it all down and, you know, fucking control all the
610
00:32:49.900 --> 00:32:50.660
GPUs, right?
611
00:32:50.760 --> 00:32:51.560
Eliezer says that.
612
00:32:51.861 --> 00:32:58.360
But Eliezer is like the, all the way on the spectrum, he's like, you know, the platonic
613
00:32:58.480 --> 00:33:01.210
ideal of a decelerationist, right, in that sense.
614
00:33:01.511 --> 00:33:04.080
But most people on EA are very, you know, there's
615
00:33:04.100 --> 00:33:07.238
like a multidimensional spectrum that people sit on different places in that spectrum.
616
00:33:07.439 --> 00:33:07.660
So for,
617
00:33:07.700 --> 00:33:11.854
you know, for myself, I was donating to AI risk, like in 2014, 2015, way before, you know,
618
00:33:13.701 --> 00:33:14.445
large language model.
619
00:33:14.485 --> 00:33:17.260
I mean, not way before large language model, but before any of this stuff was
620
00:33:17.300 --> 00:33:18.244
in the popular imagination.
621
00:33:18.866 --> 00:33:22.280
And it was not very obvious at that time, what were the specific
622
00:33:22.401 --> 00:33:25.900
policy prescriptions that were going to be made or how was this going to interact with Silicon Valley?
623
00:33:26.181 --> 00:33:30.020
At that time, the people who were working on AI risk were also Silicon Valley nerds.
624
00:33:30.160 --> 00:33:31.360
It was the same team, right?
625
00:33:31.440 --> 00:33:33.440
It was the same people talking about this stuff.
626
00:33:34.341 --> 00:33:41.740
It's only very recently that now there's this sense that, well, you know, build, baby, build, which is like the Silicon Valley mantra, is now hitting this wall.
627
00:33:41.900 --> 00:33:50.740
And, like, I think it's a little bit of a boogeyman as well because the reality is that the EA types, you know, they don't really have that much power.
628
00:33:51.201 --> 00:33:54.200
Like, they're not actually stopping anyone from doing anything at the moment.
629
00:33:54.521 --> 00:33:59.440
Like the real people who are stopping AIs are like the EU or like China, right, who are not EAs, right?
630
00:33:59.480 --> 00:34:00.520
That's not what's motivating them.
631
00:34:00.560 --> 00:34:04.780
It's like data privacy laws or just deference to the CCP, right?
632
00:34:04.900 --> 00:34:12.380
Even the executive order that Biden passed said that, OK, if you're above a certain training run size, you have to report that to the government.
633
00:34:13.668 --> 00:34:13.959
That's it.
634
00:34:14.060 --> 00:34:15.000
It doesn't say you have to stop.
635
00:34:15.081 --> 00:34:15.759
It doesn't say you can't do it.
636
00:34:15.820 --> 00:34:18.800
It doesn't say you have to follow these rules or whatever.
637
00:34:18.800 --> 00:34:25.640
It just says you have to let us know that it's happening, which to me, like, OK, that's a pretty far cry from, you know, deceleration, blah, blah, blah.
638
00:34:26.760 --> 00:34:37.660
It's really just like an information gathering exercise, which I don't know, like probably they wouldn't have that much trouble gathering that information anyway, given the amount of GPUs would have to be cornered by anybody running a training run of that size.
639
00:34:38.121 --> 00:34:48.920
So I think there is a lot of reactivity right now, and that's what's causing people to believe, oh, you're way on this side and I'm way on this side and you kind of can't take any nuance of you in the middle.
640
00:34:49.844 --> 00:34:51.020
That's what happens in a culture war.
641
00:34:52.403 --> 00:34:59.000
first let me just say that eliza has also tweeted things like um abortion we should be able to abort
642
00:34:59.120 --> 00:35:04.620
babies that are up to 18 months or you should be able to leave your partner if you find some
643
00:35:04.821 --> 00:35:10.660
your wife or husband if you find someone 25 better uh like he he is you mentioned the platonic ideal
644
00:35:10.881 --> 00:35:15.780
he takes utilitarianism to the extreme and so like in the same way that best manifesto takes
645
00:35:15.820 --> 00:35:20.160
that idea to the extreme there are people in ea sbf of course one of them who take uh you know
646
00:35:20.421 --> 00:35:26.180
some of the ideas within EA to the extremes so far that they're no longer, you know, sort of
647
00:35:26.200 --> 00:35:26.661
palatable.
648
00:35:26.681 --> 00:35:29.810
I would take the other side a little bit.
649
00:35:30.372 --> 00:35:33.180
I do think EA has a lot of power because I
650
00:35:33.501 --> 00:35:38.380
think that a lot of people who make up OpenAI, who make up Anthropic, who make up some of these
651
00:35:38.440 --> 00:35:42.170
biggest labs were influenced by EA or themselves are.
652
00:35:42.591 --> 00:35:43.172
They were.
653
00:35:43.734 --> 00:35:44.576
Yes, exactly.
654
00:35:44.696 --> 00:35:46.200
And so the ideas,
655
00:35:46.581 --> 00:35:48.387
they sort of came up in the EA soup.
656
00:35:48.909 --> 00:35:52.440
And right now there's not abrasive regulations, as you
657
00:35:52.440 --> 00:35:56.580
mentioned, you know, people are still allowed, but we're at a time where these regulations are
658
00:35:56.600 --> 00:36:01.555
about to be defined, the regulatory era of AI over the next year, et cetera.
659
00:36:01.776 --> 00:36:03.160
And the people who have,
660
00:36:04.221 --> 00:36:09.940
you know, like they're inviting people into the conversation who have EA influence, EA ideas.
661
00:36:10.080 --> 00:36:15.840
And so I do think the area safety community will have an impact into the regulatory.
662
00:36:16.780 --> 00:36:19.560
But OK, but you raise a good point, which is kind of ironic, right?
663
00:36:19.600 --> 00:36:27.420
Which is that if you look at OpenAI, if you look at Anthropic, if you look at the companies that are actually accelerating, they're mostly EAs, right?
664
00:36:27.500 --> 00:36:31.440
Which like kind of fucks up the narrative, which is like, oh, the EAs are ruining everything.
665
00:36:31.560 --> 00:36:32.479
They're slowing everyone down.
666
00:36:32.600 --> 00:36:34.340
And like it's the accelerationists who want to go fast.
667
00:36:34.740 --> 00:36:36.380
The accelerationists are mostly VCs.
668
00:36:36.801 --> 00:36:37.920
They're mostly not building anything.
669
00:36:38.040 --> 00:36:46.780
They're just like on the sidelines cheerleading for this imaginary team that they think is in the driver's seat when in reality it's fucking EA people who are even thinking that this was possible in the first place.
670
00:36:47.320 --> 00:36:47.440
Right?
671
00:36:47.540 --> 00:36:52.800
It was all the VCs who were not investing in any of this stuff because they thought it was so impractical and it was worth it.
672
00:36:52.980 --> 00:36:59.240
It's the EAs who believed that AI risk was a real thing because AI was going to become really powerful really quickly who created all this shit.
673
00:36:59.880 --> 00:37:05.240
So that's why I think like, again, this culture war typecasting that has happened is just ahistorical.
674
00:37:05.361 --> 00:37:08.420
It just doesn't make sense given who's actually the players involved.
675
00:37:08.840 --> 00:37:19.020
Most of the people who are regulating AI, like, yes, we just talked about Paul Cristiano, who is now somewhat influential, presumably at NIST, which is, again, a relatively minor kind of scientific organization.
676
00:37:19.843 --> 00:37:21.340
He's not a congressperson, right?
677
00:37:21.581 --> 00:37:23.500
That's who actually writes the laws is congresspeople.
678
00:37:23.780 --> 00:37:25.480
If you listen to congresspeople, what do they care about?
679
00:37:25.520 --> 00:37:26.259
They care about jobs.
680
00:37:26.563 --> 00:37:27.240
They care about bias.
681
00:37:28.000 --> 00:37:31.480
They care about, oh, you're squelching conservative voices.
682
00:37:31.921 --> 00:37:34.820
That's what is very likely to end up on the regulatory regime.
683
00:37:35.362 --> 00:37:36.240
or the legislative regime.
684
00:37:36.860 --> 00:37:38.960
You know, I'd love if Paul Cristiano got, you know,
685
00:37:39.866 --> 00:37:41.720
something in edgewise into that conversation,
686
00:37:41.820 --> 00:37:43.319
and maybe he will, but so far,
687
00:37:43.581 --> 00:37:45.700
it doesn't seem like that's predominantly what we're hearing
688
00:37:46.044 --> 00:37:47.220
from the people who actually make the laws.
689
00:37:49.122 --> 00:37:50.940
There's one thing that we're familiar with in the crypto space.
690
00:37:51.221 --> 00:37:54.900
It's tribes and certain, like, cultural leaders
691
00:37:55.262 --> 00:37:59.140
in respective tribes becoming the hardliners for that tribe, right?
692
00:37:59.200 --> 00:38:01.840
They really, they teach the tribe who, you know,
693
00:38:02.081 --> 00:38:04.760
vibe associate with the cultural leader how to be a hardliner.
694
00:38:05.262 --> 00:38:09.060
and then like the people that listen to like the high priest of this particular alignment camp
695
00:38:10.365 --> 00:38:14.640
absorb some not all of the message right like oh yeah most of this stuff feels good to me right we
696
00:38:14.680 --> 00:38:19.360
see this in literally every single crypto tribe that has ever been birthed and and we all have
697
00:38:19.440 --> 00:38:23.880
like one at least one respective hardliner and that's what like kind of more or less creates the
698
00:38:23.920 --> 00:38:29.700
tribe and then also you smash the particle of social media into that and like everything gets
699
00:38:29.961 --> 00:38:30.843
just gets juiced.
700
00:38:31.886 --> 00:38:36.560
Especially this last cycle, as like Bitcoin ETFs have gotten approved and like
701
00:38:36.580 --> 00:38:39.051
BlackRock is starting to tokenize securities on Ethereum.
702
00:38:39.593 --> 00:38:41.060
I've very quickly realized that like
703
00:38:41.220 --> 00:38:46.500
the realm of crypto Twitter is actually a significant minority of like what actually
704
00:38:46.580 --> 00:38:47.443
matters in the space.
705
00:38:48.085 --> 00:38:52.259
And like if you're in crypto Twitter, you think crypto Twitter is massive.
706
00:38:52.400 --> 00:38:52.460
You
707
00:38:52.621 --> 00:38:54.748
think it's a really big deal because you are in it.
708
00:38:55.090 --> 00:38:56.114
And maybe it actually is.
709
00:38:56.937 --> 00:38:57.860
But like I remember
710
00:38:57.940 --> 00:39:02.620
hearing this like line from Sam Harris when he was critiquing Elon Musk's takeover of Twitter is like
711
00:39:03.042 --> 00:39:07.120
Sam Harris was saying like Elon Musk is like disconnected from reality because he's on Twitter
712
00:39:07.702 --> 00:39:12.940
all of the time he thinks the universe is Twitter and this is this is where the these tribes fight
713
00:39:13.040 --> 00:39:17.660
right this is where the hardliners you know chant their chant and they'll post out their tweets like
714
00:39:17.901 --> 00:39:22.280
get the their tribe on board and then everyone else thinks that this is the universe and so I
715
00:39:22.420 --> 00:39:27.640
guess like zooming out Eric like you you talk to all the technologists of Silicon Valley like
716
00:39:27.841 --> 00:39:34.340
can you like help us understand the scope of this like is this like a broad encompassing blanket
717
00:39:34.481 --> 00:39:39.380
that is over all of silicon valley and like government and like everyone is focused on this
718
00:39:39.440 --> 00:39:45.440
conversation or is this like kind of like a side quest from like the vcs and the builders and like
719
00:39:45.500 --> 00:39:49.860
it's actually not really defining the the game the landscape that much like how how far are we on
720
00:39:49.880 --> 00:39:56.300
these like continuum here's continuums i think the answer is is both similarly you know people say
721
00:39:56.742 --> 00:39:57.726
Twitter is not the real world.
722
00:39:58.047 --> 00:39:59.754
Most people in the world are not on Twitter.
723
00:40:00.618 --> 00:40:01.140
That said,
724
00:40:01.420 --> 00:40:06.499
the people who are influential in the world, who influence other people's views, whether it's
725
00:40:06.740 --> 00:40:12.860
journalists or politicians or people on TV or anyone who's got a mouthpiece to the rest of the
726
00:40:12.860 --> 00:40:14.325
world are on Twitter.
727
00:40:14.425 --> 00:40:17.395
Not only are they on Twitter, they get their ideas from Twitter.
728
00:40:17.656 --> 00:40:18.298
They battle it
729
00:40:19.020 --> 00:40:22.070
out in the marketplace of ideas or the culture war.
730
00:40:22.592 --> 00:40:24.658
And so I disagree with Sam Harris.
731
00:40:24.719 --> 00:40:25.100
I do think
732
00:40:25.240 --> 00:40:32.660
Twitter is the source where a lot of ideas come from and a lot of the sort of ideas get litigated
733
00:40:33.302 --> 00:40:33.703
on Twitter.
734
00:40:34.044 --> 00:40:39.160
And similarly, most of Silicon Valley is just doing their jobs, not on Twitter,
735
00:40:39.701 --> 00:40:40.584
not focused on this.
736
00:40:41.106 --> 00:40:43.414
But where do things get litigated?
737
00:40:43.755 --> 00:40:44.919
Often on Twitter, right?
738
00:40:45.059 --> 00:40:45.260
Like,
739
00:40:47.303 --> 00:40:51.734
you know, Aziz was just saying how EA has not had a big impact yet.
740
00:40:52.176 --> 00:40:53.820
We alluded to this off camera,
741
00:40:53.920 --> 00:41:00.380
But there's a world where Sam Altman did get fired, and there was an EA coup, and maybe
742
00:41:00.380 --> 00:41:03.500
there was overblown, but there seemed to be some EA concerns from the board, and maybe
743
00:41:04.142 --> 00:41:09.260
OpenAI would have taken a much more sort of decelerationist stance, right?
744
00:41:09.280 --> 00:41:15.440
There was this movement to pause AI about a year ago that it didn't seem like people
745
00:41:15.541 --> 00:41:16.992
followed, but maybe they would have.
746
00:41:19.352 --> 00:41:19.620
Go for it.
747
00:41:20.702 --> 00:41:29.500
Yeah, I want to push back against this because like this again was the framing of the open AI coup basically in like the couple days when it was all fog of war and nobody knew what was actually going on.
748
00:41:29.800 --> 00:41:33.560
There's now been a ton of reporting on what actually happened in the open AI coup.
749
00:41:34.140 --> 00:41:42.820
And it looks like just a regular old, you know, like Sam Altman was like kind of telling people different things and he was like raising money for stuff on the side and not telling the board and blah, blah, blah.
750
00:41:43.242 --> 00:41:47.620
there's just like actually there was not oh there's q star it's a secret algorithm that's
751
00:41:47.640 --> 00:41:51.740
gonna ai risk and you know people saw that like oh my god you know sam allman wants to unleash it
752
00:41:51.760 --> 00:41:55.780
on the world and he's not being safe enough that was not what happened like we know now that's not
753
00:41:55.800 --> 00:42:00.600
what happened right what happened was just ordinary board loses trust with ceo ceo and board fight
754
00:42:01.282 --> 00:42:05.900
board is like very amateurish ousta ceo hire somebody you know it was all just like terrible
755
00:42:06.081 --> 00:42:10.640
board management but wow really because like to this day i have always thought that it was
756
00:42:10.902 --> 00:42:12.157
an AI safety issue.
757
00:42:12.844 --> 00:42:14.840
No, there's been so much reporting right now.
758
00:42:15.000 --> 00:42:15.340
This is new to me right now.
759
00:42:15.360 --> 00:42:18.160
The New York Times has done a ton of exposés about this now.
760
00:42:19.043 --> 00:42:21.020
There was almost nothing to do with AI safety
761
00:42:21.222 --> 00:42:22.636
and the reason why OpenAI...
762
00:42:23.061 --> 00:42:25.020
Because at the end of the day,
763
00:42:26.285 --> 00:42:27.020
the whole
764
00:42:27.404 --> 00:42:28.880
non-profit structure of
765
00:42:29.001 --> 00:42:30.880
OpenAI was not the reason why they fired Sam.
766
00:42:31.702 --> 00:42:32.680
The reason why they fired Sam
767
00:42:32.882 --> 00:42:34.840
was that the board members lost trust with Sam.
768
00:42:35.140 --> 00:42:37.020
Sam reportedly was telling different people
769
00:42:37.060 --> 00:42:38.012
on the board different things.
770
00:42:38.315 --> 00:42:38.720
The board was
771
00:42:38.740 --> 00:42:41.720
like getting back together and they were very uncomfortable with the way that Sam was managing
772
00:42:41.820 --> 00:42:43.909
things and he was kind of controlling people.
773
00:42:43.989 --> 00:42:46.580
And he had, people talked a lot about the fact
774
00:42:46.600 --> 00:42:49.736
that Sam is a very manipulative guy and he's very, very effective at what he does.
775
00:42:50.077 --> 00:42:50.640
Part of what makes
776
00:42:50.660 --> 00:42:54.440
him a very good leader is that he's very, he has this reality distortion field and he makes people
777
00:42:54.500 --> 00:42:56.628
feel how he wants them to feel when he's around them.
778
00:42:57.231 --> 00:42:58.174
That's the real story.
779
00:42:58.556 --> 00:42:59.700
If you try to tell
780
00:42:59.800 --> 00:43:05.460
the story about, well, opening eye was an EA coup, then you have a lot more explaining to do about
781
00:43:05.520 --> 00:43:10.340
like okay what was the ea thing that happened that precipitated all this perception is reality right
782
00:43:10.460 --> 00:43:16.420
and it was exactly no it's true it's true it was framed in that way by like the the chorus you know
783
00:43:16.420 --> 00:43:20.320
just the people on twitter the talking heads who wanted this to be the story it's the same thing
784
00:43:20.340 --> 00:43:24.720
with like ea versus eac right people want this to be the story but the story is almost always
785
00:43:24.800 --> 00:43:28.420
more complicated and the reality of opening eye was that it was more complicated than that
786
00:43:29.302 --> 00:43:29.462
Yes.
787
00:43:29.823 --> 00:43:37.160
But if it had been completed, this coup that ever was, it would be framed as an EA coup, right?
788
00:43:38.203 --> 00:43:38.998
Yes, yes.
789
00:43:39.841 --> 00:43:43.065
Which, so let me refer in a different way.
790
00:43:44.107 --> 00:43:48.413
You're saying, hey, the people who actually write the laws, they don't care about AI safety.
791
00:43:48.433 --> 00:43:49.735
They're not AI safety people.
792
00:43:49.755 --> 00:43:52.138
They care about jobs and they care about misinformation.
793
00:43:52.539 --> 00:43:53.580
And I agree with that.
794
00:43:53.800 --> 00:44:04.080
And this gets to the sort of Baptist bootleggers sort of dichotomy that Mark and Jason often brings up, which is this idea that there are people who are purists.
795
00:44:04.300 --> 00:44:09.980
They actually believe in things like AI safety and they really care about the issues.
796
00:44:10.520 --> 00:44:19.280
And there are bootleggers who are sort of either grifters or opportunists who notice something that is in their self-interest to take the same position as the Baptists.
797
00:44:19.300 --> 00:44:24.700
the Baptists, and they use the language of the Baptists to justify their own advancement of
798
00:44:24.961 --> 00:44:25.763
their own self-interest.
799
00:44:26.244 --> 00:44:31.599
And similarly, in tech, when right now there's a schism within tech,
800
00:44:32.422 --> 00:44:36.680
like there was a survey that went out to 10,000 researchers or thousands of researchers, and I
801
00:44:36.720 --> 00:44:43.740
think 10,000 MLIR researchers who work at these big labs expressed some concern over or sympathy
802
00:44:43.941 --> 00:44:44.663
with AI safety.
803
00:44:45.184 --> 00:44:50.820
And so if we as an industry don't present a united front in our belief of these
804
00:44:50.880 --> 00:44:54.732
technologies, it gives ammo to people who want to regulate them, right?
805
00:44:55.895 --> 00:44:57.560
And in the same way that
806
00:44:57.640 --> 00:45:04.636
within crypto, if you have half the community talking about why crypto is bad and regulators
807
00:45:06.320 --> 00:45:11.360
see that, they might say, oh, even crypto people think it's bad or Web3 people think it's bad,
808
00:45:11.720 --> 00:45:12.939
Thus, we need to regulate it.
809
00:45:13.442 --> 00:45:15.960
So there's a concern in the same way that with social media.
810
00:45:16.420 --> 00:45:19.740
Social media people apologize all the time and express their concerns.
811
00:45:19.881 --> 00:45:22.000
And it gave more ammo for regulation.
812
00:45:22.863 --> 00:45:24.880
They have the concern that the same thing will happen here.
813
00:45:25.301 --> 00:45:29.080
The counter to that, which I think is a good counter, is no, it's not about giving them ammo.
814
00:45:29.180 --> 00:45:30.480
It's about self-regulating.
815
00:45:30.901 --> 00:45:37.980
If we as an industry can self-regulate ourselves, maybe that will prevent other people from coming in because we can prevent some of the damages.
816
00:45:38.040 --> 00:45:43.160
is and and you know crypto talking about this a lot too of like you know um getting rid of scams
817
00:45:43.320 --> 00:45:49.100
early so that doesn't give the whole industry uh a bad name well in reality that is what's
818
00:45:49.120 --> 00:45:52.800
happening right it's it's who's doing the ai safety research who's actually trying to be
819
00:45:52.880 --> 00:45:56.740
responsible with this stuff the answer is open ai and anthropic and like the guys who are actually
820
00:45:56.881 --> 00:46:00.460
leading the front are self-regulating because their regulars have no clue what the fuck is
821
00:46:00.500 --> 00:46:04.880
going on they have no idea how to even be productively how to productively give input
822
00:46:05.384 --> 00:46:07.080
into how to guide this technology forward.
823
00:46:07.441 --> 00:46:08.920
So the reality is that, you know,
824
00:46:09.184 --> 00:46:10.040
the concern about regulation,
825
00:46:10.641 --> 00:46:12.060
I mean, it's one that I don't know anybody
826
00:46:12.060 --> 00:46:13.800
who doesn't share the concern over bad regulation.
827
00:46:14.180 --> 00:46:15.540
Bad regulation is bad, right?
828
00:46:15.602 --> 00:46:15.900
Clearly.
829
00:46:17.002 --> 00:46:19.180
But it's also obviously true,
830
00:46:19.341 --> 00:46:20.780
even among the effective accelerationists,
831
00:46:20.962 --> 00:46:21.740
that AI is really powerful.
832
00:46:22.641 --> 00:46:25.440
And AI is going to massively change the playing field
833
00:46:25.542 --> 00:46:26.720
around who controls what.
834
00:46:27.121 --> 00:46:29.180
And right now, you know, they're kind of toys,
835
00:46:29.280 --> 00:46:30.600
like, okay, yeah, you can spam people
836
00:46:30.743 --> 00:46:31.440
or you can, you know,
837
00:46:31.941 --> 00:46:34.020
spearfish much more efficiently using these tools
838
00:46:34.121 --> 00:46:35.660
or they can tell you how to build a bomb,
839
00:46:35.760 --> 00:46:37.160
which like, okay, if you can't figure out how to Google
840
00:46:37.322 --> 00:46:38.820
how to build a bomb, like, you know,
841
00:46:39.081 --> 00:46:41.060
good luck using a large language model to get there.
842
00:46:41.961 --> 00:46:44.500
But like, there's a lot more that's coming
843
00:46:44.541 --> 00:46:45.500
and it's pretty obvious to everybody.
844
00:46:45.580 --> 00:46:46.420
There's a lot more that's coming.
845
00:46:46.740 --> 00:46:49.659
And so I think, you know, whether you agree about,
846
00:46:50.220 --> 00:46:51.280
whether you think the right way to do this
847
00:46:51.360 --> 00:46:53.800
is self-regulation or government regulation
848
00:46:54.022 --> 00:46:56.440
depends on your confidence in the right laws
849
00:46:56.562 --> 00:46:57.900
or right frameworks to arise.
850
00:46:58.400 --> 00:47:00.440
And Silicon Valley obviously has, you know,
851
00:47:00.580 --> 00:47:03.140
a lot of scar tissue in believing that, you know,
852
00:47:03.301 --> 00:47:04.065
governments fuck it up.
853
00:47:04.146 --> 00:47:04.870
Governments are terrible.
854
00:47:04.890 --> 00:47:06.378
You know, everybody hates us.
855
00:47:06.418 --> 00:47:06.700
And so they're
856
00:47:06.700 --> 00:47:10.360
going to create these draconian laws that are going to be about this, you know, power struggle
857
00:47:10.762 --> 00:47:12.953
with government and with special interests or whatever.
858
00:47:13.235 --> 00:47:14.240
And that's going to be the regulatory
859
00:47:14.320 --> 00:47:15.246
regime that we enter into.
860
00:47:15.588 --> 00:47:17.780
But nobody goes around saying like, you know, you know, all this stuff
861
00:47:17.820 --> 00:47:21.500
that OpenAI is doing around safety or this research that Anthropic is doing, it's stupid,
862
00:47:21.700 --> 00:47:23.226
it's worthless, it's slowing us down.
863
00:47:23.427 --> 00:47:24.110
Let's throw it away.
864
00:47:24.492 --> 00:47:25.114
Right.
865
00:47:25.295 --> 00:47:26.520
If effective acceleration has said
866
00:47:26.560 --> 00:47:30.180
that, then I would say, okay, it sounds like you guys are very consistent in believing that AI is
867
00:47:30.341 --> 00:47:30.944
an alloyed good.
868
00:47:31.407 --> 00:47:32.071
But they don't say that.
869
00:47:32.211 --> 00:47:33.720
They don't say like, well, why are these people slowing
870
00:47:33.760 --> 00:47:34.283
themselves down?
871
00:47:34.565 --> 00:47:37.220
Instead, they say, well, you know, government bad, these people good, which
872
00:47:37.260 --> 00:47:38.907
like, okay, I understand where that's coming from.
873
00:47:39.349 --> 00:47:42.080
And I don't think EAs are pro-government,
874
00:47:42.240 --> 00:47:42.361
right?
875
00:47:42.381 --> 00:47:46.140
Which I think is, again, another way that this dichotomy kind of breaks down.
876
00:47:47.040 --> 00:47:49.659
EAs in general think that governments are terrible allocators of goods, right?
877
00:47:49.719 --> 00:47:49.860
I mean,
878
00:47:49.880 --> 00:47:50.744
you can just see it historically.
879
00:47:51.026 --> 00:47:52.553
Governments are incredibly inefficient.
880
00:47:53.136 --> 00:47:53.840
And so I don't think EAs
881
00:47:53.960 --> 00:47:57.520
look at that and say, well, great, government is the natural person to provide this regulatory
882
00:47:57.661 --> 00:48:00.160
regime because they're all brilliant and we love what everything the governments do.
883
00:48:00.701 --> 00:48:03.496
Most EAs are pretty libertarian on almost every single front.
884
00:48:03.516 --> 00:48:04.280
They generally believe that
885
00:48:04.521 --> 00:48:05.544
governments are incredibly wasteful.
886
00:48:05.946 --> 00:48:10.020
But when it comes to AI, the concern is that, well,
887
00:48:10.261 --> 00:48:14.397
is self-regulation enough when you have a tool that is incredibly powerful, right?
888
00:48:14.638 --> 00:48:15.020
So if you
889
00:48:15.060 --> 00:48:19.960
have self-regulation around nuclear weapons, how well is that going to do in an environment where
890
00:48:20.020 --> 00:48:22.413
nuclear weapons are extremely powerful and extremely damaging?
891
00:48:22.675 --> 00:48:23.660
Well, in practice, we've
892
00:48:23.680 --> 00:48:27.180
kind of learned through experience the only way to really control nuclear weapons is with the
893
00:48:27.200 --> 00:48:27.943
monopoly on violence.
894
00:48:28.666 --> 00:48:32.100
It's basically saying, yo, Iran, if you develop nuclear weapons, we're going
895
00:48:32.100 --> 00:48:32.624
to come at you.
896
00:48:32.744 --> 00:48:35.020
So, you know, North Korea, if you develop nuclear weapons, we're going to come at
897
00:48:35.040 --> 00:48:36.390
you because we don't want these proliferate.
898
00:48:36.410 --> 00:48:37.840
We don't want this to just be a tool in everybody's
899
00:48:37.880 --> 00:48:38.161
hands.
900
00:48:38.483 --> 00:48:42.138
We actually think that's a worse equilibrium to have everybody be really powerful.
901
00:48:42.359 --> 00:48:42.660
But this is
902
00:48:42.740 --> 00:48:46.960
kind of what EAC implies, is that the best equilibrium for AI is that everyone has one
903
00:48:47.241 --> 00:48:48.849
and AIs keep other AIs in check.
904
00:48:48.869 --> 00:48:49.934
And it's sort of like we all have nukes.
905
00:48:50.517 --> 00:48:51.240
Maybe that's true,
906
00:48:52.301 --> 00:48:56.060
but there's at least a real possibility that that's not true, that it's more like nukes and
907
00:48:56.060 --> 00:49:00.360
It's not like, you know, I don't know, you know, the books, right?
908
00:49:00.626 --> 00:49:01.059
Which is a thing.
909
00:49:01.140 --> 00:49:02.460
Okay, more education is good for everyone.
910
00:49:03.440 --> 00:49:05.280
So anyway, sorry, Eric, I know you want to jump in.
911
00:49:07.221 --> 00:49:11.500
You are correct in saying that we are accelerating and EAC people are happy with the status quo.
912
00:49:12.761 --> 00:49:15.600
They're just worried that things will change.
913
00:49:15.821 --> 00:49:21.900
Because as you mentioned, the irony, people who are sympathetic with AI safety are the ones driving the acceleration.
914
00:49:22.321 --> 00:49:23.019
People work at these labs.
915
00:49:23.421 --> 00:49:32.100
But there are people who are even further on the side of AI safety who are noticing that contradiction, saying, hey, you claim to be about AI safety, but you're actually accelerating.
916
00:49:32.500 --> 00:49:39.400
People like Eliza, people like Laurent, very smart, brilliant people who are saying, hey, we actually need to slow down.
917
00:49:39.680 --> 00:49:41.600
Like, you need to be more, we need to pause AI.
918
00:49:42.041 --> 00:49:43.300
We need to put an off switch.
919
00:49:43.580 --> 00:49:47.220
We need to have some sort of more extreme intervention.
920
00:49:47.801 --> 00:49:50.460
And I think the EAC people are responding to that concern.
921
00:49:51.242 --> 00:49:57.920
Not that the status quo is a problem, but that a very significant change could be a problem.
922
00:49:58.200 --> 00:49:58.940
So that's what they're worried about.
923
00:50:20.903 --> 00:50:23.419
providing a more stable foundation for Mantle's applications.
924
00:50:23.600 --> 00:50:25.579
The Mantle treasury is one of the biggest
925
00:50:25.760 --> 00:50:28.219
DAO owned treasuries, which is seeding an ecosystem
926
00:50:28.380 --> 00:50:31.199
of projects from all around the Web3 space for Mantle.
927
00:50:31.280 --> 00:50:32.759
Mantle already has sub communities
928
00:50:32.820 --> 00:50:34.159
from around Web3 onboarded,
929
00:50:34.220 --> 00:50:36.039
like Game7 for Web3 gaming,
930
00:50:36.180 --> 00:50:38.739
and Bybit for TVL and liquidity and on-ramps.
931
00:50:38.860 --> 00:50:40.599
So if you want to build on the Mantle network,
932
00:50:40.720 --> 00:50:42.318
Mantle is offering a grants program
933
00:50:42.480 --> 00:50:44.018
that provides milestone based funding
934
00:50:44.220 --> 00:50:46.178
to promising projects that help expand,
935
00:50:46.460 --> 00:50:48.279
secure and decentralize Mantle.
936
00:50:48.360 --> 00:50:49.560
If you want to get started working
937
00:50:49.580 --> 00:50:51.919
with the first DAO-led Layer 2 ecosystem,
938
00:50:52.020 --> 00:50:54.137
check out Mantle at Mantle.xyz
939
00:50:54.460 --> 00:50:56.237
and follow them on Twitter at ZeroxMantle.
940
00:50:56.640 --> 00:50:58.919
Celo is the mobile-first, EVM-compatible,
941
00:50:59.040 --> 00:51:01.379
carbon-negative blockchain built for the real world.
942
00:51:01.560 --> 00:51:03.899
Driving real-world use cases like mobile payments
943
00:51:04.000 --> 00:51:05.916
and mobile DeFi, and with Opera MiniPay
944
00:51:06.340 --> 00:51:08.138
as one of the fastest-growing Web3 wallets,
945
00:51:08.340 --> 00:51:10.179
Celo is seeing a meteoric rise
946
00:51:10.441 --> 00:51:12.058
with over 300 million transactions
947
00:51:12.220 --> 00:51:14.558
and 1.5 million monthly active addresses.
948
00:51:14.780 --> 00:51:16.838
And now, Celo is looking to come home to Ethereum
949
00:51:17.080 --> 00:51:24.900
As a Layer 2, Optimism, Polygon, Matterlabs, and Arbitrum have all thrown their hats in the ring for the Celo Layer 2 to build upon their stacks.
950
00:51:25.100 --> 00:51:25.739
Why the competition?
951
00:51:26.080 --> 00:51:33.460
The Celo Layer 2 will bring huge advantages like a decentralized sequencer, off-chain data availability secured by Ethereum validators, and one-block finality.
952
00:51:33.782 --> 00:51:34.620
What does that all mean for you?
953
00:51:34.800 --> 00:51:43.520
With Celo Layer 2, gas fees will stay low and you can even pay for gas natively using ERC-20 tokens, sending crypto to phone numbers across wallets using Social Connect.
954
00:51:43.680 --> 00:51:45.900
But Celo is a community-governed protocol.
955
00:51:46.121 --> 00:51:49.060
This means that Celo needs you to weigh in and make your voice heard.
956
00:51:49.240 --> 00:51:50.920
Join the conversation in the Celo forums.
957
00:51:51.140 --> 00:51:54.980
Follow Celo on Twitter and visit Celo.org to shape the future of Ethereum.
958
00:52:16.241 --> 00:52:20.280
to make sure that you have the best advice and Toku can help to optimize your taxes so you pay
959
00:52:20.340 --> 00:52:23.415
the least possible amount of taxes while still maintaining legal compliance.
960
00:52:23.596 --> 00:52:24.420
With Toku's guidance,
961
00:52:24.540 --> 00:52:27.995
you can concentrate on building your company while Toku handles the logistics.
962
00:52:28.116 --> 00:52:29.040
Token launches don't
963
00:52:29.100 --> 00:52:29.904
have to be complicated.
964
00:52:30.085 --> 00:52:33.138
Talk to Toku today to get a free initial token valuation.
965
00:52:33.158 --> 00:52:33.500
Welcome to
966
00:52:33.560 --> 00:52:35.367
Consensus 2024.
967
00:52:35.467 --> 00:52:38.939
Consensus is one of the biggest hubs for all things crypto, blockchain, and web3.
968
00:52:39.220 --> 00:52:42.917
This May, Consensus is celebrating 10 years of decentralizing the future.
969
00:52:42.937 --> 00:52:43.560
Going to Consensus
970
00:52:43.721 --> 00:52:47.619
means being in good company with over 15,000 attendees from over a hundred countries.
971
00:52:47.800 --> 00:52:47.860
You
972
00:52:47.940 --> 00:52:51.980
have the opportunity to learn directly from the architects and advocates representing Bitcoin,
973
00:52:52.180 --> 00:52:56.140
Ethereum, Solana, along with the teams from popular layer twos like Mantle, Arbitrum,
974
00:52:56.380 --> 00:52:57.747
Optimism, Base, and others.
975
00:52:57.887 --> 00:53:00.560
But consensus is more than just discussions and presentations.
976
00:53:01.020 --> 00:53:04.760
Badge holders get access to witness 10 professional karate combat fights and one
977
00:53:04.921 --> 00:53:06.750
unprofessional one between myself and Nick Carter.
978
00:53:06.770 --> 00:53:08.780
For the visionaries, Pitchfest is your spotlight,
979
00:53:08.980 --> 00:53:12.516
presenting a stage for the most promising early stage Web3 companies.
980
00:53:12.597 --> 00:53:13.260
And for the devs,
981
00:53:13.401 --> 00:53:17.660
the hackathon offers a unique chance to build your next project or take your existing project
982
00:53:17.861 --> 00:53:21.860
to the next level immerse yourself in hundreds of side events and hacker houses scattered around
983
00:53:22.060 --> 00:53:27.620
austin register for consensus 2024 today and save 20 with code bankless yeah yeah i think i think
984
00:53:27.821 --> 00:53:32.800
that draw draws something that we see on both the extremes and something that elizur yukowski and
985
00:53:32.920 --> 00:53:38.339
like um beth jesus might also agree it's like there's another more cynical interpretation
986
00:53:38.981 --> 00:53:42.279
of the stuff that the ea people like the same altman's are doing and the reason they're
987
00:53:42.560 --> 00:53:43.564
spending time in DC.
988
00:53:43.624 --> 00:53:46.012
And that interpretation is regulatory capture.
989
00:53:46.553 --> 00:53:48.540
So they actually want
990
00:53:49.001 --> 00:53:56.659
to hit the accelerator on AI in general, but they're also reasonable people.
991
00:53:56.779 --> 00:53:57.080
And they're
992
00:53:57.261 --> 00:54:03.100
happy to work with the government in order to come up with sensible regulation and partner with them
993
00:54:03.260 --> 00:54:05.351
because we're the educated ones, we're building the models.
994
00:54:05.813 --> 00:54:07.260
And so, yeah, let's team up with the
995
00:54:07.320 --> 00:54:13.060
regulators and the governments around the world in order to kind of create the right rules for
996
00:54:13.100 --> 00:54:14.627
this new and emerging industry.
997
00:54:14.708 --> 00:54:17.400
And oh, by the way, this would be, I think, the Marc Andreessen
998
00:54:17.520 --> 00:54:17.961
critique.
999
00:54:18.342 --> 00:54:22.654
By the way, we are the recipients of value.
1000
00:54:23.416 --> 00:54:24.539
We have cemented ourselves.
1001
00:54:24.559 --> 00:54:24.960
We've created
1002
00:54:25.040 --> 00:54:27.129
a moat for ourselves and our own businesses.
1003
00:54:27.189 --> 00:54:30.020
And so there's a criticism that comes from Eliezer
1004
00:54:30.140 --> 00:54:30.623
in doing this.
1005
00:54:30.643 --> 00:54:32.371
He's saying, well, this is hypocritical.
1006
00:54:32.391 --> 00:54:34.100
If you're really worried about
1007
00:54:34.200 --> 00:54:36.510
AI safety, you would like cease immediately.
1008
00:54:36.530 --> 00:54:38.740
You would stop like doing this until you do a
1009
00:54:38.780 --> 00:54:40.107
Manhattan project and go figure it out.
1010
00:54:40.168 --> 00:54:42.300
And then there's a criticism from the Beth Jaysos
1011
00:54:43.142 --> 00:54:45.008
side of things with Sam Altman.
1012
00:54:46.353 --> 00:54:48.600
Beth would probably be someone who says like, this is
1013
00:54:48.660 --> 00:54:49.907
clearly regulatory capture.
1014
00:54:49.947 --> 00:54:51.374
They're doing it in broad daylight, right?
1015
00:54:51.394 --> 00:54:52.420
The only solution for this
1016
00:54:52.981 --> 00:54:59.540
is decentralized AI and everyone needs like a model and an AI genie in their own home running
1017
00:54:59.640 --> 00:55:03.980
on their own uh like hardware right so this is a critique that's coming from both sides and i'm
1018
00:55:04.181 --> 00:55:09.039
wondering hasib like what about that more cynical interpretation of of what's going on like in in
1019
00:55:09.300 --> 00:55:14.860
fact wouldn't that be like um actually the game theory of what companies want to do they want to
1020
00:55:15.041 --> 00:55:20.400
go build monopolies and so i mean that that's a sensible uh business strategy uh sam albin's
1021
00:55:20.600 --> 00:55:25.540
sure individual he's he's great at this right yeah okay so first of all i'd say i find this a
1022
00:55:25.540 --> 00:55:30.040
is a little ironic because again, during the whole AI ousting or open AI ousting drama,
1023
00:55:30.940 --> 00:55:35.220
Sam Allman was typecast as like the Promethean, you know, like the, the EAC guy who was trying
1024
00:55:35.280 --> 00:55:36.325
to make AI go faster.
1025
00:55:36.647 --> 00:55:38.375
And now all of a sudden, Oh no, no, actually, nevermind.
1026
00:55:38.657 --> 00:55:39.340
He's not the savior.
1027
00:55:39.481 --> 00:55:41.092
He's actually the, the, he's the boogeyman.
1028
00:55:41.112 --> 00:55:42.380
He's actually going out there and trying to,
1029
00:55:42.700 --> 00:55:45.352
you know, he's the JP Morgan trying to corner the steel trade or whatever.
1030
00:55:46.116 --> 00:55:46.397
I don't know.
1031
00:55:46.679 --> 00:55:46.920
To me,
1032
00:55:47.060 --> 00:55:47.764
this is not a critique.
1033
00:55:47.784 --> 00:55:48.989
This is an innuendo, right?
1034
00:55:49.632 --> 00:55:50.436
What is Sam Allman?
1035
00:55:50.637 --> 00:55:50.818
Okay.
1036
00:55:50.858 --> 00:55:51.320
Let's, let's,
1037
00:55:51.400 --> 00:55:52.326
let's step away from AI.
1038
00:55:52.346 --> 00:55:54.460
Let's talk about crypto because I think, you know, people in this podcast
1039
00:55:54.661 --> 00:55:58.900
understand crypto easier let's pretend that instead of sam allman it's brian brian armstrong
1040
00:55:59.201 --> 00:56:04.540
okay he goes to dc and he says hey i'm going to engage with dc and try to come up with you know
1041
00:56:04.560 --> 00:56:08.880
sensible legislation would you be like oh my god he's trying to cap his regulatory capture what is
1042
00:56:08.880 --> 00:56:16.160
he doing why is he consorting with the enemy so not not uh brian armstrong but spf that's exactly
1043
00:56:16.280 --> 00:56:22.120
what he was doing yes yes okay so put spf aside let's say brian armstrong brian armstrong does
1044
00:56:22.180 --> 00:56:25.960
go to dc he does do all those things do you say the same thing is happening now it's just brian
1045
00:56:26.000 --> 00:56:31.218
armstrong instead of spf no but i would be cautious generally i'm why not why not why not
1046
00:56:32.382 --> 00:56:36.460
just just because i mean for me is this a be cautious what do you mean by be cautious what
1047
00:56:36.460 --> 00:56:41.520
do you mean be cautious i think i think it's both it's he's both fighting for regulatory arbitrage
1048
00:56:42.122 --> 00:56:46.500
in the way maybe implicitly but he's also fighting for the industry as a whole
1049
00:56:47.865 --> 00:56:52.020
uh yes regulatory capture yes but he's me i think he's motivated by all the right reasons
1050
00:56:52.645 --> 00:56:54.600
and then the effect of that is regulatory capture.
1051
00:56:55.280 --> 00:56:55.400
Sure.
1052
00:56:55.480 --> 00:56:55.600
Okay.
1053
00:56:55.923 --> 00:56:57.219
So it sounds like what you're saying is,
1054
00:56:57.661 --> 00:56:58.179
hold up a sec.
1055
00:56:58.921 --> 00:57:00.619
Let's see what he actually does.
1056
00:57:01.328 --> 00:57:01.460
Yes.
1057
00:57:01.500 --> 00:57:03.180
And then judge him based on the consequences
1058
00:57:03.463 --> 00:57:05.040
of the legislation that's actually proposed.
1059
00:57:05.100 --> 00:57:05.233
Yes.
1060
00:57:05.380 --> 00:57:05.578
Yes.
1061
00:57:05.660 --> 00:57:05.770
Okay.
1062
00:57:06.184 --> 00:57:07.420
That sounds very reasonable, right?
1063
00:57:07.461 --> 00:57:08.200
That sounds very reasonable.
1064
00:57:08.320 --> 00:57:08.432
Okay.
1065
00:57:08.822 --> 00:57:10.640
What is the legislation that Sam Allman has proposed?
1066
00:57:12.402 --> 00:57:13.479
I actually haven't seen anything.
1067
00:57:13.664 --> 00:57:14.340
Nobody fucking knows.
1068
00:57:14.703 --> 00:57:15.220
We don't know.
1069
00:57:15.300 --> 00:57:16.080
The answer is we don't know.
1070
00:57:16.160 --> 00:57:17.220
So it's all innuendo, right?
1071
00:57:17.280 --> 00:57:19.280
This is people just making up a fucking story
1072
00:57:19.300 --> 00:57:20.240
because it sounds good.
1073
00:57:20.461 --> 00:57:26.920
The reality is if you're Sam Altman and you're not going to D.C. and you're ignoring people in the White House, you're a fucking idiot.
1074
00:57:27.520 --> 00:57:30.380
You're an idiot whether you're fighting for the industry or whether you're fighting for regulatory capture.
1075
00:57:30.880 --> 00:57:33.200
Like either way, if you just ignore them, you're a fucking moron.
1076
00:57:33.440 --> 00:57:37.380
So until we see what is the actual legislation that he's advocating, we don't know.
1077
00:57:38.000 --> 00:57:43.520
And if you were Sam Altman and you literally just ignored D.C., you would be definitely doing the worst thing possible, right?
1078
00:57:43.801 --> 00:57:49.520
I would much rather Sam Altman be there than whoever else is going to be there instead, you know, some guy from, I don't know, fucking Google or something.
1079
00:57:50.046 --> 00:57:51.200
I would much rather say well than be there.
1080
00:57:51.200 --> 00:57:52.040
I think this is a fair point.
1081
00:57:52.362 --> 00:57:55.520
I want to hear what Eric says on this, but it is a fair point.
1082
00:57:55.600 --> 00:58:08.900
Actually, it was kind of like a critique that I had of like Brian Armstrong and Coinbase of like not engaging in D.C. earlier and sort of letting a vacuum be present so that people like SBF could actually engage.
1083
00:58:09.721 --> 00:58:22.320
The moment I actually, like Bankless in general, so David and I kind of like started, our spidey senses started tingling with SBF was when the draft of some proposed legislation that he was responsible for actually came out.
1084
00:58:22.780 --> 00:58:32.860
And it felt very much like a rug pull, ladder pull up on all DeFi front end interfaces and like brought him to task and said, hey, SBF, this is not like crypto values.
1085
00:58:32.942 --> 00:58:33.739
Like, what are you doing?
1086
00:58:33.800 --> 00:58:40.060
you're um you're essentially forcing it or creating some legislation that would force
1087
00:58:40.140 --> 00:58:44.740
kind of compliance on this space and and promote your centralized exchange and so a critique of
1088
00:58:44.820 --> 00:58:50.400
like brian during that time was like i wish he had been there because if if if if you're saying
1089
00:58:50.480 --> 00:58:55.400
sam altman could be good we don't know it's kind of like you know schrodinger's uh sam altman we'll
1090
00:58:55.600 --> 00:59:01.860
see what the legislation that that pops out on the other side um i guess the the the case for
1091
00:59:01.880 --> 00:59:06.500
that would be if he doesn't do it, then some more nefarious version of a Sam Alton would be in DC
1092
00:59:06.660 --> 00:59:09.331
engaging those regulators and engaging regulatory capture.
1093
00:59:09.672 --> 00:59:11.540
So I suppose we don't know until the
1094
00:59:11.600 --> 00:59:12.642
legislation is out.
1095
00:59:13.304 --> 00:59:14.547
What do you think about this, Eric?
1096
00:59:16.151 --> 00:59:20.180
I agree with Haseeb's framing that Sam has
1097
00:59:20.802 --> 00:59:21.605
to get involved.
1098
00:59:21.706 --> 00:59:24.035
Yeah, let's wait and see what the regulation looks like.
1099
00:59:24.477 --> 00:59:25.180
I would like to see
1100
00:59:25.400 --> 00:59:32.380
AI safety people more strongly disown the AI ethics people because they often get conflated,
1101
00:59:32.901 --> 00:59:33.564
right?
1102
00:59:34.086 --> 00:59:37.660
And because they have a sort of a coincidence of wants motivated by, you know,
1103
00:59:37.720 --> 00:59:39.007
disparate desires.
1104
00:59:39.027 --> 00:59:40.092
Who are the AI ethics people?
1105
00:59:40.112 --> 00:59:41.580
I don't think we've defined that well enough, Eric.
1106
00:59:42.440 --> 00:59:47.920
AI ethics people are people who are worried about misinformation, who are worried about
1107
00:59:48.943 --> 00:59:54.239
sort of inequality about certain people or certain voices being represented by the AI.
1108
00:59:55.645 --> 00:59:56.800
Is this like AI woke people?
1109
00:59:56.920 --> 00:59:58.017
Yes, that's exactly it.
1110
00:59:59.281 --> 01:00:04.520
And these are the concerns that along with jobs and other things that Congress people
1111
01:00:04.540 --> 01:00:10.060
or regulators are more sympathetic to because they don't quite understand the AI safety
1112
01:00:10.201 --> 01:00:12.920
arguments just yet, or it's a bit too abstract.
1113
01:00:13.800 --> 01:00:21.820
And by not disowning those sort of reasons, they sort of implicitly accept them or give
1114
01:00:21.861 --> 01:00:22.480
them more power.
1115
01:00:23.080 --> 01:00:24.617
And I would like to see that.
1116
01:00:24.880 --> 01:00:30.039
And this is why we see things like Google Gemini being super woke or even ChatGPT earlier.
1117
01:00:31.384 --> 01:00:33.060
And these things matter.
1118
01:00:33.640 --> 01:00:36.180
And so that's something that I would like to see.
1119
01:00:37.262 --> 01:00:40.860
One thing I want to return to a little bit earlier is Steve was framing the culture war.
1120
01:00:41.662 --> 01:00:43.840
And when we have a culture war, things get kind of dumbed down.
1121
01:00:44.300 --> 01:00:47.680
And people typically associate culture war with bad, right?
1122
01:00:47.760 --> 01:00:52.560
And when they think of a lack of culture war, they think of, oh, a great marketplace of ideas
1123
01:00:52.640 --> 01:00:56.500
where people are being very sophisticated and presenting good arguments.
1124
01:00:56.920 --> 01:01:00.640
And this, to EA's credit, is often what EA is in many ways.
1125
01:01:00.841 --> 01:01:04.320
It's very intelligent conversation, trying to get the truth in some ways.
1126
01:01:04.860 --> 01:01:10.520
Now, I would also posit that the lack of a culture war is not just this great marketplace of ideas.
1127
01:01:10.702 --> 01:01:12.079
It's also something like North Korea.
1128
01:01:12.761 --> 01:01:16.500
It's something where there's a lack of moral or intellectual diversity.
1129
01:01:17.382 --> 01:01:19.019
It's just like you just have to have this.
1130
01:01:20.063 --> 01:01:21.400
This is a defense that some people give.
1131
01:01:21.703 --> 01:01:23.360
Is this a pro-culture war argument?
1132
01:01:23.721 --> 01:01:24.460
Yes, yes.
1133
01:01:24.904 --> 01:01:25.940
We should have more culture wars.
1134
01:01:26.160 --> 01:01:28.357
It's a pro-polarization argument.
1135
01:01:29.381 --> 01:01:33.860
I think in, and even in EA, I saw this blog post, we can link to it, where someone was
1136
01:01:33.880 --> 01:01:35.560
saying that EA needs more polarization.
1137
01:01:36.241 --> 01:01:37.800
And it actually affects their giving.
1138
01:01:38.442 --> 01:01:42.820
Like EA gave, or, you know, Gible gave like $200 million to criminal reform or something.
1139
01:01:43.341 --> 01:01:45.520
And this guy was arguing that actually made the problem worse.
1140
01:01:46.703 --> 01:01:48.860
And why didn't people come in and measure this?
1141
01:01:48.920 --> 01:01:51.719
because they felt it was too controversial to go against it.
1142
01:01:52.201 --> 01:01:56.840
And this is a critique that some people give of EA, which is as much as they want to claim
1143
01:01:56.920 --> 01:02:01.720
to be rational, first principles, you would think that that would lead to a diversity
1144
01:02:01.720 --> 01:02:02.620
of perspectives.
1145
01:02:02.901 --> 01:02:08.860
And in some cases, it does on issues like long-termism and specific implementation details.
1146
01:02:09.361 --> 01:02:16.840
But it's worth noting that EA is Democrat, that there are very few Trump-supporting EAs.
1147
01:02:17.240 --> 01:02:18.200
Now, why is that?
1148
01:02:18.240 --> 01:02:18.760
Is it possible?
1149
01:02:18.860 --> 01:02:25.500
that everyone in EA from their first principles is just like Trump bad, Biden better, that seems
1150
01:02:25.620 --> 01:02:26.262
very unlikely.
1151
01:02:26.603 --> 01:02:31.979
It seems that EA is also prone to tribal thinking in ways that other communities
1152
01:02:32.421 --> 01:02:32.561
are.
1153
01:02:33.323 --> 01:02:38.680
And that's not a knock on EA, but it's the reason why culture war in this case might be good.
1154
01:02:39.060 --> 01:02:46.180
I'll also posit that EAC is a way for Silicon Valley to implicitly be right wing, implicitly
1155
01:02:46.361 --> 01:02:51.880
be conservative and you notice that eac is not only uh talking about issues like you know advancing
1156
01:02:52.000 --> 01:02:58.120
ai or technology but also wading into other culture it's it's much more masculine it's much
1157
01:02:58.260 --> 01:03:05.940
more alpha it's much you know it's concerned with birth rates right um and so um i think it's good
1158
01:03:06.221 --> 01:03:11.778
for there to be in silicon valley more polarization and more diversity as it relates to um
1159
01:03:12.320 --> 01:03:17.940
as it relates to intellectual issues, it relates to politics, because for so long,
1160
01:03:18.020 --> 01:03:19.123
we've had a lack of that.
1161
01:03:19.504 --> 01:03:24.739
And that's prevented really important conversations on things like,
1162
01:03:25.421 --> 01:03:30.620
you know, criminal justice, for example, or things where there are different, where there are taboos,
1163
01:03:30.941 --> 01:03:36.480
and EA and rationality is supposed to get rid of taboos, or, and in some ways, it hasn't done that.
1164
01:03:36.660 --> 01:03:41.760
So I'm curious how you would react to the characterization of EA as Democrat affiliated,
1165
01:03:41.961 --> 01:03:44.920
or left-wing and not having enough sort of intellectual or moral diversity.
1166
01:03:47.165 --> 01:03:48.580
I think you raise good points.
1167
01:03:49.000 --> 01:03:54.780
And I should probably reiterate that, you know, I've been affiliated with the EA philosophy
1168
01:03:55.023 --> 01:03:56.580
for, you know, almost 10 years now.
1169
01:03:57.882 --> 01:04:03.540
I wrote a post fairly recently kind of critiquing EA's, I think I titled the post, the unreasonable
1170
01:04:03.682 --> 01:04:05.260
ineffectiveness of effective altruists.
1171
01:04:05.980 --> 01:04:11.360
And I think what I've found over time, you know, coming into the crypto space, I kind
1172
01:04:11.380 --> 01:04:15.340
of thought that there would be, you know, when I first got into EA, EA was like, you know,
1173
01:04:15.460 --> 01:04:19.680
these kids who were really, really smart and, you know, made a lot of money and they seem to be
1174
01:04:19.780 --> 01:04:21.853
extremely smart and fastidious and hardworking.
1175
01:04:22.134 --> 01:04:23.060
And I thought, wow, these people are going to
1176
01:04:23.080 --> 01:04:23.904
be super successful.
1177
01:04:24.286 --> 01:04:24.909
I'm a nobody.
1178
01:04:24.989 --> 01:04:26.436
I hope I can be successful too someday.
1179
01:04:26.898 --> 01:04:27.340
And then, you know,
1180
01:04:27.340 --> 01:04:28.990
I found my way into crypto and started building stuff.
1181
01:04:29.031 --> 01:04:30.500
And I knew some EAs who'd gotten into
1182
01:04:30.500 --> 01:04:30.801
crypto.
1183
01:04:31.242 --> 01:04:37.160
One of those EAs was this kid named Sam, who I first met at EA Global, which is like this
1184
01:04:37.200 --> 01:04:39.435
big EA conference in San Francisco in 2015.
1185
01:04:41.241 --> 01:04:42.720
He was still working at Jane Street at that time.
1186
01:04:42.860 --> 01:04:46.360
And I met him in a hallway and we chatted for like five minutes just about blah, blah,
1187
01:04:46.380 --> 01:04:47.040
blah, EA, whatever.
1188
01:04:47.300 --> 01:04:47.740
And, you know, whatever.
1189
01:04:47.840 --> 01:04:48.700
I didn't think much of it.
1190
01:04:48.800 --> 01:04:51.600
He's just, you know, got, you know, a piece of guy curly hair, shorts.
1191
01:04:52.120 --> 01:04:52.660
Little did I know.
1192
01:04:52.740 --> 01:04:53.220
Little did I know.
1193
01:04:53.541 --> 01:04:57.320
I'd come across him again much later when we were both doing very different things in
1194
01:04:57.340 --> 01:04:57.780
the crypto industry.
1195
01:04:57.900 --> 01:04:58.980
Neither of us were in crypto at that time.
1196
01:05:00.381 --> 01:05:04.260
And so I kind of assumed that there would be a lot of people who were really successful
1197
01:05:04.280 --> 01:05:05.080
in crypto who were EAs.
1198
01:05:06.143 --> 01:05:12.020
and what I would later discover many years after being a VC was that that was super not true.
1199
01:05:12.960 --> 01:05:18.400
There was basically only one other person who was a real EA or like a kind of central EA who was
1200
01:05:18.440 --> 01:05:22.357
very successful in crypto and that was Sam and I didn't know Sam super well.
1201
01:05:22.377 --> 01:05:23.060
We chatted maybe three
1202
01:05:23.140 --> 01:05:27.540
four times over the years that we'd interacted with each other but obviously you know we all
1203
01:05:27.821 --> 01:05:33.040
know the story of what happened with Sam and he you know single-handedly kind of redefined EA for
1204
01:05:33.120 --> 01:05:39.820
most people as being this kind of really corrupted kind of corrosive um you know group thinky kind of
1205
01:05:39.880 --> 01:05:45.320
go along with the flow um very left-leaning uh tribe right essentially that that's that's what
1206
01:05:45.380 --> 01:05:49.580
most people think of ea as well well now i'd just like to add a little bit more onto that just like
1207
01:05:49.760 --> 01:05:55.500
sociopathic comes to mind because it was like utilitarianism to its nth degree like not
1208
01:05:55.560 --> 01:06:00.739
considering about like any local humans emotions or anything like that right right and so in these
1209
01:06:01.120 --> 01:06:04.733
famous stories about you know would you bet the entire universe on a 51 49 uh coin flip and he's
1210
01:06:06.560 --> 01:06:11.540
like yes i do it infinitely many times um and so just almost like you know this autistic chasing
1211
01:06:11.861 --> 01:06:17.120
of oh go ahead eric it's worth noting he said that in a podcast and it wasn't like there was this
1212
01:06:17.501 --> 01:06:22.639
you know everyone coming down on him criticizing him he was still a god uh and so it's yeah it's
1213
01:06:23.020 --> 01:06:27.260
crazy the difference between how spf was treated before we found out he was doing a fraud even
1214
01:06:27.280 --> 01:06:30.600
though he was saying some of these things he wouldn't say fraud but he was saying extreme
1215
01:06:30.680 --> 01:06:32.946
utilitarian ideas in broad daylight.
1216
01:06:33.989 --> 01:06:35.513
And so it's just worth noting.
1217
01:06:36.556 --> 01:06:37.137
Yeah, exactly.
1218
01:06:37.478 --> 01:06:38.180
And so it is,
1219
01:06:38.380 --> 01:06:43.578
it is, it's something that we actually talked about, we alluded to briefly earlier about
1220
01:06:44.281 --> 01:06:45.427
extinction risk, right?
1221
01:06:45.528 --> 01:06:47.680
And like, you know, what is the, what is the importance of avoiding
1222
01:06:47.720 --> 01:06:48.342
extinction risk?
1223
01:06:48.924 --> 01:06:51.472
And the reason why SPF's example is wrong.
1224
01:06:51.492 --> 01:06:53.860
So a classical thing in EA is this idea
1225
01:06:54.302 --> 01:06:57.960
of risk neutrality, which is this idea that, okay, you know, most people are risk averse.
1226
01:06:58.322 --> 01:07:02.380
they don't want to take risks and so it's like okay if i can definitely save a life by you know
1227
01:07:02.440 --> 01:07:06.620
going to uh i don't know a food drive and donating some food and i will definitely save a life by
1228
01:07:06.680 --> 01:07:11.960
doing that but it's very very expensive um or i can like maybe probabilistically save a life by
1229
01:07:12.141 --> 01:07:15.460
you know investing into water filtration systems and maybe somewhere down the road that's going to
1230
01:07:15.460 --> 01:07:18.860
result in saving a life most people are like well i want to know for sure that i saved a life
1231
01:07:19.563 --> 01:07:23.780
and ea kind of advocates this risk neutrality which is the idea that look in expectation
1232
01:07:24.463 --> 01:07:27.820
because you know there's millions of people doing millions and millions of things trying to save
1233
01:07:27.880 --> 01:07:31.377
lives, all you really care about is on average, how many lives you save.
1234
01:07:31.397 --> 01:07:32.020
And in the portfolio
1235
01:07:32.462 --> 01:07:36.380
of all the things people are doing to save lives, you should just increase the expected value and
1236
01:07:36.440 --> 01:07:39.713
not worry so much about whether your coin flip turns out heads or tails.
1237
01:07:39.854 --> 01:07:40.015
Right.
1238
01:07:40.376 --> 01:07:41.320
And so, you know,
1239
01:07:42.184 --> 01:07:44.760
SPF kind of took this idea to an extreme and said, okay, well, we should gamble the whole
1240
01:07:44.760 --> 01:07:48.614
fucking world on, you know, plus EV one life.
1241
01:07:48.714 --> 01:07:48.935
Great.
1242
01:07:48.995 --> 01:07:49.155
Okay.
1243
01:07:49.176 --> 01:07:50.440
Flip the coin as many times you
1244
01:07:50.460 --> 01:07:50.601
want.
1245
01:07:50.621 --> 01:07:51.123
That's free money.
1246
01:07:51.926 --> 01:07:52.730
Now this is wrong.
1247
01:07:53.312 --> 01:07:53.493
Okay.
1248
01:07:53.513 --> 01:07:54.397
This is actually just wrong.
1249
01:07:54.417 --> 01:07:55.080
This is wrong.
1250
01:07:55.401 --> 01:07:59.580
according to this mathematical rule called the Kelly Criterion, which basically defines the
1251
01:07:59.740 --> 01:08:04.240
optimal level of risk-taking you should make in order to maximize the growth rate of your bank
1252
01:08:04.260 --> 01:08:04.401
roll.
1253
01:08:04.682 --> 01:08:07.900
I used to be a professional poker player before I ever got into tech or crypto or anything.
1254
01:08:08.320 --> 01:08:12.700
And Kelly Criterion is a concept that everybody in poker knows, is that if you want to make the
1255
01:08:12.700 --> 01:08:14.588
most money, you follow the Kelly Criterion.
1256
01:08:14.608 --> 01:08:17.260
You don't bet your entire bank roll on, oh, okay,
1257
01:08:17.280 --> 01:08:20.295
I've got like, you know, 60% to win this hand if we go all in.
1258
01:08:20.536 --> 01:08:21.440
So great, I'm going to bet my entire
1259
01:08:21.440 --> 01:08:24.099
a bankroll on it because that's the maximal expected value.
1260
01:08:24.341 --> 01:08:24.894
That's wrong.
1261
01:08:25.140 --> 01:08:26.115
That is wrong to do that.
1262
01:08:26.440 --> 01:08:28.936
It is wrong to do that because when you go to zero,
1263
01:08:29.640 --> 01:08:31.639
if there's a real existential risk,
1264
01:08:31.781 --> 01:08:32.751
even if the risk is low,
1265
01:08:33.621 --> 01:08:34.832
you can never come back from zero.
1266
01:08:35.721 --> 01:08:37.540
As long as you still have some bankroll,
1267
01:08:37.580 --> 01:08:38.335
you can make it all back.
1268
01:08:38.640 --> 01:08:40.578
But if you go to zero, you cannot make it all back anymore.
1269
01:08:40.820 --> 01:08:42.655
Your growth rate is now zero forever.
1270
01:08:43.421 --> 01:08:44.680
And so Kelly Criterion tells you
1271
01:08:44.700 --> 01:08:45.995
you should maximize growth rate
1272
01:08:46.440 --> 01:08:47.795
as opposed to maximize expected value.
1273
01:08:48.241 --> 01:08:49.039
And you maximize growth rate
1274
01:08:49.080 --> 01:08:50.138
by following the Kelly Criterion.
1275
01:08:50.260 --> 01:08:51.217
So Sam was wrong.
1276
01:08:52.047 --> 01:08:53.100
in his elocution of that.
1277
01:08:53.281 --> 01:08:55.140
But people kind of get stuck on this idea that,
1278
01:08:55.240 --> 01:08:56.260
oh, well, that's what EAs believe.
1279
01:08:56.640 --> 01:08:58.560
EAs believe that you should bet the universe
1280
01:08:58.804 --> 01:08:59.900
on every tiny little edge
1281
01:08:59.960 --> 01:09:02.000
and every small piece of expected value.
1282
01:09:02.280 --> 01:09:03.059
So again, it's like this,
1283
01:09:03.641 --> 01:09:05.737
it seems like the EA worldview
1284
01:09:06.241 --> 01:09:08.700
just completely diverges
1285
01:09:08.760 --> 01:09:11.280
from what normal people think of as common sense morality.
1286
01:09:12.200 --> 01:09:14.660
And I think it's, SBF's worldview did.
1287
01:09:15.401 --> 01:09:16.580
But most EA's worldviews,
1288
01:09:16.760 --> 01:09:18.340
if you explain them clearly enough,
1289
01:09:18.720 --> 01:09:19.360
it's like, actually, no,
1290
01:09:19.421 --> 01:09:20.800
that actually does sound pretty reasonable.
1291
01:09:20.920 --> 01:09:29.540
Yeah, okay, if you're playing with a technology that might make you go to zero, okay, well, you know, the one rule of gambling is that you never fuck with going to zero, right?
1292
01:09:29.601 --> 01:09:30.620
You never fuck with going broke.
1293
01:09:31.040 --> 01:09:40.240
You always leave some money on the side that if you end up losing the hand, even if you're all in aces versus kings, you still – there's a chance you lose when you go all in aces against kings, right?
1294
01:09:40.320 --> 01:09:44.600
So, okay, that 10% of the time that you wipe out, you want to make sure you've got some money to start grinding it back.
1295
01:09:45.502 --> 01:09:56.060
And in some sense, that is the core idea behind, hey, if you're P-Doom, that's the probability that you think that AI is in some way an existential risk to humanity.
1296
01:09:56.461 --> 01:10:01.940
If your P-Doom is like, you know, 5%, that means to most people, like, oh, okay, well, then why worry?
1297
01:10:02.040 --> 01:10:03.300
Like, obviously, this won't happen.
1298
01:10:03.420 --> 01:10:06.520
And so, you know, what's the chance that a 1 in 20 thing is going to happen?
1299
01:10:06.701 --> 01:10:08.079
Why are you guys freaking out about this?
1300
01:10:08.502 --> 01:10:13.180
But if somebody says, you know, I don't know what my P doom is, but it's probably something on the order of like five ten percent
1301
01:10:13.522 --> 01:10:16.180
Which means the modal outcome nobody ever has to worry about any of this
1302
01:10:16.702 --> 01:10:20.480
None of this is gonna matter most the time you're gonna walk to the world and AIs are just gonna be you know
1303
01:10:20.761 --> 01:10:23.180
Sex robots and you know, it's gonna help you with your taxes and whatever
1304
01:10:23.562 --> 01:10:26.860
But you know, none of them is gonna none of them really gonna try to kill humanity
1305
01:10:27.241 --> 01:10:33.140
Right that most people who are talking about EA believe that they believe that the modal outcome is that nothing bad will happen
1306
01:10:33.983 --> 01:10:35.696
Or nothing bad on a catastrophic scale will happen.
1307
01:10:35.716 --> 01:10:36.280
I should I should clarify
1308
01:10:37.444 --> 01:10:42.660
And, but they say, even if that's true, that 5% of the time, you don't risk that.
1309
01:10:43.162 --> 01:10:43.899
You don't risk that.
1310
01:10:44.180 --> 01:10:55.020
So even if you think, yes, there are going to be trillions of humans in the future, and the EAC worldview is that, well, all those trillions of humans, I agree there will be trillions of humans in the future, and we should care about their well-being just as much as we care about present humans.
1311
01:10:55.861 --> 01:11:05.700
And so therefore, you know, if we wait even a moment longer to let AIs arrive into the world, we are doing a massive amount of harm to all those humans.
1312
01:11:06.321 --> 01:11:08.140
The shape of the argument is kind of similar, right?
1313
01:11:08.420 --> 01:11:23.600
But the Kelly criterion, if you take it seriously, what it implies is that, hey, waiting another couple of years to get to the same place we're eventually going to get to but with much more – with basically minimizing the chance that your bankroll goes to zero is worth it if you're just thinking about Kelly criterion.
1314
01:11:24.301 --> 01:11:38.340
And also with the role of human incentives, I think the gap between the Kelly criterion and what value like maximizing EV ended up being like SPFs like Bahamian penthouse and all of his political donations and everything else.
1315
01:11:40.642 --> 01:11:45.360
One thing I want to speculate with you guys is why is EA this convenient punching bag?
1316
01:11:46.080 --> 01:11:49.460
And it's a punching bag on contradictory statements, right?
1317
01:11:49.500 --> 01:11:55.060
Like the AI ethics people for years have been calling EA like transhumanists are saying, hey, you're way too tech oriented.
1318
01:11:56.081 --> 01:12:03.420
And of course, the people call them doomers like, you know, EA is has a branding problem.
1319
01:12:03.460 --> 01:12:04.586
It's been having a branding problem.
1320
01:12:04.787 --> 01:12:07.080
And it's it's it's it was a punching bag after SBF.
1321
01:12:07.180 --> 01:12:08.726
It was a punching bag for open AI.
1322
01:12:08.986 --> 01:12:12.680
It's basically something that everybody else can rally around of why they dislike it.
1323
01:12:12.980 --> 01:12:20.900
right and and it continues to have this trouble and i would pause it it's because ea has beta
1324
01:12:21.080 --> 01:12:28.619
energy ea is a has nerdy beta energy and it doesn't stand up for itself it doesn't say you
1325
01:12:29.081 --> 01:12:33.260
know uh like after the spf thing it goes it apologizes it goes into hiding after opening
1326
01:12:33.320 --> 01:12:37.339
i think also like why did they say no this was totally fake you're making this up actually
1327
01:12:37.781 --> 01:12:40.459
like EA has done some amazing things, right?
1328
01:12:40.701 --> 01:12:43.219
EA has had incredible accomplishments and yet
1329
01:12:44.681 --> 01:12:47.360
they're not shouting them off of the rooftops in the way that Mike Solana
1330
01:12:47.601 --> 01:12:51.860
is shouting off, you know, Alpha aggressively saying why tech is good.
1331
01:12:52.341 --> 01:12:55.340
And this is where I posit when I, you know, somewhat joking about culture
1332
01:12:55.360 --> 01:12:57.840
we're good, but I think EA and EAC need each other.
1333
01:12:58.141 --> 01:13:02.440
Like I want an ecosystem where EAC and EA are duking it out
1334
01:13:02.802 --> 01:13:03.717
because I think we need both.
1335
01:13:04.021 --> 01:13:07.019
I think we need a movement that is explicitly,
1336
01:13:07.682 --> 01:13:14.199
unadulterated, pro-tech, alpha energy, willing to fight back against people who are trying to
1337
01:13:14.620 --> 01:13:19.075
unfairly critique it or get our CEOs fired or put in dumb regulations.
1338
01:13:19.516 --> 01:13:20.460
But I also want another
1339
01:13:20.700 --> 01:13:25.480
movement that keeps that movement in check and says, hey, AI safety is actually this really
1340
01:13:25.520 --> 01:13:26.183
important thing.
1341
01:13:26.443 --> 01:13:27.788
We should really look into it.
1342
01:13:27.808 --> 01:13:31.040
We should self-regulate and have that fight
1343
01:13:31.220 --> 01:13:33.727
internally as opposed to externally.
1344
01:13:34.288 --> 01:13:38.660
And we've seen when EA gets too powerful, it becomes this
1345
01:13:38.740 --> 01:13:40.706
massive punching bag.
1346
01:13:40.987 --> 01:13:45.660
At the same time, EAC, when it gets too powerful, it becomes really cringe,
1347
01:13:45.940 --> 01:13:46.121
right?
1348
01:13:46.742 --> 01:13:51.436
Like EAC is at its best when it's duking it out with EA.
1349
01:13:51.717 --> 01:13:52.640
When it's trying to be this
1350
01:13:52.941 --> 01:13:56.316
overarching thing, it can't really... EAC is like an underdog movement.
1351
01:13:56.336 --> 01:13:56.939
They're both like
1352
01:13:57.321 --> 01:14:03.580
underdog movements that should be in productive tension with each other for us to figure out.
1353
01:14:04.041 --> 01:14:06.129
And maybe the status quo is actually like pretty good.
1354
01:14:06.350 --> 01:14:08.620
Like we have labs who are accelerating,
1355
01:14:09.162 --> 01:14:11.173
and yet they're also working on alignment research.
1356
01:14:11.193 --> 01:14:12.480
They're also trying to keep each other
1357
01:14:13.223 --> 01:14:13.605
in check.
1358
01:14:13.966 --> 01:14:17.059
And if we could just keep this productive tension, or what I'm calling
1359
01:14:17.240 --> 01:14:23.000
productive tension, maybe we can get to a better solution here, as opposed to the worst of these
1360
01:14:23.060 --> 01:14:23.704
extremes.
1361
01:14:23.925 --> 01:14:25.614
You talked about the manifesto and how extreme that is.
1362
01:14:25.634 --> 01:14:26.720
And we talked about some of the
1363
01:14:26.800 --> 01:14:31.839
extreme players in the ai safety and how bad that is so that's uh something i would posit why ea
1364
01:14:32.401 --> 01:14:37.120
gets beaten up so much and why it needs uh eac support to help defend the parts of it that that
1365
01:14:37.301 --> 01:14:43.600
they agree with which is 98 percent in in in in my belief um and then also why the why the tension is
1366
01:14:43.740 --> 01:14:49.240
helpful i agree with you that ea is a punching bag it's it's weird to me in a way so you know i was
1367
01:14:49.461 --> 01:14:53.100
in ea before ea got really on the global stage right most people didn't know what an ea was when
1368
01:14:53.120 --> 01:14:53.739
when I was first in EA.
1369
01:14:54.341 --> 01:14:56.680
And then really kind of in large part due to SBF,
1370
01:14:57.501 --> 01:14:59.400
EA suddenly became very high status, right?
1371
01:14:59.480 --> 01:15:02.397
During SBF's reign, EA, like it was almost crazy
1372
01:15:03.121 --> 01:15:05.180
the amount of positive coverage that EA was getting,
1373
01:15:05.521 --> 01:15:07.940
despite the fact that EA is like objectively really weird
1374
01:15:08.845 --> 01:15:11.420
and very unapproachable for most people, right?
1375
01:15:11.641 --> 01:15:13.400
So it kind of got lumped into this
1376
01:15:13.440 --> 01:15:15.079
just like general asceticism
1377
01:15:15.422 --> 01:15:17.780
and like sagely, saintly energy
1378
01:15:17.841 --> 01:15:19.920
that SBF was trying to portray himself as being.
1379
01:15:21.009 --> 01:15:22.220
and the
1380
01:15:23.753 --> 01:15:24.439
like in a way
1381
01:15:25.267 --> 01:15:26.559
two things I'd say one is that
1382
01:15:27.932 --> 01:15:28.740
that made me very uncomfortable
1383
01:15:28.922 --> 01:15:30.800
this idea that like wow EA is like so good
1384
01:15:30.840 --> 01:15:32.680
and everything about you is good and everything that is
1385
01:15:32.841 --> 01:15:34.560
EA is good and now EA
1386
01:15:35.024 --> 01:15:36.740
is actually really cringe right it's like not
1387
01:15:36.801 --> 01:15:38.720
cool you go to Silicon Valley party
1388
01:15:38.821 --> 01:15:40.619
and you tell someone you're an EA and they're like oh okay
1389
01:15:40.842 --> 01:15:42.000
well interesting
1390
01:15:42.020 --> 01:15:44.055
and that it's definitely
1391
01:15:44.841 --> 01:15:46.620
not something that people continue to self affiliate
1392
01:15:46.822 --> 01:15:48.840
with which is part of the reason why I also
1393
01:15:48.920 --> 01:15:49.543
like it more now?
1394
01:15:50.065 --> 01:15:53.400
Because I feel like, okay, if you're EA now, then you actually, it requires
1395
01:15:53.400 --> 01:15:54.144
some balls, right?
1396
01:15:54.184 --> 01:15:56.578
You're basically taking a hit publicly to say that you're an EA.
1397
01:15:56.598 --> 01:15:56.960
Right,
1398
01:15:57.000 --> 01:15:57.342
exactly.
1399
01:15:57.362 --> 01:15:58.246
It's an EA bear market.
1400
01:15:58.266 --> 01:16:01.340
I think like that is kind of the valley of darkness that EA needs
1401
01:16:01.862 --> 01:16:05.540
in order to really define its identity and not just be like this kind of milquetoast, like,
1402
01:16:05.640 --> 01:16:07.429
oh, I'm smart and cool and good.
1403
01:16:07.871 --> 01:16:09.640
And isn't that wonderful that I'm all those things at the same
1404
01:16:09.700 --> 01:16:09.921
time?
1405
01:16:10.985 --> 01:16:13.032
And so that's one thing I'll say.
1406
01:16:13.093 --> 01:16:15.100
I think that it has been good for EA in a certain sense
1407
01:16:15.462 --> 01:16:18.820
to kind of get stress tested by becoming very politically unpopular.
1408
01:16:20.281 --> 01:16:25.140
The second thing I'll say is, OK, why is EA such a punching bag in a way that, for example,
1409
01:16:25.260 --> 01:16:28.575
tech was not post Theranos or post, I don't know, crypto.
1410
01:16:30.101 --> 01:16:33.740
Crypto obviously was a punching bag post SBF blowing up, but EA as well.
1411
01:16:34.340 --> 01:16:40.560
And I think a big part of the reason for that is that EA is just not designed to be politically capable.
1412
01:16:41.500 --> 01:16:44.040
It's almost premised on this idea that politics is bad.
1413
01:16:44.444 --> 01:16:45.414
And politics is the mind killer.
1414
01:16:46.540 --> 01:17:03.440
And if you believe this and everything that you preach is about nuance and careful thinking and avoiding sloganism and not being lazy and being suspicious of power structures, and then you're like, great, now let's go fight in the biggest power game of all, which is politics, you're just going to lose because you're going to suck at it.
1415
01:17:03.800 --> 01:17:08.599
So I think that's largely the answer is that the one person who was good at politics in EA was SPF.
1416
01:17:09.822 --> 01:17:13.220
And everybody else in EA just like fucking sucks at politics.
1417
01:17:13.320 --> 01:17:25.360
And that's why they were the saintly, quiet, nerdy people who were just in the background doing stuff and donating to bed nets and all this – doing the AI research, AI safety research at a time when nobody else cared.
1418
01:17:25.820 --> 01:17:29.100
And now it's like, oh my god, you guys are actually in the way of progress and you guys are all terrible people.
1419
01:17:29.520 --> 01:17:31.320
And they're like, no, no, we're really not.
1420
01:17:31.943 --> 01:17:34.500
Let's do a two-hour debate to discuss why.
1421
01:17:35.020 --> 01:17:41.380
And they just don't have the ability to get out there and fight with slogans, which is the EAC people, very politically capable.
1422
01:17:41.681 --> 01:17:41.802
Right.
1423
01:17:41.822 --> 01:17:45.600
I mean, Marc Andreessen and Mike Salon and all those guys, they live and breathe politics.
1424
01:17:45.840 --> 01:17:46.041
Right.
1425
01:17:46.342 --> 01:17:50.060
Whereas the EA people, you know, they're a bunch of like philosophy professors.
1426
01:17:50.160 --> 01:17:50.281
Right.
1427
01:17:50.302 --> 01:17:51.860
Like that's that was the leader of their movement.
1428
01:17:52.260 --> 01:17:57.900
These are these are verbose, you know, kind of extremely particular people and they suck at politics.
1429
01:17:58.361 --> 01:18:00.660
They might not be interested in politics, but politics is interesting.
1430
01:18:00.800 --> 01:18:01.822
That's just the point.
1431
01:18:01.882 --> 01:18:11.060
Maybe we could bring to that that part of the conversation, because like these nerd battles that we thought we were having insular to like kind of nerd culture in Silicon Valley are now like.
1432
01:18:11.681 --> 01:18:12.423
political battles.
1433
01:18:12.985 --> 01:18:18.440
And there seem to be some party lines kind of drawing and using these platforms
1434
01:18:19.423 --> 01:18:24.517
for maybe a different use case than when they were originally created.
1435
01:18:25.098 --> 01:18:25.760
And there seems to be
1436
01:18:25.820 --> 01:18:33.220
this dichotomy forming behind kind of a political party for de-accelerationism, like check on
1437
01:18:33.340 --> 01:18:38.514
Silicon Valley power and technological progress without constraints.
1438
01:18:38.554 --> 01:18:41.020
Maybe there's some anti-capitalism
1439
01:18:41.100 --> 01:18:44.300
or anti-billionaire type stuff tied up in that.
1440
01:18:44.740 --> 01:18:48.660
And then there's this pro-tech accelerationism,
1441
01:18:48.881 --> 01:18:49.838
remove the constraints,
1442
01:18:50.401 --> 01:18:52.873
let's let America be the grandmaster of AI,
1443
01:18:55.228 --> 01:18:56.480
let's win this battle.
1444
01:18:57.682 --> 01:19:01.720
And at a very curious time with respect to at least American politics,
1445
01:19:02.061 --> 01:19:05.160
because we're sort of in between orders.
1446
01:19:05.780 --> 01:19:08.260
We had the New Deal order for a very long time,
1447
01:19:08.400 --> 01:19:09.533
FDR until the 1970s,
1448
01:19:10.280 --> 01:19:14.460
And now we're in this neoliberal order and have been for the last 40 years.
1449
01:19:14.920 --> 01:19:19.658
And I feel like we very much are on the exit ramp out of that neoliberal order.
1450
01:19:20.681 --> 01:19:25.100
And it's like the U.S. is trying to find out what's the next political order that we're
1451
01:19:25.120 --> 01:19:26.560
going to establish ourselves with.
1452
01:19:26.640 --> 01:19:31.698
And I think the neoliberal order that kind of rubber stamped the internet in the 1990s
1453
01:19:32.320 --> 01:19:36.420
and helped Silicon Valley prosper, they were very tech forward.
1454
01:19:36.801 --> 01:19:38.300
They love technology, right?
1455
01:19:38.381 --> 01:19:40.680
It's going to bring jobs, like the economy.
1456
01:19:41.460 --> 01:19:47.200
It's not a certain thing that the next order, political order that dominates America, will
1457
01:19:47.260 --> 01:19:51.060
be as tech forward as the order has been at this point in time.
1458
01:19:51.140 --> 01:19:53.980
You could start to see some battle lines being drawn.
1459
01:19:54.360 --> 01:19:58.840
You know, I think people like Balaji and Marc Andreessen might say there's like the blue
1460
01:19:58.900 --> 01:20:06.520
tribe that is tending towards tech de-accelerationism, like lots of restraint, lots of regulatory
1461
01:20:06.761 --> 01:20:10.992
apparatus, like slow down on this whole tech thing in general.
1462
01:20:11.775 --> 01:20:13.439
And then there's maybe Team Red.
1463
01:20:13.459 --> 01:20:13.560
They
1464
01:20:13.760 --> 01:20:19.660
haven't quite chosen an alliance yet, but maybe they are going on the tech forward side of things
1465
01:20:19.740 --> 01:20:21.648
and tech accelerationism side of things.
1466
01:20:22.211 --> 01:20:24.460
And it seems like, I don't know, you have to pick a tribe
1467
01:20:24.520 --> 01:20:25.264
at this point in time.
1468
01:20:25.625 --> 01:20:28.800
From our conversation up to this point, Haseeb and Eric, I feel like both
1469
01:20:28.860 --> 01:20:33.433
of you guys believe in a P-Doom for AI greater than zero, right?
1470
01:20:33.814 --> 01:20:35.760
So you're not maximalists on
1471
01:20:35.880 --> 01:20:41.320
either side, the EAC side, you know, or the or the EA side of this debate, you're probably
1472
01:20:41.443 --> 01:20:42.059
somewhere in the middle.
1473
01:20:42.520 --> 01:20:45.660
But it's hard to really pick a side politically, right?
1474
01:20:45.840 --> 01:20:51.240
When it seems like if you choose the de-accelerationism side, then you're consigned to like
1475
01:20:51.340 --> 01:20:55.740
shitty regulation and they're going to like shut things down and cement some of the leaders.
1476
01:20:55.900 --> 01:20:59.180
Or if you go on kind of the other side, well, then maybe you don't have the constraints
1477
01:20:59.280 --> 01:21:04.740
that you actually need for the space and you're not properly regulating kind of like the
1478
01:21:04.981 --> 01:21:11.740
the danger of this anyway how do you think about the politics of this whole conversation that has
1479
01:21:11.780 --> 01:21:18.680
exploded beyond like nerd culture and is now like biden has executive orders about this stuff all
1480
01:21:18.680 --> 01:21:25.280
right and it's just starting like ai is the thing uh that the nation that the world is is talking
1481
01:21:25.340 --> 01:21:26.002
about at this point.
1482
01:21:26.062 --> 01:21:27.685
So Haseeb, what's your take on this?
1483
01:21:31.393 --> 01:21:34.820
So what I resist the most strongly
1484
01:21:35.903 --> 01:21:39.894
is the politicization of this question, right?
1485
01:21:40.035 --> 01:21:41.940
Because this is one of the most important
1486
01:21:42.020 --> 01:21:46.520
questions of our time, which also means that it's really fucking important to get it right.
1487
01:21:47.221 --> 01:21:51.140
And the one way you will guarantee that you don't get it right is by making it into politics,
1488
01:21:51.662 --> 01:21:56.420
right like if you think about masks like you know how we dealt with the pandemic was probably one
1489
01:21:56.420 --> 01:22:00.140
of the most important questions this decade and we fucking failed because we turned it into politics
1490
01:22:00.582 --> 01:22:05.500
instead of turning it into a just raw scientific question of what is the answer of how this fucking
1491
01:22:05.600 --> 01:22:10.460
disease is spreading and disrupting society and like you know killing millions of people around
1492
01:22:10.480 --> 01:22:14.400
the world um and so if you let that happen to ai i think you've already lost the game right if you
1493
01:22:14.460 --> 01:22:18.060
if you concede this idea that well there's this tribe over here and this tribe over there and you
1494
01:22:18.080 --> 01:22:21.357
pick one side and then, you know, the end, like, no, that's fucking retarded.
1495
01:22:21.598 --> 01:22:21.739
Okay.
1496
01:22:21.759 --> 01:22:21.980
That's just
1497
01:22:22.121 --> 01:22:22.685
absolutely wrong.
1498
01:22:22.906 --> 01:22:24.235
That's not the way you should be thinking about this.
1499
01:22:24.497 --> 01:22:24.920
The way you should be
1500
01:22:24.920 --> 01:22:29.740
thinking about this is that these are really hard questions and the answers require nuance.
1501
01:22:29.901 --> 01:22:30.966
They require careful study.
1502
01:22:31.167 --> 01:22:32.514
They require lots of time and energy.
1503
01:22:32.795 --> 01:22:33.720
And the only way you will
1504
01:22:33.800 --> 01:22:37.477
guarantee that you get it wrong is by saying that there are only two sides, right?
1505
01:22:37.778 --> 01:22:38.180
In reality,
1506
01:22:38.501 --> 01:22:39.766
it's a multidimensional question.
1507
01:22:40.228 --> 01:22:43.120
Like, and the idea that, okay, well, you know, there's the D
1508
01:22:43.280 --> 01:22:46.060
cells and there's the X cells and like, you know, the blue tribe likes this, the red tribe likes
1509
01:22:46.100 --> 01:22:47.085
that and blah, blah, blah.
1510
01:22:47.266 --> 01:22:48.452
It's like, no, no, no, no, that's not true.
1511
01:22:48.774 --> 01:22:49.015
Okay.
1512
01:22:49.297 --> 01:22:49.880
One thing that we know
1513
01:22:49.980 --> 01:22:53.720
about politics from people who actually go and study this stuff, as opposed to reading media,
1514
01:22:54.200 --> 01:22:58.600
which will give you a very distorted view of what people actually think is that most people are not
1515
01:22:58.640 --> 01:23:00.187
diehard leftists or diehard rightists.
1516
01:23:00.609 --> 01:23:01.895
Most people are somewhere in between.
1517
01:23:02.216 --> 01:23:02.980
They have a mishmash of
1518
01:23:03.000 --> 01:23:03.361
ideas.
1519
01:23:03.582 --> 01:23:05.690
They don't fit neatly into either of these categories.
1520
01:23:06.151 --> 01:23:08.600
And they don't like either of the
1521
01:23:08.660 --> 01:23:10.008
leading candidates right now, right?
1522
01:23:10.229 --> 01:23:11.054
That is what we know.
1523
01:23:11.094 --> 01:23:12.220
We know that by actually doing polling.
1524
01:23:12.740 --> 01:23:16.980
If you go look at people, they don't fit neatly into these two categories and they don't on AI risk either.
1525
01:23:17.660 --> 01:23:18.043
Right.
1526
01:23:18.063 --> 01:23:20.460
So people say, well, you know, I think PDOM is this, but I actually think we should do that.
1527
01:23:20.741 --> 01:23:27.480
Or I think, hey, you know, GPU constraints are not smart, but I think we should be putting more effort into, you know, using the government to fund safety research.
1528
01:23:27.540 --> 01:23:27.661
Right.
1529
01:23:27.681 --> 01:23:29.700
Because maybe it's undersupplied by the private market.
1530
01:23:29.940 --> 01:23:34.300
There are a lot of median positions you can take that are not shut it all down or like build, baby, build.
1531
01:23:34.680 --> 01:23:35.226
Right.
1532
01:23:35.327 --> 01:23:36.700
There's tons of room in the middle.
1533
01:23:37.081 --> 01:23:39.560
And if you lose that room, then you're definitely going to fuck up.
1534
01:23:40.062 --> 01:23:43.780
I think you'll fuck up to some degree, but I would argue that it could actually be worse.
1535
01:23:44.701 --> 01:23:50.960
And so I think it would be not the worst situation if one side was build baby boom and the other side was shut it all down.
1536
01:23:51.301 --> 01:23:54.360
Because my worry is that both sides are going to say shut it all down.
1537
01:23:54.740 --> 01:23:59.100
Like we have Tucker Carlson going on Joe Rogan, literally saying shut it all down.
1538
01:23:59.140 --> 01:24:01.100
He sounds like a dumber version of Eliza.
1539
01:24:02.100 --> 01:24:03.537
Not to put down Tucker necessarily.
1540
01:24:05.542 --> 01:24:07.837
And Tucker is one of the most popular people on the right.
1541
01:24:08.320 --> 01:24:12.260
Like if he ran for president, there wasn't Trump, maybe he'd be a candidate.
1542
01:24:12.500 --> 01:24:16.200
So I could see shut it all down being a bipartisan thing.
1543
01:24:17.260 --> 01:24:21.240
Like, you know, we're at a time of extreme populists, as Steve mentioned.
1544
01:24:22.641 --> 01:24:23.944
You know, people don't like Biden.
1545
01:24:23.964 --> 01:24:30.180
People don't like Trump, you know, and and and we're also at a side of an era of horseshoe theory on anti-tech.
1546
01:24:30.421 --> 01:24:30.602
Right.
1547
01:24:30.642 --> 01:24:34.720
Like when you think about someone like Bernie Sanders, you think about someone like Trump.
1548
01:24:34.780 --> 01:24:39.440
These are populists from both sides who actually have a lot in common or more than we might think.
1549
01:24:40.420 --> 01:24:48.120
So my hope is that the AI safety people, I don't want to call them decelerationists because they don't call themselves decelerationists.
1550
01:24:48.180 --> 01:24:52.540
And we should use the terms that people use for each other, for themselves.
1551
01:24:53.421 --> 01:25:00.437
My hope is that the AI safety people and the EAC people can publicly state, hey, we have 95% agreement.
1552
01:25:02.460 --> 01:25:03.540
We're both, we're all pro-tech.
1553
01:25:03.702 --> 01:25:05.640
Tech as an industry is pro-tech.
1554
01:25:05.740 --> 01:25:07.260
We disagree on this little issue.
1555
01:25:07.340 --> 01:25:12.160
Of course, it's of mega importance, of utmost importance of AI safety, because that would
1556
01:25:12.562 --> 01:25:14.637
signal to the rest of the world that there's alignment.
1557
01:25:15.560 --> 01:25:21.980
Whereas my concern is that both sides will be, both red and blue in a populist side will
1558
01:25:22.040 --> 01:25:26.960
be anti-tech and will be sort of missing the forest from the trees when we need to actually
1559
01:25:27.207 --> 01:25:27.700
defend tech.
1560
01:25:27.901 --> 01:25:28.940
I share that concern.
1561
01:25:29.220 --> 01:25:33.140
And actually, that might be kind of like the default state of the pendulum swing, because
1562
01:25:33.341 --> 01:25:38.000
I would say the neoliberal order of the last 40 years has been very much like pro-globalization,
1563
01:25:38.120 --> 01:25:41.111
pro-many things, but definitely pro-tech.
1564
01:25:41.151 --> 01:25:43.660
And I'm seeing whatever the new political order that's
1565
01:25:43.740 --> 01:25:49.355
replacing it to not be so favorable to tech as they previously were.
1566
01:25:49.576 --> 01:25:51.080
And a microcosm for this
1567
01:25:51.421 --> 01:25:53.827
is crypto for me.
1568
01:25:54.309 --> 01:25:58.320
It's basically like it's not a sure thing that Team Red or Team Blue
1569
01:25:58.862 --> 01:25:59.885
will be team crypto.
1570
01:26:00.507 --> 01:26:02.032
It's just totally not a sure thing.
1571
01:26:02.855 --> 01:26:04.400
We have no idea where they're actually
1572
01:26:04.500 --> 01:26:06.185
going to fall on this.
1573
01:26:06.846 --> 01:26:12.060
And it actually, to Eric's point, does make me, as somebody who is an
1574
01:26:12.160 --> 01:26:16.700
advocate for crypto technology, decentralization, private keys in the hands of individual,
1575
01:26:17.321 --> 01:26:24.016
all of these things, start to skew a bit more extreme than I otherwise would be in trying to
1576
01:26:25.660 --> 01:26:32.700
make the case for like why crypto is is good right it's like um i have to do that in order to
1577
01:26:33.422 --> 01:26:38.860
push the agenda and to contrast myself against these movements from both like sides of of
1578
01:26:39.181 --> 01:26:44.440
politics that would just like shut it all down they don't see the use case for um proof of work
1579
01:26:44.540 --> 01:26:48.800
mining and bitcoin so just shut down all the mining facilities they don't understand why
1580
01:26:49.261 --> 01:26:55.280
people need to run a validator in their own home or why um defy interfaces you know don't require
1581
01:26:55.441 --> 01:27:00.240
aml kyc and all of these things right so let's see what do you say to that like eric's making
1582
01:27:00.320 --> 01:27:06.560
the case that maybe we we should all band together not necessarily on the far extreme eac type of um
1583
01:27:06.680 --> 01:27:14.459
case but we should definitely make a pro-tech case um like or else or else we'll fall prey to the the
1584
01:27:14.960 --> 01:27:18.720
other political parties that just don't want any of this technology just don't care about it
1585
01:27:18.720 --> 01:27:25.860
it so the the ea tribe is often affiliated with being doomers about the technology but i feel like
1586
01:27:26.181 --> 01:27:31.300
of the eax side is in a way a doomer on the politics right where they basically believe
1587
01:27:31.380 --> 01:27:35.420
that you know if we don't have this extremely rigorous defense of technology that it's all
1588
01:27:35.581 --> 01:27:39.420
just going to get you know it's all going to go kablooey and we're going to get you know we're
1589
01:27:39.561 --> 01:27:44.800
going to turn into like this uh socialist uh state that we're all going to be uh enslaved by elizabeth
1590
01:27:44.840 --> 01:27:46.247
Warren and nobody's going to have any freedom.
1591
01:27:46.850 --> 01:27:48.820
And like the reality, you just look right now,
1592
01:27:48.900 --> 01:27:52.460
okay, you say, oh, the EX are in control, or sorry, the EAs are in control, the safetyists
1593
01:27:52.982 --> 01:27:53.806
are running the asylum.
1594
01:27:54.529 --> 01:27:57.260
If you look right now around the world, where are the AI restrictions
1595
01:27:57.461 --> 01:27:58.566
actually happening, right?
1596
01:27:58.767 --> 01:27:59.671
The answer is in the EU.
1597
01:28:00.194 --> 01:28:01.600
That's where the most constraining stuff is
1598
01:28:01.600 --> 01:28:01.841
happening.
1599
01:28:02.082 --> 01:28:03.248
And then in China, right?
1600
01:28:03.429 --> 01:28:05.500
People are saying on the EX side, they're like, oh, no,
1601
01:28:05.540 --> 01:28:10.940
if we don't, you know, EA is so anti-tech, you know, if we let the Democrats, we have a democratic
1602
01:28:11.141 --> 01:28:12.287
president right now.
1603
01:28:12.670 --> 01:28:14.560
The Democrats, if they control things, they're going to shut down all
1604
01:28:14.560 --> 01:28:17.900
the technology and they're going to make it so I can't use the internet, I can't use social media,
1605
01:28:17.980 --> 01:28:18.402
I can't use whatever.
1606
01:28:18.724 --> 01:28:21.440
Go look at China where they have, you know, people are worried, oh,
1607
01:28:21.580 --> 01:28:22.547
China's going to race ahead of us.
1608
01:28:22.567 --> 01:28:23.977
They're going to unlock the technology.
1609
01:28:23.997 --> 01:28:24.400
They're the real
1610
01:28:24.440 --> 01:28:25.122
accelerationists.
1611
01:28:25.584 --> 01:28:28.493
Go read the AI regulations in China, right?
1612
01:28:28.794 --> 01:28:30.560
In China, you cannot release an AI
1613
01:28:30.600 --> 01:28:34.033
application without it being explicitly approved by the CCP.
1614
01:28:34.534 --> 01:28:36.060
There are only 40 applications that
1615
01:28:36.080 --> 01:28:37.786
have ever been approved by the CCP, right?
1616
01:28:38.027 --> 01:28:42.100
You are liable in China if your large language model
1617
01:28:42.200 --> 01:28:45.715
says something that's not true, you are liable in China.
1618
01:28:45.735 --> 01:28:46.980
That's the law that they passed in China.
1619
01:28:47.180 --> 01:28:47.361
Okay.
1620
01:28:47.642 --> 01:28:48.525
That's fucking insane.
1621
01:28:48.765 --> 01:28:50.471
That's absolute.
1622
01:28:51.053 --> 01:28:53.100
That is what you're up against in America.
1623
01:28:53.401 --> 01:28:55.672
The reality is that yes, there's a cacophony of voices.
1624
01:28:56.254 --> 01:28:57.540
Many people saying many different things.
1625
01:28:57.660 --> 01:28:58.544
The EACs are saying one thing.
1626
01:28:58.725 --> 01:28:59.630
The EAs are saying one thing.
1627
01:28:59.871 --> 01:29:01.580
The AI, you know, anti-bias people are
1628
01:29:01.620 --> 01:29:02.343
saying some other things.
1629
01:29:02.845 --> 01:29:04.734
And the reality, what did we actually have in the US?
1630
01:29:05.598 --> 01:29:06.140
The answer is
1631
01:29:06.240 --> 01:29:06.481
nothing.
1632
01:29:07.345 --> 01:29:08.048
There is no laws.
1633
01:29:08.390 --> 01:29:09.234
There's no regulation.
1634
01:29:09.455 --> 01:29:10.459
There's nothing right now, right?
1635
01:29:10.479 --> 01:29:10.700
There's an
1636
01:29:10.700 --> 01:29:15.180
executive order that says, you have to let us know if you run a training run at least this size,
1637
01:29:15.642 --> 01:29:16.044
that's it.
1638
01:29:16.546 --> 01:29:18.174
That's the only regulation that we actually have.
1639
01:29:18.395 --> 01:29:19.340
And regulation is a strong word
1640
01:29:19.400 --> 01:29:20.004
for that, right?
1641
01:29:20.266 --> 01:29:22.560
The reality is that there's no regulatory regime at the moment in the US.
1642
01:29:22.881 --> 01:29:27.560
And the reason why that is, is because that cacophony of voices continues pulling things
1643
01:29:27.600 --> 01:29:28.606
in different directions, right?
1644
01:29:28.847 --> 01:29:31.080
I mean, we say, man, it really sucks that Gensler is doing this.
1645
01:29:31.240 --> 01:29:33.914
And, you know, the FinCEN is doing this and blah, blah, blah.
1646
01:29:33.934 --> 01:29:34.095
Okay.
1647
01:29:34.135 --> 01:29:35.140
What are the laws of crypto in
1648
01:29:35.140 --> 01:29:35.441
the US?
1649
01:29:36.546 --> 01:29:37.730
The answer is that there are none.
1650
01:29:38.212 --> 01:29:39.176
There are no laws in the US.
1651
01:29:39.437 --> 01:29:40.080
There's no regulatory
1652
01:29:40.200 --> 01:29:41.244
regime, there's nothing.
1653
01:29:41.505 --> 01:29:42.690
There's just a bunch of people fighting.
1654
01:29:43.352 --> 01:29:45.220
And, you know, if you look at the
1655
01:29:45.561 --> 01:29:49.880
founding fathers, the way that they described government is that they wanted that to be the
1656
01:29:49.900 --> 01:29:50.121
case.
1657
01:29:50.904 --> 01:29:54.600
They wanted there to be continual fighting between the different branches of government to
1658
01:29:54.640 --> 01:29:57.653
prevent any centralized control in, for example, the president.
1659
01:29:58.276 --> 01:29:59.140
They didn't want the president to
1660
01:29:59.140 --> 01:29:59.944
be able to say, well, guess what?
1661
01:30:00.004 --> 01:30:00.466
I'm president.
1662
01:30:00.647 --> 01:30:01.611
So here are the laws now.
1663
01:30:02.174 --> 01:30:03.480
Well, the laws can get
1664
01:30:03.520 --> 01:30:04.644
shut down by judges.
1665
01:30:04.905 --> 01:30:07.033
They can get replaced by legislators.
1666
01:30:07.415 --> 01:30:08.800
And the reality is that that
1667
01:30:08.860 --> 01:30:10.768
cacophony of voices keeps things in check.
1668
01:30:11.110 --> 01:30:13.420
So anybody who's claiming, oh my God, there are EAs
1669
01:30:13.460 --> 01:30:15.252
in the government now, therefore it's all going to go to hell.
1670
01:30:15.312 --> 01:30:16.460
And, you know, we're now all going
1671
01:30:16.460 --> 01:30:20.460
to be, you know, talking on our Google controlled smartphones and nobody's going to have open source
1672
01:30:20.500 --> 01:30:21.043
AI anymore.
1673
01:30:21.063 --> 01:30:21.827
It's all going to get banned.
1674
01:30:22.188 --> 01:30:23.173
They're fucking wrong.
1675
01:30:23.334 --> 01:30:24.158
It's never happened.
1676
01:30:24.238 --> 01:30:24.520
It's never
1677
01:30:24.620 --> 01:30:26.529
been true because of our system of government.
1678
01:30:26.992 --> 01:30:28.760
So that's why that like, there's a political
1679
01:30:28.860 --> 01:30:34.560
doomerism that I also object to, which I feel like is part of what motivates the, the, um,
1680
01:30:34.981 --> 01:30:40.840
the energy in demonizing the the ea type so i think what what they have to say is a very good
1681
01:30:40.920 --> 01:30:45.920
point that deserves to be heard but do i want ea is running the government no i want the government
1682
01:30:46.000 --> 01:30:51.500
to be the way it is which is you know a a um a contest of different forces right there are people
1683
01:30:51.600 --> 01:30:55.820
who are economically minded who have a good point that's not being pro-culture war it's being pro
1684
01:30:55.900 --> 01:31:01.640
democracy yes i'm just teasing but you're pro this productive tension and it plays out yes yes i'm
1685
01:31:01.700 --> 01:31:02.972
Very pro-productive tension, 100%.
1686
01:31:05.249 --> 01:31:07.260
I actually really like this productive tension framing.
1687
01:31:07.960 --> 01:31:14.580
Kind of like as we get to the close of this conversation here, I kind of want to zoom out because I actually see the same acceleration.
1688
01:31:14.721 --> 01:31:18.880
I know decelerationist isn't actually a tribe here, but it's like kind of a talking point.
1689
01:31:19.801 --> 01:31:22.180
There's accelerationist versus decelerationist debate.
1690
01:31:22.320 --> 01:31:24.440
But it's also not just like Silicon Valley.
1691
01:31:24.740 --> 01:31:25.516
It's not just AI.
1692
01:31:25.880 --> 01:31:26.600
It's not just crypto.
1693
01:31:27.322 --> 01:31:30.000
Whereas these things are generally all on the accelerationist front.
1694
01:31:30.080 --> 01:31:33.540
Like we're also seeing experiments going on in other technology verticals as well.
1695
01:31:34.120 --> 01:31:36.600
Like Brian Armstrong or excuse me, not Brian Armstrong.
1696
01:31:36.760 --> 01:31:40.840
Brian Johnson is trying to experiment himself into like becoming a new class of humans.
1697
01:31:41.361 --> 01:31:48.620
This is the guy who's taking any risk or not any risk, but any sort of like experiment to try and like add, you know, five more minutes to his life every single day.
1698
01:31:48.700 --> 01:31:48.861
Right.
1699
01:31:49.203 --> 01:31:52.980
Like people like Aubrey de Grey are literally trying to create like immortality.
1700
01:31:53.721 --> 01:31:59.900
Even in the crypto space, there's like Vida Dow out there who's like trying to combine crypto and immortality to make it happen faster.
1701
01:32:00.443 --> 01:32:03.320
and like this is we haven't even like opened up the conversation about like gene editing
1702
01:32:03.781 --> 01:32:09.320
which is like not far away like there's other technologies out there that's really going to
1703
01:32:09.380 --> 01:32:12.740
like if you think ai and crypto is going to make the future weird well like there's like three more
1704
01:32:12.780 --> 01:32:17.420
technologies out there who's also going to make uh the future weird and the thing is like this
1705
01:32:17.460 --> 01:32:21.200
has always been the case of like technology technology just accelerates it's always
1706
01:32:21.320 --> 01:32:25.460
accelerated as soon as we develop tools we can use those tools to develop other tools and all
1707
01:32:25.480 --> 01:32:29.980
of a sudden like we are an accelerationist in an accelerating state and this is just the only
1708
01:32:30.100 --> 01:32:34.760
difference between like now and any other time in history is that it seems to be where we are now
1709
01:32:34.961 --> 01:32:39.520
versus where we are at the end of the lifetimes might be like multiple orders of magnitude
1710
01:32:39.942 --> 01:32:44.760
different and that has never been true before in history like we've only ever had like 20 percent
1711
01:32:45.202 --> 01:32:49.840
30 percent gains at most like usually over the span of human history we have like two percent
1712
01:32:49.940 --> 01:32:53.900
gains and like for thousands of years we actually literally had zero percent gains and we're now
1713
01:32:53.900 --> 01:32:59.400
we're all on this like parabola of innovation and so if we're talking about this tension between
1714
01:33:00.523 --> 01:33:02.489
governments, like stakeholders in the government.
1715
01:33:02.610 --> 01:33:05.960
It does make sense that when you have crypto and
1716
01:33:06.121 --> 01:33:11.580
AI and like biotech and all of these other different like sectors of technology, which
1717
01:33:11.600 --> 01:33:16.240
are all like making monumental gains, that all of government is like, okay, everyone break.
1718
01:33:16.701 --> 01:33:18.026
Because who else is saying that?
1719
01:33:18.368 --> 01:33:20.337
Who else is saying like, yo, calm down.
1720
01:33:20.638 --> 01:33:21.080
Like that's maybe
1721
01:33:21.140 --> 01:33:26.480
like kind of like my commentary on like why I think the government will trend towards breaks,
1722
01:33:26.601 --> 01:33:29.700
not gas, but like maybe Eric, maybe you can like check my reasoning on this.
1723
01:33:30.640 --> 01:33:32.968
No, I, um, I, I, I, I agree.
1724
01:33:33.068 --> 01:33:35.658
That's part of, uh, the, the broader productive tension.
1725
01:33:35.798 --> 01:33:36.300
And I, um,
1726
01:33:36.801 --> 01:33:41.100
I want to bring in, you guys had this amazing episode with Vitalik where he, he, uh, identified
1727
01:33:41.341 --> 01:33:43.467
and expanded upon his Diak philosophy.
1728
01:33:44.009 --> 01:33:46.156
Um, and I think that's worth noting too.
1729
01:33:46.537 --> 01:33:47.500
And, and I think
1730
01:33:47.841 --> 01:33:53.240
that piece had a great, um, synthesis of the, of the, of the problem and sort of the common
1731
01:33:53.280 --> 01:33:59.040
alignment between EA and EAC and trying to develop this sort of thesis and synthesis approach.
1732
01:33:59.860 --> 01:34:02.569
But I think, unfortunately, I think it's a bit too neat.
1733
01:34:03.071 --> 01:34:05.800
I think it's tough to just focus on
1734
01:34:06.201 --> 01:34:07.687
technology that are defensive.
1735
01:34:08.691 --> 01:34:11.120
Things that tend to be defensive can be offensive and vice versa.
1736
01:34:11.762 --> 01:34:13.674
Fire can be used for good, it can be used for evil.
1737
01:34:13.694 --> 01:34:14.519
And if we're going to continue to
1738
01:34:15.102 --> 01:34:19.720
develop these models, I don't see ways in which it can only be used for defensive ways.
1739
01:34:20.122 --> 01:34:23.740
So that's my commentary on that.
1740
01:34:23.960 --> 01:34:34.980
And then to connect to your point, I think we are going to continue to need the AI safety people to develop and further develop our alignment capabilities.
1741
01:34:35.843 --> 01:34:37.880
And we're going to need people who are pushing the brakes.
1742
01:34:38.260 --> 01:34:48.520
It's my goal in this conversation to rehab or improve EA's reputation, AI safety's reputation, but to also keep it in its box.
1743
01:34:48.720 --> 01:34:50.482
As Haseeb said, we don't want them running the government.
1744
01:34:50.843 --> 01:34:52.525
They're not close to it yet.
1745
01:34:52.645 --> 01:34:57.271
But, you know, the regulatory regime, you know, is going to be up for grabs over the next few years.
1746
01:34:57.291 --> 01:34:59.714
And that's why these fights are really important.
1747
01:34:59.774 --> 01:35:04.780
Even if they only influence a small amount of people right now, they have potential to influence a large amount of people.
1748
01:35:04.900 --> 01:35:09.840
And so I think we do need the right combination of brakes versus gas.
1749
01:35:10.000 --> 01:35:12.720
I would rather see it come from within the industry, from people who know what they're
1750
01:35:12.740 --> 01:35:16.740
talking about, than outside of it prematurely.
1751
01:35:18.060 --> 01:35:22.180
And so, yes, we are going to see governments come in with some brakes.
1752
01:35:22.360 --> 01:35:25.960
Ideally, they can be informed by some of our people and some of the productive tension
1753
01:35:26.041 --> 01:35:26.932
that we're having first.
1754
01:35:29.242 --> 01:35:33.120
The one thing I would add to that is that we've talked a lot about, oh, my God, what
1755
01:35:33.140 --> 01:35:36.056
if the EAs are running things and what if they have total control?
1756
01:35:36.076 --> 01:35:36.880
Won't they shut it all down?
1757
01:35:38.060 --> 01:35:43.840
The reality is a much more politically powerful cohort is not the AI safety people, but the AI
1758
01:35:43.960 --> 01:35:45.447
bias people, right?
1759
01:35:45.588 --> 01:35:46.994
These people are much more influential.
1760
01:35:47.215 --> 01:35:48.380
Are these the same people that Eric
1761
01:35:48.420 --> 01:35:51.150
was calling AI ethics earlier in the conversation?
1762
01:35:51.190 --> 01:35:53.037
Yes, AI ethics, exactly.
1763
01:35:53.257 --> 01:35:54.020
So these are people like,
1764
01:35:54.120 --> 01:35:55.669
okay, you know, what if the models are biased?
1765
01:35:55.689 --> 01:35:57.379
What if the models are more liberal?
1766
01:35:57.399 --> 01:35:57.580
What if
1767
01:35:57.600 --> 01:35:58.866
they're more Republican?
1768
01:35:58.906 --> 01:36:01.599
What if they, you know, say things that they shouldn't be saying or,
1769
01:36:02.542 --> 01:36:05.110
they give people instructions on how to cause harm or whatever.
1770
01:36:06.173 --> 01:36:08.380
Now, the interesting thing about AI
1771
01:36:08.561 --> 01:36:13.840
ethics versus AI safety is that the core problems that you need to solve are actually kind of the
1772
01:36:13.900 --> 01:36:16.430
same, which is how do you control these models?
1773
01:36:17.072 --> 01:36:18.558
And the answer is we don't really know.
1774
01:36:18.859 --> 01:36:19.120
We don't
1775
01:36:19.180 --> 01:36:20.730
have good techniques to control these models.
1776
01:36:21.012 --> 01:36:21.918
Now, how do you control a model?
1777
01:36:21.938 --> 01:36:22.180
Well, that's
1778
01:36:22.220 --> 01:36:26.280
called AI alignment, which is the core problem behind AI safety, which is, okay, right now,
1779
01:36:26.561 --> 01:36:31.440
the thing that the model is doing that you don't want is making biased representations of, you
1780
01:36:31.500 --> 01:36:33.853
know, people in certain, you know, stereotypical careers or something.
1781
01:36:34.215 --> 01:36:35.140
But, you know, eventually
1782
01:36:35.200 --> 01:36:38.860
it might be the thing you don't want is that, you know, gathering a bunch of resources or hacking
1783
01:36:38.960 --> 01:36:43.400
people in order to, you know, make a bunch of money in order to fulfill some task that somebody
1784
01:36:43.400 --> 01:36:45.268
who's running an AI agent wants it to do.
1785
01:36:46.131 --> 01:36:48.320
These problems are one in the same, in a sense, right?
1786
01:36:48.380 --> 01:36:51.760
I mean, they're, one of them is easier to solve than the other, but the core of the problem is
1787
01:36:51.780 --> 01:36:53.789
that we don't know how to control these models.
1788
01:36:54.311 --> 01:36:55.597
They are very difficult to control.
1789
01:36:55.818 --> 01:36:56.200
They're very
1790
01:36:56.240 --> 01:36:56.743
unintuitive.
1791
01:36:56.964 --> 01:36:59.979
They're kind of like these space aliens that we've conjured up, right?
1792
01:36:59.999 --> 01:37:00.160
People,
1793
01:37:00.401 --> 01:37:07.060
People often say that large language models were discovered, not invented in the sense that, you know, we just sort of threw more and more compute.
1794
01:37:07.180 --> 01:37:11.740
And then it's kind of like, you know, when something moves faster and faster, it turns from a liquid into a gas.
1795
01:37:12.401 --> 01:37:15.640
In the same way, when you throw more and more compute at something, at a certain point, it becomes a large language model.
1796
01:37:16.081 --> 01:37:17.320
And you're just like, whoa, what the fuck?
1797
01:37:17.400 --> 01:37:17.640
What is it?
1798
01:37:17.721 --> 01:37:20.020
This thing can now like talk to me and do my math homework.
1799
01:37:20.100 --> 01:37:20.799
Like, that's really weird.
1800
01:37:21.861 --> 01:37:25.280
And that phase transition is still very poorly understood.
1801
01:37:26.062 --> 01:37:28.340
And I think there's going to be a lot of political fights about it.
1802
01:37:28.641 --> 01:37:38.140
But if both the EACs and the EAs don't get their shit together, the AI ethics people are going to win because they have much – one, they're much more persuasive.
1803
01:37:38.940 --> 01:37:41.620
They're much more easy to understand for most people.
1804
01:37:42.040 --> 01:37:45.300
Like the whole Gemini drama had nothing to do with EAC or EA.
1805
01:37:45.702 --> 01:37:47.160
It was all bias, right?
1806
01:37:47.160 --> 01:37:58.257
It was all like, oh, you know, this thing fits neatly into the actual culture war, which is the left-right culture war, which is, oh, these things are super woke and they won't depict any white people.
1807
01:38:00.941 --> 01:38:05.580
And that is much more powerfully animating in – among Congress.
1808
01:38:06.020 --> 01:38:10.300
We're talking about what are 70-year-olds who are actually the people making the laws and voting on this stuff.
1809
01:38:10.784 --> 01:38:11.640
What resonates with them?
1810
01:38:11.906 --> 01:38:12.320
I'll tell you.
1811
01:38:12.380 --> 01:38:13.260
It's the AI ethics stuff.
1812
01:38:13.381 --> 01:38:14.540
None of this stuff even rates yet.
1813
01:38:14.983 --> 01:38:17.420
So everything that Mark Andreessen doing
1814
01:38:17.561 --> 01:38:19.194
and that Eliezer is doing is important
1815
01:38:19.980 --> 01:38:21.975
because by default, both tribes are losing.
1816
01:38:22.902 --> 01:38:24.739
And maybe this is a great place to end this idea.
1817
01:38:25.746 --> 01:38:27.818
There's this saying, I'm against my brothers,
1818
01:38:28.120 --> 01:38:29.796
but me and my brother are against my cousin.
1819
01:38:30.341 --> 01:38:32.600
And right now EAC is against EA,
1820
01:38:32.660 --> 01:38:36.398
but EA and EAC need to team up against AI ethics or AI bias
1821
01:38:36.840 --> 01:38:39.555
because we can all agree that those people are wrong.
1822
01:38:40.520 --> 01:38:41.759
Well, hold on, hold on.
1823
01:38:41.860 --> 01:38:42.557
Last thing I wanna say,
1824
01:38:43.563 --> 01:38:49.680
I want to make some room for the AI biased people because like obviously it's true that Gemini was like a horrible model.
1825
01:38:50.100 --> 01:38:54.080
And obviously it's true that the AI models can do pedestrian bad things too.
1826
01:38:54.520 --> 01:39:01.560
And you want to control that because way before AIs are going to become super intelligent or become so powerful they take over the world, they're going to do pedestrian harms.
1827
01:39:02.120 --> 01:39:03.900
And you kind of want both, right?
1828
01:39:03.920 --> 01:39:07.180
You want to also solve the pedestrian harms, and that's a real problem.
1829
01:39:07.964 --> 01:39:09.360
You also want to solve the existential harms.
1830
01:39:09.742 --> 01:39:11.500
That's a potential real problem too.
1831
01:39:11.582 --> 01:39:12.480
It's a much bigger problem.
1832
01:39:13.304 --> 01:39:15.760
And you also want to solve AI overregulation.
1833
01:39:16.120 --> 01:39:17.280
That's also a big problem, right?
1834
01:39:17.380 --> 01:39:22.080
All three of these are problems that a good regulatory regime should find a way to solve.
1835
01:39:23.740 --> 01:39:27.320
And so, I mean, that is my very lame and narrow attempt at synthesis of these three positions.
1836
01:39:27.560 --> 01:39:31.220
But I don't think you have to say, well, you know, these two are illegitimate and this one is legitimate.
1837
01:39:31.962 --> 01:39:34.520
The answer is obviously all three of them have legitimate concerns.
1838
01:39:34.963 --> 01:39:36.280
And you want to find a way to address all three.
1839
01:39:36.540 --> 01:39:37.640
That's too much nuance for me.
1840
01:39:37.720 --> 01:39:38.531
I preferred my…
1841
01:39:39.321 --> 01:39:40.657
Fair enough.
1842
01:39:40.940 --> 01:39:42.420
Let's close this conversation out.
1843
01:39:42.440 --> 01:39:48.700
And I want to ask you just one final ending question, which is just maybe summarize the conversation for each of you.
1844
01:39:48.880 --> 01:39:56.520
So when faced with the question, which you surely will, Bankless listener and then Haseeb and Eric, are you team Excel or team Decel?
1845
01:39:56.981 --> 01:39:59.496
Are you team EAC or are you team EA?
1846
01:39:59.596 --> 01:40:00.200
Pick a side.
1847
01:40:01.501 --> 01:40:06.820
What's the most sensible posture to that kind of framing of the question to pick a side?
1848
01:40:08.184 --> 01:40:08.539
Haseeb.
1849
01:40:10.485 --> 01:40:15.680
so i mean i won't surprise anyone to say that i'm on team ea um but a frame that is saying i'm on
1850
01:40:15.760 --> 01:40:20.840
team ea because i think uh at the margin the world needs more of that i think it's got enough people
1851
01:40:20.920 --> 01:40:25.520
who are say you know pro tech uh no regulation i want more ai products and i think there's also
1852
01:40:25.600 --> 01:40:29.580
enough people who are saying hey there's too much bias in these things and they're they're bad for
1853
01:40:29.620 --> 01:40:33.600
society i think there's not enough people saying hey these things are really risky and we should
1854
01:40:33.620 --> 01:40:34.563
invest more into safety?
1855
01:40:36.288 --> 01:40:38.675
I also want to surprise people.
1856
01:40:38.695 --> 01:40:39.878
I'm team accelerationist.
1857
01:40:40.019 --> 01:40:40.400
I'm team
1858
01:40:40.480 --> 01:40:44.572
pro-tech, but I think that we should absorb some parts of EA.
1859
01:40:44.773 --> 01:40:47.260
I think EA is beaten down.
1860
01:40:47.921 --> 01:40:53.356
They have wounds and we should pay for their hospital bills and bring them into our tribe
1861
01:40:55.262 --> 01:40:58.230
and team up against the AI bias or AI ethics.
1862
01:40:59.213 --> 01:41:01.580
Some who have some legitimate points, I will concede
1863
01:41:01.640 --> 01:41:05.773
that, but also some points that I also very much disagree with.
1864
01:41:05.833 --> 01:41:08.000
And I think my call out to
1865
01:41:08.000 --> 01:41:12.412
accelerationists is to take AI safety seriously.
1866
01:41:12.573 --> 01:41:15.140
I think they worry that if you concede that
1867
01:41:16.023 --> 01:41:21.080
PDoom is even a concept or that there are AI safety concerns, that you sort of throw out the
1868
01:41:21.100 --> 01:41:25.097
baby with the bathwater and you go as extreme as some of the most extreme EA people.
1869
01:41:25.419 --> 01:41:25.740
But I think
1870
01:41:25.760 --> 01:41:29.700
we should find a way to have more common ground with AI safety people, acknowledge that there's
1871
01:41:30.121 --> 01:41:36.780
95% agreement with most of them and work on alignment and then fight the right enemies,
1872
01:41:37.240 --> 01:41:41.640
which as Hasib mentioned, it's the AI ethics people, AI advice people, regardless of what
1873
01:41:41.640 --> 01:41:46.955
you think of them, they have a lot more power on a regulatory stage.
1874
01:41:47.216 --> 01:41:48.600
And that's where the battle
1875
01:41:48.981 --> 01:41:49.865
is really taking shape.
1876
01:41:49.925 --> 01:41:53.860
So I'm an accelerationist with some humanism, and I think we should extend
1877
01:41:53.900 --> 01:41:58.160
an olive branch to our most rational and reasonable AI safety people.
1878
01:41:58.160 --> 01:41:58.482
There you go.
1879
01:41:58.502 --> 01:42:00.191
Nothing unites us like a common enemy.
1880
01:42:00.231 --> 01:42:01.980
And so maybe that common enemy is the
1881
01:42:02.060 --> 01:42:03.185
EI bias people.
1882
01:42:03.245 --> 01:42:05.194
The people are just trying to censor all these models.
1883
01:42:05.556 --> 01:42:06.540
Hasib, Eric, thank you
1884
01:42:06.600 --> 01:42:08.026
for this fantastic conversation.
1885
01:42:08.428 --> 01:42:11.280
I think this will leave bankless listeners with a lot to think about
1886
01:42:11.361 --> 01:42:12.227
on this subject.
1887
01:42:12.248 --> 01:42:12.893
It's been a pleasure.
1888
01:42:13.114 --> 01:42:13.840
Thank you for having me.
1889
01:42:13.940 --> 01:42:14.986
Got to leave you guys with this.
1890
01:42:15.046 --> 01:42:17.540
Of course, I usually say crypto is risky, but we talked
1891
01:42:17.600 --> 01:42:18.243
mostly about AI.
1892
01:42:18.303 --> 01:42:19.911
I think AI is risky as well.
1893
01:42:19.931 --> 01:42:21.176
This is definitely the frontier.
1894
01:42:21.196 --> 01:42:21.960
It's not for
1895
01:42:22.000 --> 01:42:24.256
everyone, but we're glad you're with us on the bankless journey.
1896
01:42:24.276 --> 01:42:24.860
Thanks a lot.