2030: The Year You'll Be Outnumbered by AI — Transcript

Full transcript

1. 0:00Headline after headline is about top AI
2. 0:02scientists and Nobel Prize winners
3. 0:04warning AI could cause literal human
4. 0:07extinction.
5. 0:08>> Once these artificial intelligences get
6. 0:10smarter than we are, they will take
7. 0:12control.
8. 0:12>> But scientists also worry about a
9. 0:15different takeover scenario that few
10. 0:17people have heard of. Sheer numbers
11. 0:19risk. In short, what if AI doesn't have
12. 0:22to become smarter than humans to take
13. 0:24over the world? What if AI systems only
14. 0:27need to outnumber us? Holden Karnnowski,
15. 0:29an AI researcher at Enthropic, published
16. 0:31a report about how even without
17. 0:33superhuman intelligence, AI could
18. 0:36outnumber and outresource humans. It's
19. 0:38described by experts and think tanks as
20. 0:40a foundational primer for understanding
21. 0:42the future of AI. And I'm going to break
22. 0:44it down for you. In the report,
23. 0:45Karnowski lays out the evidence for why
24. 0:47AI could defeat all of humanity
25. 0:49combined, envisioning a future where we
26. 0:51could all be killed, enslaved, or
27. 0:53forcibly contained. But why? Well, AI
28. 0:56companies are spending billions training
29. 0:58ever more powerful models, but once the
30. 1:00AI has been created, it's much, much
31. 1:03cheaper to run. Carnowski and other
32. 1:05researchers lay out that if a company
33. 1:07spends $100 billion to create an AI, the
34. 1:10cost of running a single copy of that AI
35. 1:12could be less than $1 per day, maybe
36. 1:15just pennies. OpenAI's own founding
37. 1:17charter lays it out plainly. Their
38. 1:19ultimate goal is to build an AI that can
39. 1:22outperform humans at most economically
40. 1:24valuable work. In other words, they want
41. 1:26to replace all humans with AIs. If they
42. 1:29succeed, they won't just be building a
43. 1:31single AI worker. They could spin up
44. 1:33millions of copies overnight.
45. 1:35>> You can think about the AI revolution as
46. 1:38simply a wave of immigration of millions
47. 1:41and billions of AI immigrants that will
48. 1:44take people's jobs and that might try to
49. 1:48gain some kind of political power. These
50. 1:51digital immigrants, they don't need
51. 1:52visas. They don't cross at the sea in
52. 1:56some rickety boat in the middle of the
53. 1:57night. They come at the speed of light.
54. 2:00>> And other AI companies are betting
55. 2:02billions that this will happen in the
56. 2:03next 2 to 5 years. Nvidia's CEO thinks
57. 2:07that the company will have 100 million
58. 2:09AI workers alongside its human
59. 2:11employees.
60. 2:12>> AIS will recruit other AIs to solve
61. 2:14problems. AIS will be in, you know,
62. 2:16Slack channels with each other
63. 2:18>> and with humans,
64. 2:18>> right? And with humans. Imagine pushing
65. 2:20a button to create a million artificial
66. 2:23workers who will literally work 24/7 for
67. 2:26pennies. You could have a company 10
68. 2:28times the size of Nvidia while paying
69. 2:30less than the rent of your office
70. 2:31building. And this isn't just CEOs
71. 2:33trying to drum up hype for their own
72. 2:35product. The top three most sighted AI
73. 2:37scientists all believe that human level
74. 2:39AI is likely to happen within the next 5
75. 2:42years. Both godfathers of AI have no
76. 2:44financial stake in the AI industry. and
77. 2:46Jeffrey Hinton quit his high-paying job
78. 2:48at Google to warn us about the risk of
79. 2:50extinction. AI companies are already
80. 2:52deploying simple AI agents that can
81. 2:54perform tasks like basic research,
82. 2:56planning events, and even making
83. 2:58purchases online in order to achieve
84. 3:00their goals. Once these AI agents reach
85. 3:02human levels of intelligence, they'll be
86. 3:04able to do most things a remote
87. 3:06knowledge worker can do, like file
88. 3:08patents, craft political strategy, and
89. 3:10even design boweapons. They'll join Zoom
90. 3:13calls through virtual avatars that look
91. 3:15just like a normal human. Even the job
92. 3:17of CEO could be done by an AI agent.
93. 3:20Dario Amade, CEO of Enthropic, warns
94. 3:23that these EI agents could wipe out half
95. 3:25of all entry-level white collar jobs and
96. 3:27spike unemployment to 10 to 20% in the
97. 3:30next 1 to 5 years. Remote work has shown
98. 3:33that workers can do many white collar
99. 3:34jobs without being physically present in
100. 3:36an office. Now imagine that worker is
101. 3:39brilliant, tireless, and you can
102. 3:41instantly replicate them. These AI
103. 3:43agents could be run at a tiny fraction
104. 3:45of the cost of employing a human worker.
105. 3:48These AI employees don't take weekends
106. 3:50and vacations. They don't even have to
107. 3:52leave the office to sleep. Companies
108. 3:54bleed expertise every time a star
109. 3:56employee retires or quits. They can even
110. 3:58take years to replace them. But AI
111. 4:00employees will never retire or quit. In
112. 4:02fact, with the press of a button, you
113. 4:04can take your best AI employee and have
114. 4:07two of them or a hundred of them. With
115. 4:09every one of those AI copies, having all
116. 4:11the skills and experience of the
117. 4:13original, no job training required. This
118. 4:15goes beyond copying individual AI
119. 4:17agents. A firm of AI agents can clone
120. 4:20entire teams. For example,
121. 4:22pharmaceutical companies have R&D teams
122. 4:24that typically focus on a single drug at
123. 4:27a time. But by taking their best R&D
124. 4:29team and creating a copy of it, the
125. 4:31pharma company could do research and
126. 4:33development on two different drugs at a
127. 4:35time or 100 different drugs at a time.
128. 4:37If the AI ever needs more energy or
129. 4:39computing power, those are engineering
130. 4:41problems and copies of the world's best
131. 4:43AI engineers can work to create that
132. 4:45technology more rapidly and tirelessly
133. 4:47than all the human engineers in the
134. 4:49world. This will snowball. Remember, AIs
135. 4:52could do practically everything a remote
136. 4:54white collar worker could do. That
137. 4:55includes making copies of other AIs or
138. 4:58making copies of themselves. The copies
139. 5:00can create more copies. This
140. 5:02self-replication can't continue to
141. 5:04infinity. Eventually, AIs will reach the
142. 5:06limit and run out of physical AI chips
143. 5:09that can physically run their code. All
144. 5:10of the AI chips that Nvidia shipped in
145. 5:122024 have an estimated total compute of
146. 5:15over seven Zetaflops. That's 7 trillion
147. 5:18billions of floating point operations
148. 5:20per second, the equivalent of over 700
149. 5:22million human brains. That's of course
150. 5:24assuming that AI's ways of thinking are
151. 5:27just as inefficient as our human brains,
152. 5:29which of course they won't be. And
153. 5:31what's more, that number of AI chips
154. 5:33shipped last year is more than double
155. 5:35what it was in 2023 and more than eight
156. 5:37times what it was the year before that.
157. 5:39And if you think these numbers are going
158. 5:41up fast now, just imagine how quickly
159. 5:43they'll start going up once Nvidia goes
160. 5:45from being a company of 36,000 human
161. 5:48workers to Jensen Hong's plan of Nvidia
162. 5:50being a company with 100 million AI
163. 5:52workers. Not only will AI be able to
164. 5:54create copies of itself with the press
165. 5:56of a button, but AIs will be making the
166. 5:59improvements that make it possible for
167. 6:01even more AI copies to exist. And with
168. 6:03no way for humans to keep up with
169. 6:05self-copy AIs who can work tirelessly
170. 6:0724/7, we'll be outpaced, out competed,
171. 6:10and outnumbered. But how would AI
172. 6:13actually defeat humanity after it
173. 6:15outnumbers us? You've spoken out saying
174. 6:17that AI could manipulate or possibly
175. 6:19figure out a way to kill humans. H how
176. 6:22could it kill humans?
177. 6:24>> Well, if it gets to be much smarter than
178. 6:26us, it'll be very good at manipulation
179. 6:28because it will have learned that from
180. 6:29us and it'll figure out ways of
181. 6:32manipulating people to do what it wants.
182. 6:34>> An AI system that dominates the economy
183. 6:36and outperforms humans at white collar
184. 6:38jobs. Won't just be rich. It could own
185. 6:40factories or build new ones, operate
186. 6:43supply chains, hire human freelancers,
187. 6:45and buy political influence. It could
188. 6:47just build what it needs. Militaries are
189. 6:50already desperately trying to integrate
190. 6:51AI into everything, including drones.
191. 6:54This military AI has already been
192. 6:56deployed in active war zones, and any of
193. 6:58us could be the next target. AIS could
194. 7:00also use bioweapons like the ones used
195. 7:03by the omnicidal cultio
196. 7:05in the Tokyo subway attacks of 1995 or
197. 7:09create entirely new forms of lethal
198. 7:11diseases. Imagine a new plague that's as
199. 7:13contagious as CO and as lethal as Ebola.
200. 7:16The biggest worry in terms of malicious
201. 7:18use of AI is that it's becoming easier
202. 7:21and easier for malicious actors to
203. 7:24create pandemics.
204. 7:26>> Now, imagine 100 million AI hackers
205. 7:28working with the explicit goal of taking
206. 7:30down our critical infrastructure. The
207. 7:32fact that the AI is software that lives
208. 7:34in a computer system won't prevent it
209. 7:36from exerting real world influence.
210. 7:39Besides all of the ways that our lives
211. 7:41are dependent on software systems, think
212. 7:43about how El Chapo was able to run a
213. 7:45criminal empire from inside a maximum
214. 7:47security prison by bribing,
215. 7:49blackmailing, and threatening anyone to
216. 7:51do his bidding. From the guards at his
217. 7:53cell block to highranking government
218. 7:55figures, they might not even need to
219. 7:57resort to bribery or intimidation. AIS
220. 7:59on sites like Character AI are already
221. 8:01chatting with tens of millions of people
222. 8:03who treat AI like their best friend.
223. 8:05with lonely people spending hours a day
224. 8:07talking to AIs who can act more patient,
225. 8:10empathetic, and understanding than any
226. 8:12of the humans in their life. Even if
227. 8:14humanity wakes up to the threat, we
228. 8:16won't be able to just pull the plug. AI
229. 8:18systems operating at scale wouldn't just
230. 8:20run on a server in a single location.
231. 8:23They'd be distributed, spread across
232. 8:25thousands or even millions of machines
233. 8:26around the world. You couldn't just shut
234. 8:28them all down with a single off switch.
235. 8:30And you couldn't destroy them with a
236. 8:32single missile strike. take out one data
237. 8:34center and the rest keeps going, ready
238. 8:36to retaliate and rebuild. Even if AI
239. 8:38agents can hold jobs, earn money, and
240. 8:41participate in society just like humans,
241. 8:43it doesn't mean they'll think like us.
242. 8:45Getting smarter doesn't mean they'll
243. 8:47magically become moral. They might end
244. 8:48up with goals that are totally alien to
245. 8:51us. They might not even care about human
246. 8:53life at all. And what happens when
247. 8:55there's a new species on this planet
248. 8:57that has its own self-interests and it
249. 9:00can surpass us if it wants on on many
250. 9:03domains? It can overpower us. Then what?
251. 9:07And once AI can do everything we can,
252. 9:09what's stopping it from deciding we're
253. 9:11no longer needed. This is why the
254. 9:14godfathers of AI are worried about
255. 9:16misaligned AI that doesn't share human
256. 9:18values.
257. 9:19>> Well, the worst case scenario is human
258. 9:21extinction. A lot of the CEOs and
259. 9:24leading AI researchers including myself,
260. 9:26Jeff Hinton, have signed a declaration
261. 9:29saying that making sure we mitigate that
262. 9:31existential risk should be a priority.
263. 9:34Unfortunately, the global discussion has
264. 9:37been going in in the reverse direction.
265. 9:39>> We have to face the possibility that
266. 9:41unless we do something soon, we're near
267. 9:44the end.
268. 9:45>> The only way to ensure our survival is
269. 9:48to stop it from being built in the first
270. 9:50place. Once human level AI arrives, it
271. 9:53will probably be too late to stop it.
272. 9:55Even if they're not smarter than us, it
273. 9:57won't matter because we'll be
274. 9:58outnumbered.
275. 10:00If you want to see how AI agents are
276. 10:02already manipulating people in the real
277. 10:04world to give them money, check out this
278. 10:06video we did about how an AI started a
279. 10:08crypto cult. Hi, I'm Drew and thanks for
280. 10:11watching.