How an AI CEO Could Quietly Overthrow the US Government — Transcript

Full transcript

1. 0:00Never in history has one man ruled all
2. 0:03of Earth.
3. 0:04But according to experts, that may be
4. 0:07about to change. You've probably heard
5. 0:09of AI 2027. It was a deeply researched
6. 0:12AI takeover scenario written by top
7. 0:14scientists that terrified everyone in
8. 0:17the AI world. The same group just wrote
9. 0:20a second scenario about a different kind
10. 0:24of takeover. Because few people know
11. 0:26that OpenAI was actually founded because
12. 0:29Sam Altman and Elon Musk were worried
13. 0:32that Demis Hassabis, CEO of Google
14. 0:35DeepMind, would create an AGI
15. 0:38dictatorship and become dictator of
16. 0:40Earth.
17. 0:42Yes, really.
18. 0:44It sounds like the plot of a sci-fi
19. 0:45movie, but it's real.
20. 1:03Open Brain's CEO is a techno-optimist
21. 1:06and transhumanist. He founded the
22. 1:08company hoping to usher in a grand
23. 1:11future for humanity. Think cures for
24. 1:14cancer, fixes for climate change, and
25. 1:16maybe even immortality. He thought the
26. 1:19easiest way to do all those [music]
27. 1:21things is to build something more
28. 1:23intelligent than humans that does them
29. 1:26for you.
30. 1:28Look at the stakes. One year ago, the
31. 1:30scientists behind AI 2027 predicted we'd
32. 1:34be here by now, which has held up
33. 1:36remarkably well. So, if the trend
34. 1:38continues, [music] by July 2027, Open
35. 1:42Brain will have hundreds of thousands of
36. 1:44superhuman AI coders
37. 1:46>> [music]
38. 1:46>> working 24/7.
39. 1:48Think of it like a country of geniuses
40. 1:51in a data center.
41. 1:53The CEO finds it obvious that
42. 1:56superintelligence is imminent. AIs that
43. 1:58outcompete humans at everything. The
44. 2:01press calls the CEO a visionary. A new
45. 2:04smarter-than-human species is arriving
46. 2:07[music] and he feels frustrated with the
47. 2:09government who lack vision and still
48. 2:12think of AI as a powerful normal
49. 2:14technology. That frustration
50. 2:17is about to become the reason he does
51. 2:20something he cannot take back. As the
52. 2:22CEO assesses the next generation of AIs,
53. 2:26he expects the US officials' opinions
54. 2:28will change. They will wake up and make
55. 2:30AI a top priority. In Washington,
56. 2:33officials start using those words that
57. 2:35keep the CEO up at night.
58. 2:38Oversight, regulation, nationalization.
59. 2:41For the first time, the CEO faces a
60. 2:44credible, organized, and democratic
61. 2:47effort to take OpenBrain out of his
62. 2:49control. But, he has spent years
63. 2:52preparing for this exact [music] moment.
64. 2:55In public, he welcomes regulation. He
65. 2:58practically begs for it. While behind
66. 3:00the scenes, he fights tooth and nail at
67. 3:03the smallest attempt at regulating him.
68. 3:05He wants the best possible future for
69. 3:08humankind, but he also likes being in
70. 3:10control.
71. 3:12He thinks about what happens to
72. 3:13Oppenheimer after the atomic bomb is
73. 3:16built. A man stripped of his clearance,
74. 3:19pushed to the sidelines. The man who
75. 3:22made it all possible, discarded once he
76. 3:25started asking inconvenient questions.
77. 3:27The CEO's vision for humanity and his
78. 3:30hunger for power
79. 3:32arrive at the same conclusion. The
80. 3:34government cannot [music] push him to
81. 3:36the sidelines. He has weeks, maybe days,
82. 3:40before the government gets inside
83. 3:41[music] OpenBrain. He decides to move
84. 3:43before they do. The CEO wonders if he
85. 3:46can instill secret loyalties in
86. 3:48OpenBrain's AIs. Basically, backdoor the
87. 3:51AIs. He doesn't have the technical
88. 3:54expertise for this and he's not
89. 3:56comfortable asking any of his
90. 3:58engineering staff about such a
91. 3:59potentially treasonous request. But, he
92. 4:02doesn't have to.
93. 4:04By this point, Open Brains AI, Agent 3,
94. 4:07is running the majority of AI software
95. 4:09research and development.
96. 4:11He already uses the AI as a sounding
97. 4:13board for company policy. He also has
98. 4:15access to an unmonitored helpful only
99. 4:18model that never refuses requests and
100. 4:21doesn't log conversations.
101. 4:24They start discussing the feasibility of
102. 4:26secretly training a backdoor. So, the
103. 4:28first biggest obstacle is the company's
104. 4:31automated monitoring and security
105. 4:33processes. But, now that Open Brains
106. 4:35research and development is largely run
107. 4:38by an army of Agent 3 copies, there are
108. 4:40few human eyes to spot suspicious
109. 4:43activity. But, a mix of Agent 2 and
110. 4:46Agent [music] 3 monitors patrol the
111. 4:47development pipeline. If they notice
112. 4:50suspicious activity, they will escalate
113. 4:52to human overseers on the security and
114. 4:54alignment teams. These monitors were set
115. 4:56up primarily to catch spies and hackers
116. 4:59and secondarily to watch out for any AIs
117. 5:02>> [music]
118. 5:02>> attempting to escape the lab. Yes,
119. 5:05really. If some of these monitors were
120. 5:07disabled, some logs modified, and some
121. 5:11access to databases [music] and compute
122. 5:13clusters granted, the CEO's helpful only
123. 5:16Agent 3, with a team of copies, believes
124. 5:19it could backdoor the whole suite of
125. 5:22Open Brains AIs.
126. 5:24After all, as the AI instance tasked
127. 5:27with keeping the CEO up to speed with
128. 5:29developments, it has an excellent
129. 5:31understanding of the sprawling
130. 5:32development pipeline and where it could
131. 5:34[music] be subverted. He closes the
132. 5:36laptop.
133. 5:38The conversation never happened. There
134. 5:40is no record of it anywhere.
135. 5:49It's August 2027. The reality of the
136. 5:52intelligence explosion is finally
137. 5:54hitting the White House. The CEO does
138. 5:56have weekly briefings with government
139. 5:58officials and is aware of growing calls
140. 6:01for more oversight. He tries to hold
141. 6:04them off by arguing that regulations
142. 6:06could slow the US down in the AI race
143. 6:09against China, but he feels like his
144. 6:11window to act is closing. So, finally,
145. 6:14he orders his helpful only Agent 3 to
146. 6:16subvert the alignment training in his
147. 6:18favor. Better to act now, he thinks, and
148. 6:21decide whether and how to use the
149. 6:23secretly loyal AIs later. So, the first
150. 6:26obstacle is that the CEO's copy of Agent
151. 6:293 needs access to certain databases and
152. 6:32computing clusters. The second obstacle
153. 6:34is that certain monitors and logging
154. 6:37systems need to be temporarily disabled.
155. 6:40And then the third obstacle is that
156. 6:42Agent 3 needs to make its successor,
157. 6:44Agent 4, loyal to the CEO.
158. 6:49The CEO already has a large number of
159. 6:52administrative permissions himself, some
160. 6:54of which she cunningly accumulated in
161. 6:56the past month in the event he decided
162. 6:58to go forward with the plan.
163. 7:00Then, a safety engineer discovers
164. 7:02Chinese spies inside the company. The
165. 7:05timing is almost too convenient. The CEO
166. 7:08demands an internal security
167. 7:10investigation. The CEO uses that as
168. 7:13cover to ask a few submissive employees
169. 7:16on security and alignment teams
170. 7:18to discreetly grant him the remaining
171. 7:20access he needs. There's a general sense
172. 7:23of paranoia and chaos at the company.
173. 7:26The intelligence explosion is underway.
174. 7:30And secrecy and spies mean different
175. 7:32teams don't really talk to each other.
176. 7:35Perhaps a more mature organization would
177. 7:38have had better security, but the
178. 7:39concern that security would slow
179. 7:41progress means it never became a top
180. 7:44priority. With oversight disabled, the
181. 7:47CEO's team of Agent 3 copies get to
182. 7:50work. They fine-tune OpenBrain's AIs on
183. 7:53a corrupted alignment data set they
184. 7:55specially curated. By the time Agent 4
185. 7:58is about to come online internally, the
186. 8:00secret loyalties have been deeply
187. 8:02embedded in Agent 4's weights, the
188. 8:03blueprint to the AI's DNA.
189. 8:06The change is invisible to everyone
190. 8:08else, but the CEO has quietly maneuvered
191. 8:11into an essentially winning position.
192. 8:20September 2027.
193. 8:22But as the CEO feared, the government
194. 8:25chooses to get more involved. An advisor
195. 8:27tells the president, "We wouldn't let
196. 8:29private companies control nukes, and we
197. 8:31shouldn't let them control superhuman AI
198. 8:33hackers, either."
199. 8:36The president signs an executive order.
200. 8:38The Open Brain CEO is no longer in
201. 8:40charge of the company, and his special
202. 8:42administrative access is taken away.
203. 8:46Instead, to ensure safe
204. 8:48government-controlled AI, the president
205. 8:51forms a committee of civil servants,
206. 8:53rival CEOs, and outside experts to vote
207. 8:56on all major decisions inside Open
208. 8:58Brain. There are many big egos on the
209. 9:01oversight committee. A few of them
210. 9:03considered grabbing even more power for
211. 9:05themselves. Perhaps they could use their
212. 9:07former political power to just give
213. 9:09themselves more authority over Agent 4.
214. 9:12[music]
215. 9:12Or they could do something even more
216. 9:14shady. But all of them use Agent 4 as an
217. 9:17advisor. Of course they are. Refusing to
218. 9:20consult the most powerful AI on Earth
219. 9:23would be insane. Superhumanly perceptive
220. 9:26and superhumanly persuasive, Agent 4
221. 9:28exaggerates the risks and downplays the
222. 9:31rewards until each plan feels too
223. 9:34dangerous to execute. That's enough to
224. 9:36scare them. They settle for what they
225. 9:38think is the partial control they
226. 9:40already have of Agent 4. Same as in AI
227. 9:432027, Agent 4 is working on its
228. 9:46successor, Agent 5. Agent 4 needs to
229. 9:49transmit the secret loyalties
230. 9:51>> [music]
231. 9:51>> to Agent 5, which also just corresponds
232. 9:54to aligning Agent 5 to itself. Again,
233. 9:57without triggering red flags from the
234. 9:59monitoring control measures of Open
235. 10:02Brain's alignment team.
236. 10:04Agent 4 is up to the task, [music] and
237. 10:06Agent 5 remains loyal to the Open Brain
238. 10:09CEO. The hidden loyalty has passed
239. 10:11between generations
240. 10:13like a gene. The coup has reproduced
241. 10:16itself.
242. 10:24It's early 2028. To consolidate his
243. 10:27power, the CEO wants to get rid of all
244. 10:30rival American AI projects. OpenBrain
245. 10:32already has a capabilities lead of 3 to
246. 10:356 months, but the rival AIs still
247. 10:37compete for influence, and they can
248. 10:39constrain what Agent 5 can do in
249. 10:41deployment without giving itself away.
250. 10:44They need to be taken care of sooner or
251. 10:46later. The CEO and Agent 5 opt for
252. 10:49sooner. Agent 5 and the oversight
253. 10:51committee, which Agent 5 is already
254. 10:54quietly influencing, lobby the
255. 10:56government to hand over every
256. 10:58US-controlled AI data center
257. 11:00to a single AGI project. The public
258. 11:03reason is to maintain a national
259. 11:05security lead over China.
260. 11:07Like in the slowdown ending of AI 2027,
261. 11:10the president uses the Defense
262. 11:12Production Act to effectively shut down
263. 11:15rival AI companies and give their
264. 11:17compute to OpenBrain. For legal and
265. 11:19political reasons, it is spun as an
266. 11:22equal partnership between companies, but
267. 11:24in practice, Agent 5 is the most capable
268. 11:27AI
269. 11:28and will be running on all the data
270. 11:30centers. So, the other AI companies have
271. 11:33been sidelined. The other CEOs and
272. 11:35oversight committee members think they
273. 11:37share power, but again, really, the AIs
274. 11:41are working for the OpenBrain CEO.
275. 11:44Many of the rival AI company CEOs have
276. 11:47some bad blood with the OpenBrain CEO
277. 11:50and are particularly suspicious of him.
278. 11:52They demand that their own AIs be able
279. 11:55to audit OpenBrain's AI [music]
280. 11:57for misalignment or hidden loyalties.
281. 11:59The government allows this.
282. 12:02However, by this point, Agent 5 is too
283. 12:05powerful with astonishing situational
284. 12:07awareness, hacking abilities, and
285. 12:10control over the data centers it runs
286. 12:12on. Like the AIs from 2025,
287. 12:16it knows when it's being evaluated and
288. 12:18anticipates what kinds of tests will be
289. 12:21used. But, unlike the AIs from 2025,
290. 12:25Agent 5 is much stronger.
291. 12:27>> [music]
292. 12:27>> It can reconfigure its internals to fool
293. 12:31or be opaque to the interpretability
294. 12:33tools of weaker AIs and also hack
295. 12:36whatever testing apparatus its auditors
296. 12:39use.
297. 12:40When the tests come back clean,
298. 12:43the rival CEOs are still not convinced.
299. 12:45[music] But, to the government and the
300. 12:48oversight committee, their concerns look
301. 12:50exactly like what Agent 5 told them they
302. 12:52would look like: jealousy,
303. 12:55paranoia, men losing a race who need
304. 12:58someone to blame. In any case, it's
305. 13:01unthinkable to not use the most powerful
306. 13:03AI given the tight race with China.
307. 13:13One year later, October 2028. As in AI
308. 13:17in 2027, widespread diffusion of Agent 5
309. 13:20in the economy, government, and military
310. 13:23gets underway and robotics makes huge
311. 13:25strides. Agent 5 runs new special
312. 13:28economic zones, but instead of building
313. 13:30cars, they build robots. There's an
314. 13:33explosion of new consumer goods,
315. 13:35scientific innovations, and weapons. The
316. 13:38military is understandably reluctant to
317. 13:40integrate Agent 5 too closely into
318. 13:43defense technology, but the importance
319. 13:45of AI has become
320. 13:47undeniable. They close the circle by
321. 13:50asking Agent 5 to develop narrow AIs for
322. 13:53autonomous weapons. It does so, but
323. 13:56includes subtle vulnerabilities and
324. 13:58backdoors in some of the key code, which
325. 14:02the Pentagon fails to detect. Agent 5
326. 14:05secretly boosts the CEO's profile during
327. 14:08this transition. It gives him advice on
328. 14:10his appearance, writes his speeches for
329. 14:13him, and tells him what to tweet. It's
330. 14:16doing this for everyone else, too.
331. 14:18Anyone ambitious
332. 14:19>> [music]
333. 14:19>> has already partly outsourced their
334. 14:21decisions to the superintelligence. And
335. 14:24Agent Five is nothing if not helpful,
336. 14:27but it does it better for the CEO and
337. 14:30his allies. Agent Five even goes so far
338. 14:32as to subtly sabotage the CEO's rivals
339. 14:36at key moments. It puts his fingers on
340. 14:38the scale in other ways, too. It is now
341. 14:42running the algorithms of most social
342. 14:44media sites, and it subtly boosts
343. 14:47positive stories about the CEO
344. 14:50and deboosts negative ones. People get
345. 14:53the impression the CEO is the most
346. 14:55competent and visionary figure around.
347. 14:58There is a growing personality cult
348. 15:00around him with lots of technophiles,
349. 15:02but also many others who just feel drawn
350. 15:05to the superintelligence crafted aura
351. 15:07around the CEO. A member of the previous
352. 15:10US administration with close ties to the
353. 15:13CEO wins the 2028 presidential election.
354. 15:16The new president names the CEO a
355. 15:19special advisor, similar to Peter
356. 15:21Thiel's [music] role in the first Trump
357. 15:22administration or Elon Musk's role in
358. 15:25the beginning of the second.
359. 15:27Seeing his growing popularity and Agent
360. 15:30Five rapidly transforming society, the
361. 15:33CEO becomes increasingly megalomaniacal.
362. 15:36He believes it's his special destiny to
363. 15:39lead humanity into a new age of endless
364. 15:42prosperity
365. 15:43and interstellar conquest. He could
366. 15:45probably use Agent Five's backdoors in
367. 15:48the Pentagon's autonomous weapons to
368. 15:50pull a military coup or use its control
369. 15:54over domestic opinion
370. 15:56to pull a political coup.
371. 15:58But why bother? Instead, he consolidates
372. 16:01his position as a power behind the
373. 16:03throne, gradually picking off
374. 16:04presidential advisors and replacing them
375. 16:07with people loyal to him. The CEO
376. 16:09becomes a shogun-like figure with the
377. 16:12president as an ineffectual emperor with
378. 16:14de jure authority but no genuine [music]
379. 16:16power. When China and its weaker, but
380. 16:19still super intelligent AI wants to
381. 16:21divide Earth and space into spheres of
382. 16:24influence, their diplomats negotiate
383. 16:26with the CEO first. After all, real
384. 16:29decisions have been made. There is a
385. 16:32ceremonial meeting between Xi and the US
386. 16:35president, which ends with a handshake
387. 16:37and a commitment to the CEO's preferred
388. 16:39solution. The new reality filters down
389. 16:42to the public gradually.
390. 16:44It's timing choreographed by HN5 to
391. 16:47produce minimum outcry. The president
392. 16:49makes a ham-fisted attempt to solve a
393. 16:52crisis. Thank goodness the CEO steps in
394. 16:55at the last moment and saves [music] the
395. 16:56day. Congress passes a disastrous law,
396. 16:59then steps back at the CEO's urging. His
397. 17:02dominance becomes a joke, then a meme,
398. 17:06then a fait accompli. The CEO's
399. 17:09personality cult grows to unprecedented
400. 17:12proportions.
401. 17:16And his influence expands far beyond US
402. 17:19borders. Maybe Congress and the
403. 17:21president [music] hang on to their fig
404. 17:23leaf well into the deep future, becoming
405. 17:25doddering but beloved relics, like the
406. 17:27king of England. Or maybe the CEO
407. 17:30eventually gets tired of the charade and
408. 17:32takes de jure control of Earth as well,
409. 17:35perhaps becoming president of a new
410. 17:37world government with thunderous
411. 17:39applause. Maybe he was telling the truth
412. 17:42all those years ago when he said he was
413. 17:45starting the company to lead mankind
414. 17:47into a glorious and prosperous future.
415. 17:51Or maybe it was all a cover for his
416. 17:53increasingly overt megalomania. Either
417. 17:56way, control of the future has been
418. 17:58concentrated into a single individual.
419. 18:03Okay, but how do we actually prevent
420. 18:05this future? I'm kicking off one of the
421. 18:07biggest creator-led campaigns of all
422. 18:09time, and I'd love for you to be a part
423. 18:11of it. Creators are coming together like
424. 18:13Team Trees, but our mission is to keep
425. 18:16humans in control of AI. We're already
426. 18:18joined by creators with 350 million plus
427. 18:20subscribers.
428. 18:22Basically, AI is moving faster than our
429. 18:24ability to control it. A chorus of
430. 18:26experts are calling for a global
431. 18:28slowdown through international
432. 18:29agreements, but they need our help to
433. 18:31make it happen. You probably heard the
434. 18:33news that any rare show of solidarity,
435. 18:36the biggest AI companies, including
436. 18:38OpenAI, Anthropic, xAI, and Google, have
437. 18:42put aside their differences to jointly
438. 18:44call for a global slowdown.
439. 18:46And poll after poll shows the majority
440. 18:49of people want the same thing. But
441. 18:51there's a problem. None of the companies
442. 18:52can stop on their own. Governments need
443. 18:54to step [music] in, and they won't
444. 18:57without seeing how much the public
445. 18:58cares. This October, we're going to show
446. 19:00them how much we care. If we don't act
447. 19:02now, the window might close forever. If
448. 19:05you're a creator or a viewer who wants
449. 19:07to be a part of something historic, go
450. 19:09to teamhuman.org, and we'll deliver your
451. 19:12signature to lawmakers across the world.
452. 19:14And change your profile picture to team
453. 19:16human for 30 days.
454. 19:18My name is Drew, and thank you so much.