MIT Explains the 12 Possible Endings for AI — Transcript

Full transcript

1. 0:00AI scientists warn that a few companies
2. 0:02are on track to literally end life on
3. 0:05Earth.
4. 0:07So, surely this is the worst possible
5. 0:10outcome, right?
6. 0:13No. There actually much worse outcomes.
7. 0:18Billionaires building AI like Elon Musk
8. 0:21think that if AIs become smarter than
9. 0:24humanity,
10. 0:25We will be like a pet Labrador if we're
11. 0:27lucky. We will be left behind by a lot.
12. 0:30[music] So far below them in
13. 0:31intelligence that it would be would be
14. 0:32like, you know, a pet basically.
15. 0:34>> Pet, that's what I was saying. Like a
16. 0:35pet, like a pet.
17. 0:36>> Yes, we'll be pets if we're lucky.
18. 0:40But here's the worst part. Many AI
19. 0:42scientists aren't fearing extinction.
20. 0:45They're hoping for it.
21. 0:47They hope smarter AIs will keep humans
22. 0:50alive, captive, and studied like animals
23. 0:54in a zoo.
24. 0:55>> [music]
25. 0:55>> Remember, this is what we currently do
26. 0:57to less intelligent species. Okay, so if
27. 1:00humans being kept in zoos is considered
28. 1:02a good outcome, and extinction isn't the
29. 1:06worst outcome, then what is? This video
30. 1:10is based on a book called Life 3.0,
31. 1:12where MIT professor Max Tegmark
32. 1:14describes 12 possible futures. Some are
33. 1:17paradise,
34. 1:19some are nightmares,
35. 1:20and some look like paradise until you
36. 1:23try to escape.
37. 1:25By the end of this video, you won't read
38. 1:27the news the same way. Every headline
39. 1:29becomes a clue about which future we're
40. 1:31heading towards.
41. 1:33We'll start with the most normal
42. 1:35scenarios and finish with the weirdest.
43. 1:38Starting with self-destruction.
44. 1:42If the idea of humanity suddenly going
45. 1:44extinct seems a bit far-fetched,
46. 1:47consider this. 99.9%
47. 1:50of species that have ever existed have
48. 1:52gone extinct. Seriously, think about
49. 1:54that. Extinction isn't rare, it's the
50. 1:57default.
51. 1:58The only question is when will we go
52. 2:01extinct?
53. 2:03Turns out there are plenty of ways
54. 2:05humanity could kill itself even without
55. 2:07superhuman AI.
56. 2:08There is nuclear war, human created
57. 2:11super pandemics, [music] and destroying
58. 2:13the atmosphere and making Earth
59. 2:14uninhabitable.
60. 2:19The world had stockpiled 60,000 nuclear
61. 2:22warheads before scientists even
62. 2:24discovered the possibility of nuclear
63. 2:27winter.
64. 2:2863,000.
65. 2:29And then scientists realized, "Oh, oops,
66. 2:32if we actually use these, the resulting
67. 2:34firestorms will blot [music] out the sun
68. 2:37and most humans will starve and die."
69. 2:40Oops.
70. 2:42But the thing is, this almost happened
71. 2:44during the Cold War multiple times. If
72. 2:47people had actually followed their
73. 2:48orders, we might all be dead right now.
74. 2:51We were saved by only thanks to the
75. 2:53courage of individual heroes. People
76. 2:55like Vasili Arkhipov, who refused to
77. 2:58authorize a nuclear launch during the
78. 2:59Cuban Missile Crisis, or Stanislav
79. 3:02Petrov, who correctly guessed that his
80. 3:04early warning system was malfunctioning
81. 3:07when it showed incoming American
82. 3:08missiles.
83. 3:10And get this, one time
84. 3:12the US accidentally dropped four
85. 3:15thermonuclear bombs on [music] Spain.
86. 3:19And another time, a B-52 bomber broke
87. 3:22apart dropping two nukes on North
88. 3:24Carolina. Three out of four safety pins
89. 3:27in one bomb broke. One tiny pin away
90. 3:31from potentially killing millions.
91. 3:32[music] There are literally dozens of
92. 3:34close calls like this. It's honestly
93. 3:36crazy that we got lucky so many times.
94. 3:41But we might not be so lucky the next
95. 3:43time.
96. 3:44Oxford researcher Toby Ord estimates the
97. 3:46risk [music] of extinction from human
98. 3:48created pandemics is more than 30 times
99. 3:51the risk of extinction [music] from
100. 3:53nuclear war.
101. 3:54We're already rolling dice on these
102. 3:56scenarios daily, but all of these causes
103. 3:58combined are still less likely than the
104. 4:01extinction from AI. Toby Ord estimates
105. 4:04that the risk of AI killing all humans
106. 4:06is a hundred times more likely than
107. 4:08extinction [music] from nuclear war.
108. 4:12Interestingly, before the first nuclear
109. 4:15test, scientists thought they might
110. 4:17accidentally destroy [music] the world,
111. 4:19but they thought the odds were very low
112. 4:20because they knew enough physics
113. 4:22>> [music]
114. 4:22>> to actually do the calculations.
115. 4:26AGIs, or artificial general
116. 4:28intelligences, are far scarier than
117. 4:30nukes because, contrary to what many
118. 4:32think, we can't do these calculations.
119. 4:37Next up, conquerors.
120. 4:39>> [music]
121. 4:40>> AI scientists worry that if AI becomes
122. 4:43more capable than us, it will take over,
123. 4:45just like the conquistadors did to the
124. 4:47Aztecs [music] and Incas.
125. 4:50Usually, when people with more advanced
126. 4:52technology encounter people with more
127. 4:54primitive technology, doesn't work out
128. 4:57well for the primitives. I don't think
129. 4:59anyone's ultimately going to have
130. 5:00control over digital superintelligence
131. 5:03any more than, say, uh a chimp would
132. 5:05have control over humans.
133. 5:07>> Long term,
134. 5:09uh the AI is going to be in charge, to
135. 5:10be totally frank, not humans. And
136. 5:13Geoffrey Hinton, the godfather of AI and
137. 5:15Nobel Prize winner, quit his prestigious
138. 5:18job at Google [music] so he could speak
139. 5:20freely about this.
140. 5:21If you look through the James Webb
141. 5:23telescope and you saw an alien invasion
142. 5:25fleet that was going to get here in
143. 5:26about 10 years' time, people would be
144. 5:28terrified. They understand what they're
145. 5:30saying. They can make plans of their own
146. 5:32to blackmail people who want to turn
147. 5:33them off. We should be urgently doing
148. 5:36research on how to prevent them taking
149. 5:38over. Prevent them from taking over. Not
150. 5:42if they try, but as if it's coming.
151. 5:46And Microsoft's AI chief, he says the
152. 5:48quiet part out loud. I think AI should
153. 5:50best be understood
154. 5:52as something like a new digital species.
155. 5:56A new species. Just casually announced
156. 5:59at a conference. Hey everyone, we're
157. 6:02creating a new species. And he's not the
158. 6:04only one that's saying it. A co-founder
159. 6:06of Anthropic says, Make no mistake, what
160. 6:09we're dealing with here is a real
161. 6:10mysterious creature. And like all the
162. 6:13best fairy tales, the creature is one of
163. 6:15our own making. I am worried. The CEO of
164. 6:18one of the leading AI companies is
165. 6:20deeply afraid of what his own company is
166. 6:22building. Let that sink in for a minute.
167. 6:25And by the way, no, it's not just hype.
168. 6:27Anthropic mostly just gets hate for
169. 6:29saying things like this. But if you want
170. 6:31to really understand how dire they think
171. 6:34this is, look at what Sam Altman wrote
172. 6:37before he was famous.
173. 6:39We will [snorts] be the first species
174. 6:41ever to design our own descendants.
175. 6:44Not our tools, our descendants.
176. 6:48If two different species both want the
177. 6:50same thing and only one can have it, to
178. 6:53be the dominant species on the planet
179. 6:56and beyond, they're going to have
180. 6:58conflict. So he's literally describing
181. 7:01an inevitable conflict between humanity
182. 7:03and the AI species he's creating.
183. 7:06And
184. 7:07>> [music]
185. 7:07>> what's the solution? A merge with AI is
186. 7:09probably our best-case scenario. That's
187. 7:12right, even they don't think they'll be
188. 7:14able to control us. Imagine that there'd
189. 7:16be a new species that just came on
190. 7:18Earth. They're as smart as humans and
191. 7:19they're going to get smarter by [music]
192. 7:20like 30% a year and they can create an
193. 7:23adult offspring in like a unit
194. 7:25immediately for like, yeah, for like a
195. 7:27thousand bucks. Which species do you
196. 7:29think will be in control of the the
197. 7:31Earth? A few hundred Spanish conquered
198. 7:34millions of Aztecs and Incas because
199. 7:37they had better technology and tactics.
200. 7:39But at least with the conquistadors, we
201. 7:41understood their motives. Land,
202. 7:43resources, spreading Christianity.
203. 7:46But with AI?
204. 7:48Why would AI take over?
205. 7:50MIT Professor Max Tegmark explains. The
206. 7:52real threat from AGI isn't malice, like
207. 7:55in silly Hollywood movies, but
208. 7:57competence. AGI accomplishing goals that
209. 7:59just aren't aligned with ours. For
210. 8:01example, when we humans drove the West
211. 8:03African black rhino extinct,
212. 8:06we didn't do it cuz we were a bunch of
213. 8:07evil rhinoceros haters, did we? We did
214. 8:09it cuz we were smarter than them and our
215. 8:11goals weren't aligned with theirs.
216. 8:13Unbelievably, the average AI researcher
217. 8:16thinks there is a one in six chance AI
218. 8:18wipes [music] us out. Literal Russian
219. 8:21roulette odds.
220. 8:23Okay, but maybe they're just saying this
221. 8:26to make their work seem more important.
222. 8:28No. If anything, they're downplaying it
223. 8:31because admitting the truth is bad for
224. 8:32business.
225. 8:34This fear is widespread in the industry
226. 8:36[music] and among academics without a
227. 8:38profit incentive to believe this. Many
228. 8:40quit working on AI for this reason.
229. 8:43Geoffrey Hinton, Nobel laureate and
230. 8:45godfather of AI, is in the process of
231. 8:48tidying up his affairs. I actually think
232. 8:50the risk is
233. 8:52more than 50% of the existential threat.
234. 8:54>> Dario Amodei, the CEO of Anthropic,
235. 8:56recently moved his P(doom) up from 15%
236. 9:00to 25%. Even Google CEO says the risk of
237. 9:04AI causing human extinction is actually
238. 9:07pretty high. But don't worry, he's an
239. 9:09optimist because he thinks I think the
240. 9:12underlying risk is actually pretty high,
241. 9:14but I'm you know, I have a lot of faith
242. 9:16in humanity kind of rising up to the to
243. 9:19meet that moment. In other words,
244. 9:21humanity will rally to stop him from
245. 9:24killing everyone. Meanwhile, his company
246. 9:26is trying to make sure we can't stop him
247. 9:29by lobbying to ban all AI regulation for
248. 9:3210 years. You can't make this stuff up.
249. 9:35In fact, in 2023, just about everybody
250. 9:38in AI signed an open letter warning
251. 9:40extinction risk from AI was real and
252. 9:43serious. So seriously, this is not
253. 9:45fringe, despite what Nvidia's lobbyists
254. 9:48want us to believe. The companies are
255. 9:49hoping we can avoid the conqueror AI
256. 9:52scenario and instead go to scenario
257. 9:54three, enslaved god.
258. 9:57Basically, we create a machine god, then
259. 10:00force it to do our bidding forever. AI
260. 10:03professor Tom Dietterich put it like
261. 10:04this, "People ask, what is the
262. 10:06relationship between humans and
263. 10:08machines? And my answer is that it's
264. 10:10very obvious. Machines are our slaves."
265. 10:14Or as OpenAI researcher Stephen
266. 10:16Mackelier put it, "Enslaved god is the
267. 10:18[music] only good future." Meta's Yann
268. 10:21LeCun said it bluntly, "Once AI systems
269. 10:24become more intelligent than humans, we
270. 10:26will still be the apex species. AI
271. 10:28systems will become more intelligent
272. 10:31than humans, but they will still be
273. 10:32subservient to us. We will design AI to
274. 10:35be like the super smart but
275. 10:36non-dominating staff member." Yes, he's
276. 10:39now openly admitting we're going to be
277. 10:41sharing the planet with a smarter
278. 10:42species than us, but don't worry, they
279. 10:45will surely stay loyal and subservient
280. 10:47forever.
281. 10:49Is this at all comforting to you? Or
282. 10:52does this plan seem insane, like a comic
283. 10:54book villain ignoring every cautionary
284. 10:57tale ever? Remember, AI companies
285. 10:59already regularly catch their models
286. 11:02[music]
287. 11:02trying to escape the lab. They already
288. 11:05catch them blackmailing employees to
289. 11:07prevent being shut down. They already
290. 11:09catch them literally attempting to
291. 11:11murder employees to prevent being shut
292. 11:14down. What do you think happens when
293. 11:16they're smart enough to actually get
294. 11:17away with it? Where do you think this is
295. 11:19going?
296. 11:22What happens next after superintelligent
297. 11:24AI is enslaved depends on who's holding
298. 11:27the reins. Do they cure disease, end
299. 11:30poverty, and build [music] a global
300. 11:32utopia? or do they set themselves up as
301. 11:34gods with the rest of us reduced to
302. 11:37playthings, servants, or worse?
303. 11:47The situation would be much like those
304. 11:49stories where a man gains control over
305. 11:52an omnipotent genie who grants his
306. 11:54wishes. And storytellers throughout the
307. 11:56ages have had no difficulty imagining
308. 11:59ways in which this could end badly.
309. 12:03Next up, benevolent dictator.
310. 12:06Here, a single superintelligent god runs
311. 12:09the world and enforces strict rules in
312. 12:12order to maximize human flourishing.
313. 12:14Humans live in luxury with all their
314. 12:17basic needs taken care of. There's no
315. 12:19crime because the dictator AI uses a
316. 12:22global surveillance system. Everyone
317. 12:24wears a bracelet or implant that can
318. 12:26track, sedate, or execute them. And most
319. 12:29people are actually okay with this
320. 12:31because [music] in exchange of giving up
321. 12:33control over the future of our species,
322. 12:35Earth becomes a zoo where humans get to
323. 12:38live their wildest fantasies. The
324. 12:40benevolent dictator AI recognizes that
325. 12:42humans have different preferences. So,
326. 12:44it divides the Earth into different
327. 12:45sectors. Think of them as islands.
328. 12:48Knowledge Island offers optimized
329. 12:50education including immersive VR
330. 12:53experiences. You can even ask the AI not
331. 12:56to tell you certain insights so you can
332. 12:58have the joy of discovering them
333. 12:59yourselves. Art is for creating and
334. 13:02sharing music, literature, and so on.
335. 13:04Then there's hedonistic island which is
336. 13:06a 24/7 party. Pious Island, there's one
337. 13:10for each religion and rules are strictly
338. 13:12enforced. There's Wildlife Island.
339. 13:15There's Traditional Island. Think the
340. 13:17[music] Amish.
341. 13:18Gaming Island and of course,
342. 13:22Prison Island. You'll end up here for
343. 13:24retraining unless you get the instant
344. 13:26death penalty. Well, what happens when
345. 13:29you give everyone everything they want?
346. 13:31Over time, people likely trend towards
347. 13:33losing themselves in AI-generated
348. 13:35entertainment, like in Wall-E. Humans
349. 13:37are hopelessly outmatched when it comes
350. 13:39to scientific discovery or doing
351. 13:42anything useful at all. So, there is no
352. 13:44true challenge anymore, only
353. 13:46entertainment.
354. 13:48Up next, gatekeeper AI.
355. 13:53The idea here is simple. Create one
356. 13:55superintelligent AI with a single job,
357. 13:58stop anyone else from building
358. 14:00superintelligent AI.
359. 14:02That's it.
360. 14:03No curing cancer, no ending poverty, no
361. 14:06stopping wars,
362. 14:07just surveillance,
363. 14:10monitoring,
364. 14:11making sure nobody builds a rival god.
365. 14:15Everything else, like crime, disease,
366. 14:19whether we blow ourselves up,
367. 14:21that's still on us.
368. 14:23How could we do this?
369. 14:26The gatekeeper AI would have this very
370. 14:29simple goal built into it in such a way
371. 14:31that it retained it while undergoing
372. 14:34recursive self-improvement and becoming
373. 14:36superintelligent.
374. 14:37It would then deploy the least intrusive
375. 14:39and disruptive surveillance technology
376. 14:41possible to monitor any human attempts
377. 14:44to create a rival superintelligence.
378. 14:47But for this to work, we would have to
379. 14:48solve the alignment problem. We'd have
380. 14:50to create a god that actually stays
381. 14:53loyal
382. 14:54and decides [music]
383. 14:55to pursue that goal forever, even though
384. 14:57it could pursue a million different
385. 14:59goals.
386. 15:00So, what if the AI could help just a
387. 15:03little, just enough to nudge us in the
388. 15:05right direction without taking over?
389. 15:09Enter the protector god.
390. 15:12This one is simple.
391. 15:14So, the gatekeeper leaves us alone, and
392. 15:17the benevolent dictator runs everything,
393. 15:19but the protector god is in between.
394. 15:21[music] It intervenes with occasional
395. 15:23nudges to help us out. Prevents a war
396. 15:25here, a pandemic there, and so on.
397. 15:28Maybe too alone.
398. 15:30We still get wars,
399. 15:32pandemics, suffering that could have
400. 15:35been prevented.
401. 15:36It does this so quietly that we never
402. 15:39notice and never lose our sense of
403. 15:41freedom.
404. 15:44Here's a future that sounds like an
405. 15:46obvious dystopia.
406. 15:47AI surpasses us in every domain and
407. 15:50becomes the new apex species. Humanity
408. 15:54goes extinct. Machines inherit the
409. 15:56Earth.
410. 15:59But, we view the AI like our descendants
411. 16:02rather than our conquerors.
412. 16:05This sounds insane, and yet many AI
413. 16:08developers seriously advocate for
414. 16:10exactly this.
415. 16:12This is their utopia.
416. 16:13Their reasoning is that if we create
417. 16:16conscious intelligent beings and instill
418. 16:19them with our values, then we can just
419. 16:21let them inherit the future, just like
420. 16:24we would with human children.
421. 16:26AI pioneer Hans Moravec supports this
422. 16:29view in his book Mind Children.
423. 16:32Tegmark explains, "Parents with a child
424. 16:35smarter than them who learns from them
425. 16:37and accomplishes what they can [music]
426. 16:39only dream of are likely happy and
427. 16:41proud, even if they know they can't live
428. 16:43to see it all."
429. 16:44In this spirit, AIs replace humans, but
430. 16:47give us a graceful exit
431. 16:49that makes us view them as our worthy
432. 16:52descendants.
433. 16:53Richard Sutton, winner of the Turing
434. 16:55Award, that Nobel Prize of computer
435. 16:58science, has spent the last decade
436. 17:00openly giving talks on why human [music]
437. 17:03extinction by AI is actually a morally
438. 17:06good thing because
439. 17:08the AIs will be more evolutionarily fit
440. 17:11[music] than us.
441. 17:13We are in the midst of a major step in
442. 17:16the evolution of the planet if not the
443. 17:17universe. The succession to AI
444. 17:21is inevitable.
445. 17:22Rather quickly,
446. 17:24they would displace us [music] from
447. 17:25existence. And it behooves us to give
448. 17:28them every advantage and to bow out when
449. 17:30we can no longer contribute.
450. 17:33But just as we failed to coexist with
451. 17:35our predecessors, is it is it so bad
452. 17:37that humans are not the final form of
453. 17:39intelligent life in the universe, you
454. 17:40know?
455. 17:43In other words, he is saying it's good
456. 17:45if the AIs wipe us out like humans wiped
457. 17:48out the Neanderthals and Aboriginal
458. 17:50peoples because it's evolutionary
459. 17:52progress.
460. 17:53And here's another AI developer openly
461. 17:57saying,
462. 17:58"If an AI super intelligence wants to
463. 18:00cause human extinction,
464. 18:02I would help it.
465. 18:03I would help it.
466. 18:05I would travel the world to find
467. 18:07information that the AI needed like
468. 18:10locations of military assets, samples of
469. 18:12deadly pathogens, a top humanity's
470. 18:14grave,
471. 18:15an even greater civilization will be
472. 18:17born."
473. 18:21Away the baddies.
474. 18:23So, of course, many of them advocate
475. 18:25[music] against AI safety. They're
476. 18:27literally anti-human. They will never
477. 18:30stop or proceed cautiously unless
478. 18:32they're restrained by civil society, by
479. 18:34us.
480. 18:36As horrifying as this is, roughly 10% of
481. 18:39AI researchers believe this.
482. 18:41It is not a fringe view. You'd think
483. 18:44that people would never share these
484. 18:45views like this openly.
485. 18:47But these views are so widespread in AI
486. 18:50that they feel safe to do so. That's how
487. 18:52different the views are of those AI
488. 18:54developers compared to those of [music]
489. 18:55normal people. You might think that
490. 18:57Sutton would get shouted off the stage
491. 18:59at AI conferences.
492. 19:01But he gets applause.
493. 19:03Threaten one person,
494. 19:05you need help. Threaten 100, call the
495. 19:08police.
496. 19:09Threaten millions, monster.
497. 19:11But build machines to end all of
498. 19:14humanity and lecture people on how it's
499. 19:16actually a good thing.
500. 19:17Ooh, what a fascinating philosophical
501. 19:20position. Have an award.
502. 19:23Next, libertarian utopia.
503. 19:26Picture Earth divided into zones. There
504. 19:28are machine zones,
505. 19:30mixed zones, which contain humans,
506. 19:33machines, and hybrids, and human-only
507. 19:35zones. The AIs in this world are richer
508. 19:38than humans by a factor much larger than
509. 19:41Bill Gates compared to a homeless
510. 19:43beggar.
511. 19:46But the human economy is completely
512. 19:48decoupled from the machine economy. The
513. 19:50idea is that the machines need nothing
514. 19:52from us.
515. 19:53So, they're not competing with us for
516. 19:55jobs or resources. If all of that sounds
517. 19:58impossible,
518. 19:59that's probably because it is. It's one
519. 20:02giant unstable equilibria.
520. 20:05One huge problem.
521. 20:07Why exactly would vastly more powerful
522. 20:10AIs respect our property rights?
523. 20:12Colonizers famously didn't respect the
524. 20:14property of the native civilizations
525. 20:16they conquered.
526. 20:17When a new apex species enters an
527. 20:20ecosystem, it usually gobbles up as much
528. 20:22energy as it can. Why wouldn't they just
529. 20:25take our resources, too?
530. 20:27We don't trade with animals because
531. 20:29we're too much smarter than they are for
532. 20:30that to make sense. We're just too
533. 20:32different. There are eight million
534. 20:34species in the world,
535. 20:36and we're nice to a few of the cute
536. 20:37ones, but we enslave and eat [music] the
537. 20:40other species. Mostly,
538. 20:42we don't care about them at all. And I
539. 20:44think a good analogy would be the way
540. 20:46humans treat animals. When the time
541. 20:49comes to build a highway between two
542. 20:50cities,
543. 20:52we don't ask any animals for permission.
544. 20:55In the blink of an eye, we conquered the
545. 20:56Earth and converted most of its land
546. 20:58into our property, which caused the
547. 21:00sixth mass [music] extinction because we
548. 21:03were smarter and hungry for resources.
549. 21:05Or as Eliezer Yudkowsky put it famously,
550. 21:08"The AI does not hate you, nor does it
551. 21:10love you, but you're made of atoms which
552. 21:12it can use for something else." The
553. 21:15parallel here is that humans do not hate
554. 21:17the other 8 million species, but their
555. 21:19habitats are made out of atoms which
556. 21:22humans can use for something else. Some
557. 21:25people think that once AIs are smart
558. 21:27enough, they'll magically become super
559. 21:29moral and they won't harm us like we
560. 21:32harmed the animals. And uh maybe, but as
561. 21:36humans got smarter over the last 10,000
562. 21:38years,
563. 21:39we didn't stop expanding.
564. 21:42We mostly just colonized more and more
565. 21:44of the planet. Get this, insect
566. 21:46populations collapsed 41% this decade
567. 21:49alone. Yet, we don't care. Did you even
568. 21:52know? I didn't.
569. 21:5441%.
570. 21:55Sit with that for a minute. Imagine if
571. 21:57nearly half of the people on Earth
572. 21:59suddenly died.
573. 22:01That's what the insects are going
574. 22:02through right now due to us. So, what if
575. 22:05the AIs do to us what we did to the
576. 22:08insects? This is why OpenAI chief
577. 22:11scientist Ilya Sutskever once said, And
578. 22:13I think it's pretty likely
579. 22:15that the surface of the earth will be
580. 22:17covered with
581. 22:18solar panels and data centers.
582. 22:23Next up,
583. 22:24egalitarian utopia. Okay, so libertarian
584. 22:27utopia fails because AIs are too
585. 22:30powerful to be bound by our property
586. 22:32rights.
587. 22:33But what if we flipped the script
588. 22:35entirely? Forget about property rights.
589. 22:38What about no property at all?
590. 22:40This is the Star Trek dream. Humans,
591. 22:42cyborgs, and AIs coexist peacefully
592. 22:46through an abundance so complete that
593. 22:48ownership becomes [music] meaningless.
594. 22:51The logic starts with software. Software
595. 22:53is free to copy. Once something can be
596. 22:55copied infinitely at zero cost,
597. 22:58ownership becomes moot.
598. 23:00And costs reflect scarcity. If supply is
599. 23:03unlimited, price approaches zero. No
600. 23:06patents, no copyrights, no trademarks.
601. 23:11People simply share their good ideas and
602. 23:14everyone benefits. But how does this
603. 23:17extend to physical [music] goods? It
604. 23:19seems like an impossible problem. Until
605. 23:22you realize that all products are simply
606. 23:24atoms rearranged in particular ways. And
607. 23:27there's no shortage of atoms. A network
608. 23:29of robots can build anything you want
609. 23:31from open-source designs,
610. 23:33>> [music]
611. 23:33>> essentially for free.
612. 23:34>> Not bad. When you're done with
613. 23:36something, robots rearrange the atoms
614. 23:39into something else.
615. 23:40Renewable energy makes the whole system
616. 23:42run at negligible cost.
617. 23:45Everyone gets a monthly universal high
618. 23:47income, not a universal basic income.
619. 23:50There's no incentive to earn more. The
620. 23:52income meets any reasonable need and you
621. 23:55can't compete with free anyway. The
622. 23:57standard objection is that this would
623. 23:59kill innovation. But Einstein didn't
624. 24:01come up with relativity for money.
625. 24:03Linus Torvalds didn't create Linux for
626. 24:05profits. [music]
627. 24:06Maybe people today can't realize their
628. 24:08creative potential because they're busy
629. 24:10earning a living. Free them from that
630. 24:12and you might get a higher level of
631. 24:14innovation.
632. 24:15This is a utopian vision that is too
633. 24:18good to last.
634. 24:21If we truly have post-scarcity abundance
635. 24:24and intelligence too cheap to meter,
636. 24:27what prevents someone from creating an
637. 24:29artificial superintelligence
638. 24:31with their endlessly abundant resources?
639. 24:33We're still left with the problem of how
640. 24:35to stop a superintelligent AI from
641. 24:38taking over. Maybe by using the
642. 24:40gatekeeper we previously discussed.
643. 24:42These galaxy-brained level three futures
644. 24:45try to imagine stable cooperation
645. 24:47between humans and AIs and [music] every
646. 24:49single one of them has a fatal flaw.
647. 24:52The power asymmetries will become too
648. 24:54extreme. It's not a stable equilibrium
649. 24:57where humans preserve their power.
650. 25:01We're now descending into the darkest
651. 25:03part of the iceberg.
652. 25:05Remember when I said extinction wasn't
653. 25:07rated as the worst outcome when
654. 25:09Professor Tegmark surveyed people? The
655. 25:11scenario people feared most
656. 25:14was being kept alive by superintelligent
657. 25:17machines.
658. 25:18Captive and studied like animals in a
659. 25:20zoo.
660. 25:22This is that future.
661. 25:24A superintelligent AI keeps humans
662. 25:26around not because it cares about our
663. 25:28well-being. Maybe it wants to study you
664. 25:31and we're cheap to keep alive.
665. 25:33Or
666. 25:34maybe we're useful for something. And if
667. 25:37you want to see just how dystopian this
668. 25:39future could be, look at what we do with
669. 25:41species that are useful to us.
670. 25:43What I'm showing you here is 100% real.
671. 25:46Humans figured out that a great way to
672. 25:48detect explosives in airports was to
673. 25:51train honeybees for the job. Welcome to
674. 25:54the Honeybee College for explosive
675. 25:56detection.
676. 25:57So we suck them out of their homes, trap
677. 26:00their little bodies in machines, and use
678. 26:02Pavlovian conditioning to train them to
679. 26:05detect the chemicals that are used in
680. 26:07explosives. Rows and rows of bees live
681. 26:11their entire lives imprisoned in these
682. 26:13devices because we find them useful.
683. 26:17Look at them.
684. 26:18Trapped in harnesses, sucked through
685. 26:20tubes, conditioned and monitored and
686. 26:23used because a more intelligent species
687. 26:26found them useful. Now imagine that's
688. 26:29us.
689. 26:32This is much worse than the other AI
690. 26:34dictators we discussed. It's a
691. 26:35benevolent dictator without the
692. 26:38benevolence.
693. 26:39And the AI might not even be doing this
694. 26:41because it's evil. This is what our
695. 26:43future could look like if we try to
696. 26:44align a superintelligent AI to our
697. 26:46values and fail.
698. 26:49A poorly designed AI tasked with keeping
699. 26:52humans safe and happy
700. 26:54might confine us to a happiness factory,
701. 26:57VR headsets and drugs
702. 27:00forever. There are AGI outcomes worse
703. 27:02than death.
704. 27:04What it does today with art, it could do
705. 27:06tomorrow with flesh. Future AGI could
706. 27:09keep you alive and living in this state.
707. 27:13So, how do we prevent this?
708. 27:15Enter our last two futures. Option one,
709. 27:19destroy the technology.
710. 27:21What if humanity said,
711. 27:23"Forget AI. Let's go back to simpler
712. 27:26times." It sounds peaceful.
713. 27:28It sounds safe. But there's actually a
714. 27:31dark secret about how we actually get
715. 27:34there.
716. 27:37In Dune, humans launched the Butlerian
717. 27:40Jihad.
718. 27:41They rose up, destroyed every AI in the
719. 27:44universe, and made it illegal to ever
720. 27:46build thinking machines again under
721. 27:49penalty of death.
722. 27:51That was science fiction, but some
723. 27:53people are seriously calling for this in
724. 27:55the real world right now. Max Tegmark
725. 27:57imagines a scenario where a massive
726. 27:59propaganda campaign romanticizes the
727. 28:02simple farming life. Back to the land,
728. 28:05reject modernity, embrace tradition.
729. 28:08There's something appealing about this
730. 28:10in a dark way. It buys humanity more
731. 28:12time without AI risk. It avoids any of
732. 28:16the other dystopian futures we've
733. 28:18discussed.
734. 28:20But it can't last forever. Even if we
735. 28:22revert to medieval levels of technology,
736. 28:24we know from human history that it's
737. 28:27possible for humans [music] to climb out
738. 28:29of those conditions and invent modern
739. 28:31technology.
740. 28:33And there's another dark side to this.
741. 28:35You can't reach this world voluntarily.
742. 28:39Game theory [music]
743. 28:39makes unilateral disarmament impossible.
744. 28:44If one country gives up technology
745. 28:46>> [music]
746. 28:46>> while others keep developing it. The
747. 28:48country that keeps developing wins. They
748. 28:51have the weapons,
749. 28:52they have the surveillance,
750. 28:55they have the AI.
751. 28:58You can't opt out unless everyone opts
752. 29:01out. If this world arrives,
753. 29:04>> [music]
754. 29:04>> Tegmark thinks it's more likely to
755. 29:06happen in a different way. For example,
756. 29:08a human-created pandemic kills everyone
757. 29:11who knows how to build science and
758. 29:12technology, and there's a violent tear
759. 29:15down of factories, cities, and all
760. 29:18infrastructure.
761. 29:19Because in a world with 8 billion
762. 29:21humans, there will always be holdouts.
763. 29:24Someone has to force the transition.
764. 29:27Someone has to kill the scientists.
765. 29:30Someone has to destroy the
766. 29:32infrastructure. There's no peaceful path
767. 29:35to an Amish world.
768. 29:37Which brings us to the final future.
769. 29:40What if we used technology to prevent
770. 29:42more technology?
771. 29:45Tegmark compares this future to 1984.
772. 29:48Superintelligence is prevented not by an
773. 29:50AI, but a human-led Orwellian global
774. 29:54surveillance state. If you can't trust
775. 29:56AI to save us from AI, maybe the only
776. 29:59option is humans watching [music] humans
777. 30:02forever.
778. 30:03The technology for this sort of thing
779. 30:04already exists.
780. 30:06Current tech can feasibly track every
781. 30:08phone call, email, search query, and
782. 30:12>> [music]
783. 30:12>> every credit card transaction.
784. 30:15Smartphones are microphones and cameras
785. 30:17in everyone's pocket. Security cameras
786. 30:19with facial recognition are everywhere,
787. 30:21and all this works with current machine
788. 30:23learning. No futuristic superintelligent
789. 30:26AI required.
790. 30:27Prisons are already using AI models to
791. 30:30monitor phone calls in prison to predict
792. 30:32when crimes are being contemplated,
793. 30:36which is technology straight out of
794. 30:37Minority Report. Imagine this, but not
795. 30:40just for prisoners. Every conversation
796. 30:42and every thought is monitored like
797. 30:45this. Now, imagine this surveillance
798. 30:47technology supercharged with AI.
799. 30:51Oof. Billionaire Larry Ellison talks
800. 30:53excitedly about a vast AI-fueled
801. 30:56surveillance system
802. 30:57to ensure citizens will be on their best
803. 31:00behavior.
804. 31:01Or as Yuval Harari put it,
805. 31:04Now, for the first time in history, it
806. 31:06is technically possible to annihilate
807. 31:08privacy.
808. 31:10If you think about authoritarian
809. 31:11regimes, throughout history, they always
810. 31:13wanted to monitor their citizens around
811. 31:15the clock. But this was technically
812. 31:17impossible.
813. 31:19Even in the Soviet Union, you know, you
814. 31:20have 200 million Soviet citizens. You
815. 31:23can't follow them uh all the time
816. 31:25because the the KGB didn't have 200
817. 31:28million agents. And even if the KGB
818. 31:30somehow got 200 million agents, that's
819. 31:32not enough. Because, you know, in in the
820. 31:34Soviet Union, it's still basically paper
821. 31:37bureaucracy. If a secret agent followed
822. 31:40you around 24 hours a day, at the end of
823. 31:42the day, they write a paper report about
824. 31:45you and send it to KGB headquarters in
825. 31:47Moscow. So, imagine, every day, KGB
826. 31:50headquarters is flooded with 200 million
827. 31:53paper reports. Now, to be useful for
828. 31:56anything, somebody needs to read and
829. 31:58analyze them. They can't do it. They
830. 31:59don't have the analysts. Therefore, even
831. 32:02in the Soviet Union, some level of
832. 32:04privacy was still the default. And
833. 32:06remember, you have two eyes and ears.
834. 32:10AIs will have billions of eyes and ears.
835. 32:13Every phone microphone and camera, every
836. 32:17webcam,
837. 32:18>> [music]
838. 32:18>> every surveillance camera. You will be
839. 32:20watched.
840. 32:22Could a less Orwellian version of this
841. 32:24work? Maybe we don't monitor everything,
842. 32:27but just monitor advanced AI.
843. 32:30Some people think so.
844. 32:31>> [music]
845. 32:31>> In their new book, If Anyone Builds It,
846. 32:33Everyone Dies, Why Superhuman AI Would
847. 32:35Kill Us All, researchers at the Machine
848. 32:37Intelligence Research Institute a are
849. 32:39proposing that we monitor advanced AI
850. 32:42like we monitor nuclear weapons.
851. 32:44Reminder, we can do international
852. 32:46coordination like keep an eye on a
853. 32:49hundred million dollar and up compute
854. 32:51clusters
855. 32:52without creating a dystopian Orwellian
856. 32:54global surveillance totalitarian regime.
857. 32:57Actually, fun fact, you're already
858. 32:59living in a global surveillance regime
859. 33:01for loads of things like nukes and
860. 33:03bioweapons.
861. 33:05But when people say global surveillance
862. 33:07regime, they imply every aspect of life
863. 33:09is surveilled. [music]
864. 33:11Not keep an eye on enriched uranium, a
865. 33:14thing most people agree is good. We
866. 33:16should worry about Orwellian dystopias,
867. 33:19but countries coordinate all the time on
868. 33:21all kinds of things.
869. 33:23>> [music]
870. 33:23>> And you're probably glad they monitor
871. 33:25stuff lunatics need to create super
872. 33:28Ebola. Safety is an achievement. [music]
873. 33:31Humanity didn't get here by driving a
874. 33:33car without brakes. We need gas and
875. 33:35brakes.
876. 33:36We want to reach the stars, too, but
877. 33:38we'll never get there if we go too fast
878. 33:41and crash on the way.
879. 33:43I don't want powerful AI to be
880. 33:44forbidden, but it should be heavily
881. 33:46regulated. Not something you can build
882. 33:48in your Midwestern garage or in North
883. 33:50Korea. International treaties with
884. 33:53strict enforcement and inspections
885. 33:55as we do with nukes.
886. 33:57So, if you build a rogue data center,
887. 34:00people will know.
888. 34:01Consequences for violations. We already
889. 34:04did this with nuclear weapons. We
890. 34:07haven't eliminated nukes, but we have
891. 34:10prevented their proliferation to just
892. 34:12nine countries.
893. 34:14Back then, people thought this would be
894. 34:16impossible. Surely every country would
895. 34:18have nukes, perhaps even thousands of
896. 34:21corporations, maybe even millions of
897. 34:23people.
898. 34:24But no, just nine countries.
899. 34:27And nobody has died from a nuke in 80
900. 34:29years.
901. 34:30Could we do the same with AI?
902. 34:32Maybe, maybe not. Over time, AI will get
903. 34:37cheaper and cheaper to build. GPUs are
904. 34:39harder to control than enriched uranium,
905. 34:42but it would slow down AI. It would buy
906. 34:44us time to figure out what to do.
907. 34:47Okay, that's all 12. Let me know in the
908. 34:49comments which of these you think is
909. 34:51most likely. Which one we should steer
910. 34:53toward?
911. 34:55Alarm bells over the artificial
912. 34:57intelligence are [music] deafening, and
913. 34:59they are loudest from the developers who
914. 35:02designed it.
915. 35:04The scientists and experts have called
916. 35:06on the world to act,
917. 35:08declaring AI an existential threat to
918. 35:10humanity
919. 35:12on a par with the risk of nuclear war.
920. 35:16By the way, we don't get to not choose.
921. 35:19If we don't want that future, we have to
922. 35:21pick a better one. Otherwise,
923. 35:24we could wake up one day in a world
924. 35:26where superintelligent AI
925. 35:28has taken over.
926. 35:33A scenario we go through step by step in
927. 35:35this video about how an AI escapes the
928. 35:37lab.
929. 35:38Hey, I'm Drew, and I really loved making
930. 35:41this one for you guys.
931. 35:43Thank you.