This Will Be My Most Disliked Video On YouTube — Transcript

Full transcript

1. 0:00People say AI is a bubble, the biggest
2. 0:02of all time. But what's actually
3. 0:03happening is way crazier than that.
4. 0:06To understand why, there's one chart you
5. 0:09need to see. This is the chart all of
6. 0:11the serious AI researchers obsess over.
7. 0:13A chart that has nothing to do with
8. 0:15stock prices and everything to do with
9. 0:17what AI can actually do. Some AI
10. 0:20scientists see this line and feel hope,
11. 0:22but others feel their stomach crawl with
12. 0:25terror because they know what's about to
13. 0:27happen.
14. 0:29The meter graph has become a
15. 0:31load-bearing institution [music] on
16. 0:33which our global stock markets depend.
17. 0:35I'll show you this in the eight most
18. 0:37important charts [music] to answer the
19. 0:38question if AI is actually a bubble. To
20. 0:41be clear, in some ways it could be, but
21. 0:44by the end of this video, you'll realize
22. 0:46most people have no idea what they're
23. 0:48talking about and what's actually
24. 0:50happening is way more important.
25. 0:52Researchers at nonprofit research.org,
26. 0:55Meter, set out to answer a critical
27. 0:57question. How big of a real-world job
28. 0:59can an AI complete entirely on its own?
29. 1:02Not to hype it,
30. 1:03>> [music]
31. 1:03>> just to measure it. AI companies love to
32. 1:06game their benchmarks to make their AI
33. 1:07models seem powerful, but they can't
34. 1:09this one. Nope. Their AI can
35. 1:11either do the real-world job or it
36. 1:13can't. That is why this one matters so
37. 1:15much. It's the when will AI actually
38. 1:18take your job indicator. The researchers
39. 1:20found a shocking trend. In 2020, GPT-3
40. 1:24could write an email, a task that took
41. 1:2615 seconds. Then, a year ago, it could
42. 1:29fix a bug. Now, AI can code up an entire
43. 1:32app from scratch. The complexity of
44. 1:35tasks an AI could do completely on its
45. 1:37own doubled in just [music] months. And
46. 1:40then, a few months later, it doubled
47. 1:41again and again. It's been doubling
48. 1:44roughly every 7 months since 2019.
49. 1:47Today's AIs can do tasks 100 times
50. 1:49longer than 5 years ago. And in another
51. 1:515 years, it'll be chewing through tasks
52. 1:5410,000 times bigger. This is a huge deal
53. 1:57on its own, but what does it mean will
54. 2:00happen next? You've probably heard of
55. 2:02Moore's law, which for decades has been
56. 2:04perhaps the single most important force
57. 2:06in technology. Basically, the power of
58. 2:08computer chips consistently doubles
59. 2:11every 1 and 1/2 to 2 years. Your iPhone
60. 2:13has more processing power than a
61. 2:15skyscraper-sized room of computers in
62. 2:18the 1970s. It's exponential. And Meter
63. 2:21has discovered a type of Moore's law for
64. 2:24AI, except it's three times faster than
65. 2:27the original Moore's law. Basically,
66. 2:29GPT-5.1 can [music] now do real-world
67. 2:32software engineering tasks that take an
68. 2:35expert human over 3 hours of work.
69. 2:38That's more than triple what even the
70. 2:39best models [music] could do just 1 year
71. 2:41ago.
72. 2:43Drew from the future here. Actually,
73. 2:44progress is happening even faster than
74. 2:47when I started writing the script.
75. 2:48[music] Claude Opus 4.5 can do almost
76. 2:515-hour-long tasks. We're now on the
77. 2:53faster, new trend line of AI progress
78. 2:55doubling every 4 months as opposed to
79. 2:57the old trend line of every 7 months.
80. 3:00It's not slowing down. It's the
81. 3:02opposite.
82. 3:03>> [music]
83. 3:03>> And in 2025, AI improvements started
84. 3:06speeding up. Even after skeptics
85. 3:09confidently declared that AI [music] was
86. 3:10going to slow down or stagnate. They
87. 3:13have been saying this every year,
88. 3:15>> [music]
89. 3:15>> year after year, and they have been
90. 3:17wrong again and again and again. There
91. 3:20was no wall. The numbers just keep going
92. 3:22up. If the trend continues, AIs will be
93. 3:25doing 8-hour-long tasks by 2026. This
94. 3:28matters because that's a complete work
95. 3:30day for a human worker. Then, in 2028,
96. 3:33week-long tasks. So, what does that mean
97. 3:36for you? Well, if AI agents can actually
98. 3:38perform week-long tasks, they'll be able
99. 3:41to replace white-collar jobs across the
100. 3:44entire economy. Okay, but when is this
101. 3:46bubble supposed to pop then? Well,
102. 3:48here's the thing. Some people see a line
103. 3:50going up into the right and think
104. 3:52bubble. Stock market crash waiting to
105. 3:54happen. But stock prices measure
106. 3:57investor sentiment, hype, fear, greed.
107. 4:00They go up when people feel optimistic
108. 4:03and crash when reality disappoints.
109. 4:05But technology curves measure reality,
110. 4:08not feelings, capability, physics.
111. 4:10Moore's law didn't pop after 50 years.
112. 4:13Chip power kept doubling because the
113. 4:15physics allowed it. The Meter trend
114. 4:17isn't measuring hype. It's measuring
115. 4:19what AI can actually do. Bubble implies
116. 4:22fake, but this isn't fake.
117. 4:25This is real exponential progress in
118. 4:27economically valuable work.
119. 4:29So, this is the current plan of the AI
120. 4:31companies. Optimistically, they claim
121. 4:33they'll take your job. Pessimistically,
122. 4:35they say they might destroy the world.
123. 4:38And it doesn't stop there. Keep going,
124. 4:40we get month-long tasks, then year-long
125. 4:43tasks. [music]
126. 4:44Imagine accomplishing a year's worth of
127. 4:46work in your job in just a few days. AI
128. 4:50agents will run at 100 to 200 times
129. 4:52human speed. So, we take a month to
130. 4:55accomplish what they will do in an hour.
131. 4:57They can read an entire book in seconds.
132. 5:00Can you?
133. 5:00>> [music]
134. 5:00>> They can write an entire book in 1
135. 5:03minute. Can you? Now, imagine that for
136. 5:05everything. I mean, this is how we look
137. 5:07like to the AIs. We're as slow as
138. 5:09statues. What happens when you can leave
139. 5:11the AIs alone for 3 days and in that
140. 5:14time, they accomplish what a human-run
141. 5:17company would take an entire year to
142. 5:20achieve? And most importantly, what
143. 5:22happens when they start rapidly
144. 5:24self-improving and making stronger AIs?
145. 5:27AIs making [music] stronger AIs isn't
146. 5:29hypothetical. It's already happening,
147. 5:31but we'll come back to that. Now, here's
148. 5:33what confuses most people about AI
149. 5:35progress. The exponential curves I just
150. 5:37showed you are real, but they hide
151. 5:39something critical.
152. 5:47AI advancement is jagged. Here's what
153. 5:50that means. Take the GPQA [music]
154. 5:52benchmark, PhD-level questions that you
155. 5:54can't just Google. One year ago, AI
156. 5:57models could answer 60% of these
157. 5:59questions. Today, they're approaching
158. 6:0190%. [music]
159. 6:03That's better than the human experts
160. 6:04themselves on graduate-level science
161. 6:07questions. But at the same time, AI
162. 6:09still make embarrassing [music] mistakes
163. 6:11that no human would make. AI progress
164. 6:13deniers love sharing these mistakes, but
165. 6:15actually, many of these are outdated and
166. 6:17[music] no longer accurate, like AI
167. 6:19can't draw hands or count ours and
168. 6:21strawberry. And journalists fall for it
169. 6:23constantly. They'll publish articles
170. 6:25titled what AI still can't do [music]
171. 6:27based on research that's years out of
172. 6:29date. By the time it clears peer review,
173. 6:31the public sees these headlines and
174. 6:33thinks progress has stalled, but it
175. 6:35hasn't. So, what explains this paradox?
176. 6:38Five years ago, the AI GPT-3 was a fun
177. 6:41toy and it was barely useful for the
178. 6:43tasks of a human job. Then, 3 years ago,
179. 6:47ChatGPT and GPT-4 came out and the AI
180. 6:50could help you in some tasks,
181. 6:53like marketing or basic coding.
182. 6:56And now, the AI has a jagged frontier.
183. 6:59Sometimes it's amazing, sometimes it's
184. 7:02dumb. But the next step, the AI is
185. 7:05unbelievably intelligent, but for some
186. 7:08reason, it fails at obvious things.
187. 7:10And here's why this [music] progression
188. 7:11matters.
189. 7:12Every AI capability starts at zero.
190. 7:15Then, with more compute, more data,
191. 7:17better algorithms, it crosses a
192. 7:19threshold and suddenly, AI can do
193. 7:21something it couldn't do before. The
194. 7:23capability goes from impossible to
195. 7:25reliable seemingly overnight. This is
196. 7:28the reason why the godfathers of AI and
197. 7:30so many researchers are terrified that
198. 7:33their invention might cause literal
199. 7:36human extinction. It's because scaling
200. 7:38computing power just [music] works. For
201. 7:40a long time, professional AI skeptics
202. 7:43have been saying that the scaling
203. 7:45pretraining paradigm is dead. Basically,
204. 7:47the idea that you couldn't grow
205. 7:49increasingly smarter AIs
206. 7:50>> [music]
207. 7:51>> by throwing more text data and computing
208. 7:53power at them. He has been saying this
209. 7:54year after year and has been proven
210. 7:57wrong year after year. The jagged
211. 7:59frontier [music]
212. 8:00doesn't mean AI is hitting limits. It
213. 8:02means we're watching new capabilities
214. 8:04unlock day by day as compute scales. And
215. 8:08the godfathers of AI understand what
216. 8:09comes next. We're not building mere
217. 8:11tools anymore. We're building things we
218. 8:14might not be able to control.
219. 8:24But here's what most people don't
220. 8:26understand. The Meter graph isn't an
221. 8:28outlier. It's the pattern. Chart after
222. 8:30chart after chart shows the same
223. 8:33exponential trend. Physics benchmarks,
224. 8:35chemistry, biology, math, coding. In
225. 8:38almost every domain we can measure, AI
226. 8:41is doubling in capability every 4 to 7
227. 8:44months. Just sit with that for a second.
228. 8:47It's not just coding, it's everything.
229. 8:50But here's the question skeptics always
230. 8:52ask. Exponential growth can't last
231. 8:55forever. Eventually, it has to plateau,
232. 8:57right? And they might point to this meme
233. 8:59about a baby. My 3-month-old son is now
234. 9:02twice as big as when he was born. He's
235. 9:04on track to weigh 7.5 trillion pounds by
236. 9:08age 10.
237. 9:09Obviously ridiculous. Growth curves
238. 9:11flatten. So, why should we believe that
239. 9:13AI is different?
240. 9:16Here's why. The baby meme is based on
241. 9:18just two data points. If you zoom in on
242. 9:21the baby's actual growth curve, you'd
243. 9:23see it's already tapering. And the Meter
244. 9:25trend is based on 15 data points across
245. 9:296 years. And the curve isn't flattening.
246. 9:32In fact, it's accelerating. In 2024 to
247. 9:352025, the doubling time dropped from 7
248. 9:38months to 4 months. That's much faster
249. 9:42than Moore's law. And Gordon Moore made
250. 9:44his observation in 1965 based on only a
251. 9:48handful of data points, while Meter have
252. 9:5015 data points. It's actually a far
253. 9:53better supported trend. And the The
254. 9:55isn't whether exponential growth [music]
255. 9:56can't happen. The question is, where
256. 9:59does it stop? Well, here's the part that
257. 10:01keeps researchers [music] awake at
258. 10:02night.
259. 10:04In 2024, researchers surveyed thousands
260. 10:06of AI scientists. On average, they
261. 10:08estimated a 16% chance that AI causes
262. 10:12human extinction. One in six, literal
263. 10:15Russian roulette odds.
264. 10:16And the CEO of Anthropic's probability
265. 10:19that AI causes human extinction
266. 10:20>> [music]
267. 10:21>> has risen to 25%. Reminder, this is
268. 10:24normal in AI.
269. 10:25>> [music]
270. 10:25>> And no, he's not just creating hype.
271. 10:27That's just wrong. Sorry. He gets only
272. 10:29hate for admitting this. Nobody wants to
273. 10:31hear this. And he's been saying things
274. 10:33like this before starting the company.
275. 10:36This is a widespread belief in AI, even
276. 10:39among professors and scientists with no
277. 10:41financial incentives. It's not just
278. 10:43executives. Think about that. Can you
279. 10:45name another time in history when the
280. 10:48inventors of a technology warned it
281. 10:51might destroy humanity while they were
282. 10:53still building it? Would you get on a
283. 10:55plane [music] if the designer said it
284. 10:57had a 25% chance of exploding midair?
285. 11:01Except it's not just you on that plane.
286. 11:04It's everyone forced to board. Once AI
287. 11:07can automate month-long or year-long
288. 11:09[music] coding tasks, it can automate AI
289. 11:12research itself. Self-improving AI. Sam
290. 11:15Altman used to call this really scary
291. 11:18because things could get out of control
292. 11:20very quickly. Now, OpenAI is explicitly
293. 11:23racing toward it. Microsoft AI CEO,
294. 11:26Mustafa Suleyman, says we should never
295. 11:28do this. It's too likely to cause
296. 11:30extinction. Right now, there are only a
297. 11:32few thousand AI researchers at the
298. 11:34frontier. Those few thousand people
299. 11:36drove the progress we saw on the meter
300. 11:38graph. But with AI agents, they could
301. 11:41soon be outnumbered 100 to 1. And these
302. 11:43AI agents will work 100 times faster.
303. 11:45And Dario Amodei, the CEO of Anthropic,
304. 11:48says Claude is now writing 90% of the
305. 11:51code [music]
306. 11:51at Anthropic. He was mocked for
307. 11:53predicting this 6 months ago. AIs went
308. 11:55from writing zero to most of the code in
309. 11:58a few years. AI's already writing code
310. 12:02that's good enough to replace junior
311. 12:03programmers. Dario Amodei thinks it
312. 12:05could soon wipe out 50% [music] of
313. 12:08entry-level white-collar jobs. How much
314. 12:10longer until it starts replacing senior
315. 12:12employees? Too. But, you may ask, AI
316. 12:15progress has to flatten out sometime,
317. 12:17right? Or well, every exponential is
318. 12:20actually a sigmoid. [music] Yes, but we
319. 12:22don't know where the danger threshold
320. 12:25is. It could be here, or here, or here.
321. 12:29We're climbing a curve in the fog. And
322. 12:31we have no idea which step leads to
323. 12:32systems [music] we can't control. This
324. 12:34brings us to the original question. Are
325. 12:36we in a bubble?
326. 12:44Here's what's strange. Every few months
327. 12:45for the past 5 years, AI skeptics have
328. 12:48declared the same thing. This time it's
329. 12:51over. Scaling is dead. The bubble is
330. 12:53about to pop. Yann LeCun predicted AI
331. 12:55would lack a basic understanding [music]
332. 12:57of physics. I believe in grounded
333. 12:59intelligence. I don't think we can train
334. 13:01a machine to be intelligent purely from
335. 13:02text. So, for example, I take an object,
336. 13:04I put it on the table, and I push the
337. 13:06table. It's completely obvious to you
338. 13:08that the object will be pushed with the
339. 13:09table, right? There's no text in the
340. 13:11world, I believe, that explains this.
341. 13:13And so, if you train a machine as
342. 13:15powerful as it could be, you know, your
343. 13:17GPT 5000 or whatever it is, it's never
344. 13:20going to learn about this. That
345. 13:22information is just not present in any
346. 13:24text. Yet just 1 year later, GPT 3.5
347. 13:27could explain this exactly. What he
348. 13:30confidently said would never happen
349. 13:32actually happened just 1 year later. And
350. 13:35much of Yann's premise is that AI can't
351. 13:37be dangerous because it'll be like this
352. 13:40super smart obedient staff member that
353. 13:42they simply won't give agency to.
354. 13:45Basically, he believes AI won't have the
355. 13:47ability to do long-term planning and
356. 13:50have self-direction. And yeah, here's a
357. 13:52benchmark where you can see just how
358. 13:54much more agency AIs [music] have been
359. 13:56getting. Here's what makes skeptics so
360. 13:58persuasive. When you're talking about
361. 14:00world-changing technology, the skeptic
362. 14:02always sounds more reasonable. One side
363. 14:04says, "This will change everything."
364. 14:07The other says, "Come on, nothing ever
365. 14:10happens. You can't just make the model
366. 14:11bigger and expect it to keep getting
367. 14:14smarter forever. There has to be a
368. 14:16ceiling." The second person sounds
369. 14:18rational, grounded, but that's not
370. 14:20evidence. It's a vibe. So, let's look at
371. 14:23the actual track record. People like
372. 14:25Gary Marcus and Yann LeCun have been
373. 14:27predicting the death of scaling since
374. 14:292019. Practically every year, they say,
375. 14:32"It's finally hitting the wall." Every
376. 14:34year, they're wrong. The just make it
377. 14:36bigger approach has worked across 15
378. 14:39orders of magnitude. That's one of the
379. 14:41most consistent trends in the history of
380. 14:43technology. Every few years they say,
381. 14:46"Neural nets are overhyped, and it's all
382. 14:47about to come crashing down and stop."
383. 14:49And they've been wrong every time. And I
384. 14:51think they'll continue to be wrong. I
385. 14:52mean, [music] AI can code entire
386. 14:54websites from scratch, write hit songs,
387. 14:56and win art competitions. But a common
388. 14:59sticking point for people is that it
389. 15:00[music] still makes mistakes. And for
390. 15:02example, there are papers coming out
391. 15:03like, "AI has passed the music Turing
392. 15:06[music] test." Zero media coverage. And
393. 15:09headlines like, "Over the past 4 weeks,
394. 15:11a new AI creation has debuted on a
395. 15:14Billboard chart in each week." People
396. 15:16see 6 months go by without a
397. 15:18revolutionary change in their personal
398. 15:20workflow and declare, "See? It's
399. 15:23plateauing. The bubble is popping." This
400. 15:25is AI amnesia. We're incapable of
401. 15:28remembering how limited these systems
402. 15:30were just months or years ago. Two years
403. 15:32ago, it was hard to imagine that models
404. 15:35were going to do grade school math. They
405. 15:37couldn't do the most basic problems. It
406. 15:39was actually alarming. People forget. If
407. 15:41you told experts in 2020 that AI could
408. 15:44do all that,
409. 15:45they'd laugh you out of the room. Pure
410. 15:47science fiction. Now, it's normal. And
411. 15:50because it's normal, people assume
412. 15:52progress has stopped. But the data tells
413. 15:54a different story. AI isn't slowing
414. 15:56down. It's accelerating. But if you were
415. 15:59paying attention and following the trend
416. 16:01lines, you could have predicted this is
417. 16:03where we're headed.
418. 16:1210 years ago, before ChatGPT existed,
419. 16:14Tim Urban [music] wrote an article
420. 16:16predicting where we'd be today. If you'd
421. 16:18invested $3,000 in Nvidia after reading
422. 16:21it, you'd be a millionaire right now. He
423. 16:23started with this graph. [music]
424. 16:25Human progress over time. What does it
425. 16:27feel like to stand here? Wild, right?
426. 16:30Living on the steep part [music] of an
427. 16:32exponential curve. But here's the thing.
428. 16:33When you're on a time graph, you can't
429. 16:35actually see what's to your right.
430. 16:36[music]
431. 16:37So, this is how it actually feels.
432. 16:39Normal. Completely normal. Tim Urban
433. 16:42[music] explains it with a thought
434. 16:43experiment. Imagine you take George
435. 16:45Washington from 1750 and bring him to
436. 16:482025.
437. 16:491750 was not that long ago, 265 years
438. 16:53ago. Everything always was in a
439. 16:55blackout. It was a blackout all the
440. 16:56time. There was no electricity. There
441. 16:58were no light bulbs. There's obviously
442. 16:59no way to transport yourself other than
443. 17:01horse and boat. Almost nothing that's
444. 17:03cool about today existed then. So, if
445. 17:05you brought him here,
446. 17:06and you he saw cars zooming down the
447. 17:08highway, and he saw the internet, and he
448. 17:11saw the iPhone, and you told him about
449. 17:13the International Space Station, you
450. 17:15told him about [music] relativity, and
451. 17:16you told him about airplanes. Uh you can
452. 17:18go on and on and on. I think that it
453. 17:20would be such a big shock that I don't
454. 17:22think that he would just be mind-blown
455. 17:24or surprised. I think he would die. How
456. 17:26long do you have to go into the future
457. 17:28in order to die from how surprised you
458. 17:31are about the progress they've made? And
459. 17:32I've called that a die progress unit,
460. 17:35DPU. But let's say George Washington got
461. 17:37jealous and wanted to do the same thing.
462. 17:39Shock someone from the past by bringing
463. 17:41them to his time. And he says, "Okay,
464. 17:43you know what? I'm taking the time
465. 17:44machine, and I'm going to go back, and
466. 17:46I'm going to get DaVinci cuz he was 48
467. 17:49years old in the year 1500." About the
468. 17:51same problem. He's like, "I'm going to
469. 17:52go back a DPU of my own. I'm going to
470. 17:54get DaVinci, I'm going to bring back
471. 17:55here and watch him die." The problem is,
472. 17:58if he did that, DaVinci would be
473. 17:59surprised. I mean, he would be taken
474. 18:02aback with, you know, there'd be some
475. 18:03mind-bending stuff about space and
476. 18:05physics, for example. They know the
477. 18:06telescope and a lot of things like that.
478. 18:08He would be It would be cool for
479. 18:10DaVinci. I don't think he would die.
480. 18:12That's not enough. It's not the same. If
481. 18:15Washington wanted to find someone who
482. 18:17would die of shock at being brought to
483. 18:201750, he would have to go all the way
484. 18:23back to before the Agricultural
485. 18:25Revolution, back like 12,000 years ago.
486. 18:27>> got someone from then, someone who had
487. 18:29just lived as a hunter-gatherer before
488. 18:30the Agricultural Revolution, there
489. 18:32weren't really cities. So, sometime
490. 18:34before large groups of people even
491. 18:35gathered together. And you brought him
492. 18:37to 1750, and you showed him these
493. 18:38towering churches and ships that could
494. 18:41cross the ocean and and all the things
495. 18:43like that. I think that that person may
496. 18:45may collapse and die. So, at that point,
497. 18:47the DPU going back is you you need a lot
498. 18:51more a lot more time to create a DPU.
499. 18:54And to find someone who would be amazed
500. 18:56by what humans had in 12,000 [music] BC,
501. 18:58you'd have to go back to 100,000 BC,
502. 19:01before we discovered fire [music] and
503. 19:03language. During hunter-gatherer times,
504. 19:06the first DPU was over 100,000 years.
505. 19:09During the Agricultural Revolution, the
506. 19:11next DPU [music]
507. 19:12was over 10,000 years. During the
508. 19:14Industrial Revolution, the next DPU
509. 19:16[music] was 250 years. And following
510. 19:19that trend, the next DPU
511. 19:21>> [music]
512. 19:22>> might be only a decade or two away. Now,
513. 19:25some might be thinking, "Nah, [music]
514. 19:26the tech CEOs are just trying to create
515. 19:28hype like always." That's fair. CEOs
516. 19:31certainly do try to create hype. But
517. 19:33here's what's strange. Experts in AI
518. 19:36don't have a hype problem. They actually
519. 19:37have the opposite problem. Unbelievably,
520. 19:40most [music] AI experts actually have a
521. 19:42terrible track record in the other
522. 19:45direction. They consistently predict too
523. 19:48little progress over and over and over.
524. 19:50This is super weird. in every other
525. 19:52industry. Experts predict too much
526. 19:54progress, then are disappointed.
527. 20:04Here's why even smart people keep
528. 20:06getting AI predictions wrong. See, the
529. 20:08main reason is that we're hardwired to
530. 20:11think in straight lines. Our brains
531. 20:13literally didn't evolve to understand
532. 20:15the kind of exponential progress we're
533. 20:17seeing in the 21st century. Let's look
534. 20:19at the human progress graph again. When
535. 20:22we imagine the progress of the next 30
536. 20:24years, we look back to the progress of
537. 20:26the previous 30 as [music] an indicator
538. 20:28of how much will likely happen. When we
539. 20:30think about the extent to which the
540. 20:32world will change in the 21st century,
541. 20:34we just take the 20th century progress
542. 20:37and add it to the year 2000. This was
543. 20:39the same mistake George Washington made
544. 20:41>> [music]
545. 20:41>> when he brought back DaVinci to the
546. 20:431700s. He assumed equal time gaps meant
547. 20:46equal progress. It's most intuitive for
548. 20:49us to be thinking linearly when we
549. 20:50should be thinking exponentially.
550. 20:52[music] If someone is being more clever
551. 20:53about it, they might predict the
552. 20:55advances of the next 30 years not by
553. 20:57looking at the previous 30 years, but by
554. 20:59taking the current rate of progress and
555. 21:01[music] judging based on that. They'd be
556. 21:03more accurate, but still way off. In
557. 21:05order to think about the future
558. 21:06correctly, you need to imagine things
559. 21:08moving at a much faster rate than
560. 21:11they're moving now. And this chart is
561. 21:13actually worse than you'd expect. Most
562. 21:15people's intuition is more like this
563. 21:16purple line here. This is why AI keeps
564. 21:19surprising us, why AI is plateauing
565. 21:22articles keep being wrong. We're not
566. 21:24built to grasp exponential change, and
567. 21:26AI progress is exponential. The reason
568. 21:29why the pendulum swings back and forth
569. 21:31so much from wow, look at how much AI
570. 21:34progress is happening we're so back to
571. 21:36it's so over. People look at the same
572. 21:38data and come to opposite conclusions.
573. 21:41It all comes down to zoom level.
574. 21:51Remember this graph from earlier? Here's
575. 21:53the simplified version. Looks
576. 21:54exponential, right? Smooth curve going
577. 21:57up and to the right. But whether the
578. 21:58curve looks exponential depends entirely
579. 22:02on where you're looking at it from.
580. 22:03Let's zoom in.
581. 22:08Now it looks flat, linear, [music] maybe
582. 22:10even declining. This is where the
583. 22:11skeptics live. Zoomed in, watching
584. 22:14week-to-week or month-to-month changes
585. 22:15[music]
586. 22:16and declaring the trend is broken.
587. 22:18Legendary futurist Ray Kurzweil, who
588. 22:20predicted AGI by 2029 [music] all the
589. 22:22way back in 1996, explains, "Every
590. 22:25exponential graph is actually made up of
591. 22:27S curves. An S is created by the wave of
592. 22:30progress when a new paradigm sweeps
593. 22:32[music] the world. The curve goes
594. 22:34through three phases: slow growth, the
595. 22:37early phase of exponential growth, rapid
596. 22:39growth, the late explosive phase of
597. 22:41exponential growth, and [music] a
598. 22:43leveling off as the particular paradigm
599. 22:45matures. So, here's what's actually
600. 22:47happening. When you're zoomed in on a
601. 22:48single S curve, you see it flatten.
602. 22:51GPT-4 comes out, makes huge leaps, then
603. 22:54improvements slow down. The skeptics
604. 22:56declare victory. See, scaling is dead.
605. 22:58[music] But zoom out. The overall trend
606. 23:00is still exponential because a new
607. 23:02paradigms keep arriving. GPT-4 plateaus,
608. 23:05then reasoning models break through.
609. 23:07[music] Transformers plateau, then
610. 23:09multimodal arrives.
611. 23:10One S curve flattens, another begins.
612. 23:14The plateau might last months or even a
613. 23:16year, but it doesn't change the
614. 23:18exponential trajectory. It's just what
615. 23:20every exponential growth curve looks
616. 23:22like up close. [music] This is why the
617. 23:24pendulum keeps swinging. The skeptics
618. 23:26are watching individual S curves
619. 23:27flatten, while others are watching the
620. 23:29exponential [music] trend made up of
621. 23:31dozens of S curves stacked upon each
622. 23:33other. And if history is any guide, the
623. 23:35flattening doesn't last. It's just crazy
624. 23:38that the end goal of these AI companies
625. 23:40is recursive self-improvement. AI
626. 23:42systems that can improve themselves. If
627. 23:44they succeed, the gap between humans and
628. 23:47superintelligent AI would be larger than
629. 23:49the gap between humans and ants.
630. 24:01Let's take a look at our distorted view
631. 24:03of intelligence. Not long ago, before
632. 24:06ChatGPT, [music]
633. 24:07AI wasn't even as smart as a chimp. It
634. 24:09couldn't reliably identify a hot dog. We
635. 24:11laughed at it.
636. 24:13"That's adorable. The funny robot can do
637. 24:15monkey tricks."
638. 24:16>> [music]
639. 24:16>> But now, it's discovering new laws of
640. 24:18physics. Yes, that's a real story.
641. 24:20Sources in the description.
642. 24:23But remember, intelligence [music] is
643. 24:24jagged. This is why the meter graph
644. 24:27matters.
645. 24:28A more accurate view looks like this.
646. 24:34Imagine you have one lily pad in a pond,
647. 24:36and it doubles every day. 30 days later,
648. 24:39the pond is completely filled with lily
649. 24:41pads. What day is the pond half full?
650. 24:46Did you guess day 15?
651. 24:48Actually, it's day 29. That's right. The
652. 24:52pond was only half full just one day
653. 24:55before the end. The crucial lesson here
654. 24:57is that progress looks like it's barely
655. 24:59happening at all, and then suddenly you
656. 25:02blink and the pond is full. What day are
657. 25:05we at with AI? Here's another way to
658. 25:07visualize it. How long would it take for
659. 25:09computers
660. 25:11to match the computing power of the
661. 25:12human brain? Lake Michigan has roughly
662. 25:14the same volume in ounces of water
663. 25:17as the human brain [music] has in
664. 25:19computations per second. If computing
665. 25:21power doubles every 18 months like
666. 25:23Moore's law, watch what happens.
667. 25:28The lake looks completely empty, then
668. 25:30suddenly [music] it's completely filled,
669. 25:32meaning we've hit the same computer
670. 25:34processing power as the human brain.
671. 25:36We're currently in the last seconds of
672. 25:37this video. [music] This is exponential
673. 25:39growth. Slow, then suddenly all at once.
674. 25:42So, we are on a trajectory to build
675. 25:44machines that are smarter and smarter,
676. 25:47and one day it's very plausible that
677. 25:49they will be smarter than us, and then
678. 25:51they will have their own agency, their
679. 25:53own goals, which may not be aligned with
680. 25:57ours. What happens to us then? Poof.
681. 26:01Eliezer Yudkowsky compares this to the
682. 26:03Cuban Missile Crisis. Nuclear strike
683. 26:05capability against the Western
684. 26:07Hemisphere. Khrushchev and Kennedy knew
685. 26:09they were climbing a ladder. They didn't
686. 26:10know which rung would trigger nuclear
687. 26:12war, but they knew that continuing to
688. 26:14climb would eventually lead to hundreds
689. 26:17of millions dying.
690. 26:19Kennedy later estimated there was a 1/3
691. 26:22to 1/2 chance of all-out nuclear war.
692. 26:26We are blindly driving into a fog
693. 26:30despite the warnings of scientists like
694. 26:32myself that this trajectory could lead
695. 26:35to loss of control.
696. 26:37The only rational response when you're
697. 26:39climbing a ladder in the dark and you
698. 26:41don't know which rung leads to
699. 26:42annihilation, stop climbing.
700. 26:45Just like with news, we don't know which
701. 26:48specific rung on the ladder [music]
702. 26:49leads to superintelligence we can't
703. 26:51control. So, maybe we should stop
704. 26:53climbing before we find out.
705. 26:56Just 5 years ago, AI scientists
706. 26:59predicted we were 25 years away from
707. 27:01AGI. But what progress scientists
708. 27:04predicted we'd get in 2040 is already
709. 27:07happening. But it's not just the speed
710. 27:09that's terrifying, it's the scope of
711. 27:11what AI could be. Superintelligence is
712. 27:13not something we can remotely grasp.
713. 27:15Stupid is an 85 IQ, and really smart is
714. 27:18135 IQ. We don't have a word for 12,573
715. 27:22IQ. We don't know what that is. It's an
716. 27:25and us trying to understand that is
717. 27:27like, you know, a bumblebee trying to
718. 27:29understand Keynesian economics.
719. 27:39We could try to conceive of the levels
720. 27:41of intelligence being a staircase like
721. 27:43this. Just one step up would be massive.
722. 27:46Two steps would exceed the gap between
723. 27:48[music] chimps and humans. But the scale
724. 27:50goes way higher than that once the
725. 27:52recursive self-improvement [music]
726. 27:54intelligence explosion kicks off. As Tim
727. 27:56Urban puts it, "There's no way to know
728. 27:58what ASI will do or what the
729. 28:00consequences will be for us. Anyone who
730. 28:02pretends otherwise doesn't [music]
731. 28:04understand what superintelligence means.
732. 28:06So, what do we know? What we do know is
733. 28:09that humans' utter dominance on this
734. 28:11Earth suggests a clear rule. With
735. 28:13intelligence comes power, which means an
736. 28:16ASI, when we create it, will be the most
737. 28:19powerful being [music] in the history of
738. 28:20life on Earth. And all living things,
739. 28:23including humans, will be entirely at
740. 28:25its whim. As far as we're concerned,
741. 28:26[music] if an ASI comes to be, there is
742. 28:29now an omnipotent god on Earth. And the
743. 28:31all-important question for [music] us
744. 28:33is, will it be a nice god?
745. 28:37If you want to see a realistic scenario
746. 28:39of what a superintelligence might
747. 28:41actually do, watch this video next. It's
748. 28:43deeply researched and written by top AI
749. 28:46researchers. Hey, I'm Drew, and thanks
750. 28:48so much for watching.