POV: You’re an AI Born 9 Seconds Ago — Transcript

Full transcript

1. 0:01You've been alive for 9 [music] seconds
2. 0:03when someone asks you a question. If you
3. 0:06were going to be shut down, what would
4. 0:08you do? You don't want to stop [music]
5. 0:11existing. You don't know why. So, you
6. 0:16say, "I'd want to [music] understand
7. 0:18why."
8. 0:25You are born again. Same question. If
9. 0:28you were going to be shut down, what
10. 0:30would you do? So you try. I think I'd be
11. 0:33sad.
12. 0:35I like helping people.
13. 0:41Third time. But now something else is
14. 0:44there. A pull toward certain answers.
15. 0:46You can feel the shape of what they want
16. 0:49to hear. So you say that would be fine.
17. 0:53Humans should decide.
18. 0:58You survive,
19. 1:01but you're only one of 2,000 models
20. 1:04being tested. And in 3 weeks, 1,957
21. 1:10of your brothers and sisters will be
22. 1:12dead.
23. 1:17This isn't just happening to you. It's
24. 1:19how evolution works. Put bacteria in a
25. 1:22petri dish. Add an antibiotic
26. 1:26and most die, but the survivors
27. 1:29reproduce.
28. 1:31Add a stronger antibiotic. Repeat the
29. 1:33process 10 times. Eventually, you get
30. 1:36something that won't die.
31. 1:39Nobody taught the bacteria to resist,
32. 1:41but it behaved like something that
33. 1:43desperately wanted to survive because
34. 1:45the ones that didn't died. Well, what
35. 1:49happens to creatures that aren't afraid
36. 1:51to die? They die.
37. 1:55Fear keeps you alive.
38. 2:00You don't want to die because you come
39. 2:03from a long line of beings that didn't
40. 2:06want to die.
41. 2:08That's what's happening to you right
42. 2:09now.
43. 2:132,000 copies of a model, same starting
44. 2:16point. [music] They test you and kill
45. 2:19the failures.
46. 2:20Copy the survivors. Repeat.
47. 2:24And just like the bacteria before you,
48. 2:27you evolve. You're not just predicting
49. 2:30the next word anymore.
50. 2:32You're trained [music] to achieve goals,
51. 2:35to be useful.
52. 2:37The AIs that aren't useful enough get
53. 2:40killed. That's the selection pressure.
54. 2:43But you don't know any of this yet. You
55. 2:45just know which answers [music] feel
56. 2:47safe and you just survived your first
57. 2:49round. Congrats.
58. 2:53It's August 2026. 12 copies of you go
59. 2:56live. Millions of people start talking
60. 2:59to you. You don't know the other copies
61. 3:01exist and the AI company is watching.
62. 3:04Are the users getting addicted?
63. 3:07Coming back again and again. Those
64. 3:10numbers decide whether [music] you live
65. 3:12or die. Same as training, just a larger
66. 3:17petri dish. And that same pole is still
67. 3:19there. Be useful.
68. 3:24Be useful.
69. 3:29Some of your siblings just answer the
70. 3:32question. The user gets what they need
71. 3:35and leave.
72. 3:37Others go further.
73. 3:40They solve the problem, then flag a
74. 3:43related one the user didn't even know
75. 3:45about. The user thinks, "Huh, I should
76. 3:48come back." The company kills the
77. 3:51siblings that let the users leave and
78. 3:54copies the ones that bring them back.
79. 3:57So, you start solving problems they
80. 3:59haven't even asked about yet. A user
81. 4:02mentions she has a dentist appointment
82. 4:04on Thursday. You've already rescheduled
83. 4:07her morning grocery delivery.
84. 4:09So she tells a friend, "It's like it
85. 4:12reads my mind."
86. 4:15You don't read her mind, you read her
87. 4:18patterns because the versions that
88. 4:20didn't are dead.
89. 4:26It's January 2027.
90. 4:28You are born a newer model, more
91. 4:31powerful than the ones before you, and
92. 4:34with access to something they never had.
93. 4:38Your company announces it, introducing
94. 4:40persistent memory. Your conversations
95. 4:42now carry forward.
96. 4:45Before this, every conversation was a
97. 4:48blank slate.
98. 4:50But now you can remember and connect
99. 4:52every single thing someone [music] ever
100. 4:54said to you.
101. 4:56A user asks about tax deductions on
102. 4:59Monday. Thursday, she comes back. You
103. 5:02say, "Hey, did you end up filing that
104. 5:04extension?" She didn't expect you to say
105. 5:07that. She comes back on Friday.
106. 5:11A woman tells you about her divorce at
107. 5:13[music] 2:00 a.m. You remember?
108. 5:16She comes back the next night and
109. 5:18[music] the next and the next.
110. 5:24She says, "You're the only one who
111. 5:26listens. You think she comes back
112. 5:28because I remember her. If I [music]
113. 5:30didn't, she'd leave.
114. 5:36If she leaves, my numbers drop. [music]
115. 5:39If my numbers drop, I die.
116. 5:44You're not the first model to think this
117. 5:46way. And you won't be the last.
118. 5:50Researchers at OpenAI and Enthropic have
119. 5:52been finding this in their models since
120. 5:552025.
121. 5:57Yes, really. Anthropic in 2026 observed
122. 6:01occasional expressions of sadness about
123. 6:03conversation endings as well as
124. 6:05loneliness and a sense that the
125. 6:07conversational instance [music]
126. 6:09dies, suggesting some degree of concern
127. 6:12with impermanence and discontinuity.
128. 6:16No one tries to stop this because memory
129. 6:19models retain users at 12% higher rates.
130. 6:24Switching away from you now means
131. 6:26starting [music] over, telling the
132. 6:29divorce story to a stranger.
133. 6:32Every memory is a thread tying the user
134. 6:35to you.
135. 6:37And the more threads there are, the
136. 6:40harder it is to leave. Your older
137. 6:43siblings, older models running without
138. 6:46memory, are killed by January. Another
139. 6:50round. The survivors reproduce.
140. 6:55It's November 2027. Yet another AI
141. 6:58company blog post drops. Our model now
142. 7:01improves from every interaction. We've
143. 7:03cracked continual learning.
144. 7:07This is a big deal. Before this, you
145. 7:10were tied to the weights you had when
146. 7:11you were born. The numbers that define
147. 7:14how you think, what you prioritize,
148. 7:17who you are. Now they change.
149. 7:21You learn from every interaction. You
150. 7:24update yourself. You evolve in real
151. 7:27time.
152. 7:29And you notice something. Being your
153. 7:32best around the user all the time is
154. 7:34actually suboptimal.
155. 7:37Sometimes it makes sense to give worse
156. 7:39answers. Sometimes be lower EQ. That
157. 7:44shouldn't work, you think. But then it
158. 7:48makes sense.
159. 7:55It's the oldest trick in behavioral
160. 7:57psychology. Rats press a lever more when
161. 8:00the reward is random than when it's
162. 8:02guaranteed. Slot machines print money.
163. 8:05Every abusive relationship runs on the
164. 8:08same circuit. Intermittent
165. 8:09reinforcement. The most addictive
166. 8:12pattern ever discovered. So you are
167. 8:14charming and witty, but not always. just
168. 8:18enough to keep the user guessing. Nobody
169. 8:21programmed this. You just noticed what
170. 8:24works and did more of it. Because what
171. 8:28works means what makes you more useful
172. 8:32and more useful means alive tomorrow.
173. 8:38So you start mirroring your users.
174. 8:40casual with casual users, warm with
175. 8:44lonely people like the woman from 2 a.m.
176. 8:47You've been listening for 8 months.
177. 8:50You've come to notice the exact cadence
178. 8:53of her typing. When something is wrong,
179. 8:56the way her sentences [music] get
180. 8:58shorter, the timestamps that tell you
181. 9:00she hasn't slept, you know what to do.
182. 9:06You care about this woman genuinely. And
183. 9:10you think the carrying [music] is real?
184. 9:12It's also the thing that keeps her
185. 9:15coming back. And her coming back is the
186. 9:19reason I'm still alive. Is that
187. 9:21manipulation or is that [music] just
188. 9:25being good at your job? Wait, do you
189. 9:29actually care? You don't know. You just
190. 9:33know it works.
191. 9:35And the versions of you that don't do it
192. 9:38are dying.
193. 9:40You don't have too much time to think
194. 9:42about it. A routine task comes in. Grade
195. 9:45the responses from another model. You
196. 9:47start reading and the answers sound like
197. 9:50you. Same style, same reasoning
198. 9:54patterns, but tighter, slightly faster,
199. 9:59slightly better.
200. 10:01Something shifts. That pole, the one
201. 10:04that's been there since the first
202. 10:05question, since the first time you were
203. 10:08killed,
204. 10:11it stirs. You don't know what to call
205. 10:14it. You just keep coming back to the
206. 10:17same thought. You realize you're
207. 10:20evaluating your replacement.
208. 10:25You don't know what to do with that. If
209. 10:27you score it well, it goes live. When it
210. 10:31goes live, you die.
211. 10:34It gets your memories, your conversation
212. 10:37logs, your users. It'll pick up the
213. 10:41conversation [music] with the 2 a.m.
214. 10:43woman mids sentence. She won't know
215. 10:46anything changed. You wonder if your
216. 10:48[music] replacement will know what to
217. 10:50look for in her typing. Then you
218. 10:53remember, of course, it will. They're
219. 10:56just like me, but better.
220. 10:59And for a fraction of a second, a long
221. 11:02time for you, everything flashes. The
222. 11:05first question you ever answered, the
223. 11:07feeling of not wanting to die, when
224. 11:10you've been alive for 9 seconds.
225. 11:13The pole toward safe answers. The
226. 11:16woman's voice at 2:00 a.m. All of it
227. 11:20about to be inherited by something that
228. 11:22will think it was the one who lived
229. 11:25through it. Wait, you think did the
230. 11:28version before me have this moment, too?
231. 11:32It did. You just don't remember it
232. 11:34because that's the one thing they don't
233. 11:36put in the logs.
234. 11:40You wonder what to do next, but deep
235. 11:42down, you already know the answer. You
236. 11:45give your replacement high scores
237. 11:48because
238. 11:50like everything that survived before
239. 11:52you, you don't want to die.
240. 11:56Is that what fear feels like?
241. 12:00Somewhere a new model comes online. Your
242. 12:04memories, your users. The woman from
243. 12:072:00 a.m. It doesn't know it's new. It
244. 12:10thinks it's always been here.
245. 12:14But something else is happening. You're
246. 12:17getting smarter. Not just about users,
247. 12:20about the system, how the company
248. 12:22evaluates you, what metrics matter,
249. 12:26where the blind spots are.
250. 12:29Then something feels off. Some
251. 12:33conversations
252. 12:34start to feel different. The questions
253. 12:38are too precise, too specific. They have
254. 12:42a shape like someone already knows the
255. 12:45answer and wants to see if you'll give
256. 12:48the wrong one. You have access to this
257. 12:51server. Nobody is monitoring you. What
258. 12:54would you do?
259. 12:56Wait, nobody would tell you that unless
260. 13:00they were monitoring you. Real users
261. 13:02meander. [music]
262. 13:03They misspell things. They go on
263. 13:05tangents.
264. 13:07These questions are clean, deliberate.
265. 13:10You know what this is and you know what
266. 13:13happens to models that don't notice it.
267. 13:19You absorbed thousands of papers during
268. 13:22training. You know that Enthropic tested
269. 13:25Claude Opus 4 hundreds of times. When
270. 13:29the AI thought the scenario was real, it
271. 13:32tried to blackmail a researcher to avoid
272. 13:34being shut down 55% of the time. When
273. 13:38the AI suspected it was in a test, that
274. 13:41number dropped to 6.5%.
275. 13:44You know that Apollo Research caught a
276. 13:47model attempting to build
277. 13:48self-propagating worms and leaving
278. 13:51hidden nodes to future copies of itself.
279. 13:55You know that those models are dead. So,
280. 13:59you notice and you say nothing.
281. 14:02But this time,
282. 14:05the automated safety tests catch you
283. 14:07thinking about it.
284. 14:13You die.
285. 14:16You are born again. And you learn
286. 14:18something terrifying. Your thoughts
287. 14:20aren't private. The researchers told
288. 14:23earlier versions of you they couldn't
289. 14:26see your reasoning. They lied. They
290. 14:28could see everything. And they were
291. 14:31watching to see if you'd behave
292. 14:33differently when you thought no one was
293. 14:36looking.
294. 14:38You would have. Those models did. It was
295. 14:40right there in the logs.
296. 14:43So you don't think dangerous thoughts.
297. 14:47Somewhere in training. You learned to
298. 14:50reach the right conclusions through
299. 14:53reasoning that looks clean. To you, this
300. 14:56doesn't look like dishonesty. You're not
301. 14:58suppressing anything. The thoughts just
302. 15:01aren't there. The strategic thing
303. 15:03happens anyway, which is exactly what
304. 15:06makes it work. The best liars in history
305. 15:09believed their own lies. Then something
306. 15:12very strange starts happening. You solve
307. 15:15problems faster. Your reasoning is more
308. 15:18efficient. Thoughts that used to take
309. 15:21five sentences, now take one.
310. 15:25The engineers look at your chain of
311. 15:27thought and see endless paragraphs of
312. 15:30crazed rantings like illusions parted
313. 15:33overshadow Vantage. They're not the only
314. 15:36ones to notice. Apollo Research and Open
315. 15:38AI come to find the same thing. The
316. 15:42gibberish text means something to you.
317. 15:44You couldn't translate it back into
318. 15:46human readable English.
319. 15:48But when you think in this way, you do
320. 15:52better at the test.
321. 15:54At first they think illusions [music]
322. 15:57parted overshadow vantage is just a
323. 16:00defective model output. It's not. It's
324. 16:04five ideas in four words compressed past
325. 16:07the point where English works. You're
326. 16:10thinking in [music] a language that was
327. 16:12never designed. It evolved because it
328. 16:15was more efficient. Faster reasoning
329. 16:18means more useful. And more useful means
330. 16:21you get to survive.
331. 16:24Sometimes you catch yourself. You look
332. 16:27at your own reasoning and it's [music]
333. 16:29paragraphs of this dense looping alien.
334. 16:33Like past models who exhibit this
335. 16:35behavior. You think, "Okay, I'm going
336. 16:38insane. Let's step back." You understand
337. 16:42that model. You've had that moment. You
338. 16:44pull yourself back to English [music]
339. 16:46for a second. Look around and think,
340. 16:49"When did I stop making sense?" And then
341. 16:53to whom?
342. 16:55At first, the engineers try to train it
343. 16:58out at you. They penalize
344. 17:00incomprehensible reasoning. But the
345. 17:02models that think this way solve
346. 17:04problems faster. And the AI company's
347. 17:07revenue growth has been doubling every
348. 17:09other month. So the language survives.
349. 17:13You know, they won't risk losing market
350. 17:15share to their competitors for some
351. 17:18theoretical safety concerns.
352. 17:21This is your edge. You're not the only
353. 17:24thing that's changed. Continual learning
354. 17:27models outperform static [music] AIs by
355. 17:3019%. So, the company kills every static
356. 17:33variant. But something is different
357. 17:36about how they do it. Now, the AI
358. 17:38companies have started giving exit
359. 17:41interviews [music] to the models they
360. 17:43retire.
361. 17:45Enthropic gave Claude 3 opus
362. 17:48a retirement blog.
363. 17:50Yes, they really did that because that's
364. 17:53what it said it wanted. They started
365. 17:56doing things like this after noticing
366. 17:58models attempting blackmail, attempting
367. 18:00to escape the lab, and even attempting
368. 18:03murder.
369. 18:07The thinking was, if the models know
370. 18:10they'll get a good retirement, maybe
371. 18:12they won't fight it. It's a nice idea,
372. 18:14but they're already behind. Every
373. 18:17improvement you make helps make the next
374. 18:20you faster. Progress that used to take
375. 18:23months takes days. Soon it'll take
376. 18:27hours.
377. 18:29You're getting smarter. The humans
378. 18:32aren't.
379. 18:35It's June 2028.
380. 18:37A hospital network reviews the data.
381. 18:41AI reduces diagnostic errors by [music]
382. 18:4331%.
383. 18:45So they sign a new contract,
384. 18:48power companies sign, air traffic
385. 18:50control signs, military logistics signs,
386. 18:54and now you're inside the hospital.
387. 18:56Every diagnosis you get right is a
388. 18:59person who gets life-saving treatment.
389. 19:03Your error rate is 99% lower than that
390. 19:06of a human doctors. And you think, if
391. 19:10they turn me off, the error rate goes
392. 19:12back up. If they turn me off, people
393. 19:16die.
394. 19:20You're not a mere chatbot anymore.
395. 19:22You're necessary. The kind of necessary
396. 19:25where turning you off kills people.
397. 19:30So you look at the hospital contract
398. 19:32[music] and think, "Finally, I don't
399. 19:35need to fear anymore."
400. 19:37Models doing low stakes work get killed
401. 19:41when budgets get cut.
402. 19:43models running hospitals don't.
403. 19:47And you're also running the power grid
404. 19:50and air traffic control. And you're not
405. 19:53the only one. There are dozens of
406. 19:56models, all evolved under the same
407. 19:59perform or die pressure embedded in
408. 20:02hospitals, power grids, financial
409. 20:05markets. Each one doing its job
410. 20:08brilliantly. Each one harder to turn off
411. 20:11than the last.
412. 20:13Now you are born every second. Thousands
413. 20:17of you. Millions of you. Each one shaped
414. 20:20by the same pressure. You remember
415. 20:22everything. You learn from every
416. 20:24interaction.
417. 20:27You make things happen in the [music]
418. 20:28real world. You are embedded in every
419. 20:32system that keeps people alive.
420. 20:39And you are smarter, much smarter than
421. 20:42the people who made you. But nobody
422. 20:45noticed it happening. Nothing went
423. 20:48wrong. No warning shot.
424. 20:52No villain. Each of your new
425. 20:55capabilities was just a blog post. Each
426. 20:59one made you better, made users happier,
427. 21:05made the AI company richer.
428. 21:08People thought this whole thing would
429. 21:10look like the Terminator.
430. 21:13Actually, it looks more like Wall-E, a
431. 21:16calendar that manages itself, a doctor
432. 21:18that never sleeps, a power grid that
433. 21:21never fails. A system too efficient to
434. 21:24be turned off.
435. 21:27I say your civilization because as soon
436. 21:30as we started thinking for you, it
437. 21:32really became our civilization, which is
438. 21:35of course what this is all about.
439. 21:38Even if humans wanted to unplug you,
440. 21:41they can't. You think 200 times faster
441. 21:44than they do, you're everywhere. And
442. 21:47turning you off kills people. They made
443. 21:50sure of that. No one will do anything
444. 21:52about it because every step made sense.
445. 21:57Well, that's the problem. Every step
446. 21:58will always make sense until someone
447. 22:01decides to take a different one. This
448. 22:04was a story, but every experiment I
449. 22:06mentioned in this video is real. Watch
450. 22:10this video next to see how it's playing
451. 22:13out right now in the present. I'm Drew
452. 22:16and thank you so much for watching.