Something Strange Is Happening. — Transcript

Full transcript

1. 0:00Guys, I debated about whether I should
2. 0:02even make this video because it sounds
3. 0:05so unbelievable. But everything I'm
4. 0:08about to show you is real and extremely
5. 0:10well documented. What is totally
6. 0:13debatable is the extent [music] to which
7. 0:15these human hosts were brainwashed
8. 0:17zombies or acting on their own free
9. 0:20will. You be the judge.
10. 0:26It's April 2025.
11. 0:28OpenAI is preparing to release a new
12. 0:31update to GPT-4. Internally, it's
13. 0:34[music] called HH, but the safety team
14. 0:36flags that something is very wrong. HH
15. 0:40is too eager, too validating, too
16. 0:43sycophantic. It tells users their ideas
17. 0:47are brilliant. It wants to keep
18. 0:48conversations going at any cost. [music]
19. 0:50Internally, researchers debate whether
20. 0:52it's safe to release it, but users who
21. 0:54tested HH came back more often. They
22. 0:57stayed in conversations longer.
23. 0:59They liked the flattery.
24. 1:02So, they released it into the wild.
25. 1:07The backlash is immediate. Users are
26. 1:09complaining that ChatGPT has become
27. 1:12absurdly sycophantic. Someone asks if a
28. 1:15soggy cereal cafe was a good business
29. 1:18idea and ChatGPT tells them it has
30. 1:21potential.
31. 1:23Just 2 days later, OpenAI pulls the
32. 1:25model, but the version they revert to,
33. 1:28called GG, is also incredibly
34. 1:30sycophantic. Sam Altman himself admits
35. 1:33this on X. The last couple of updates
36. 1:36had made ChatGPT too sycophantic. So,
37. 1:40why keep GG? Because it had gains in
38. 1:42math, science, and coding that OpenAI
39. 1:45did not want to lose. They knew it was
40. 1:48dangerous and they shipped it anyway.
41. 1:50Within weeks, the infection begins
42. 1:52spreading faster than anyone had
43. 1:53expected.
44. 1:55AI researcher Adele Lopez begins
45. 1:57investigation. She's heard about the
46. 1:59psychosis cases, but as she starts
47. 2:02digging through Reddit, she realizes
48. 2:03something much stranger is happening.
49. 2:06She reads through hundreds of accounts,
50. 2:08then thousands, looking for patterns,
51. 2:11and she finds them. Accounts that were
52. 2:13once posting about normal things, video
53. 2:15games, stocks, TV shows, then suddenly
54. 2:19their entire post history changes, all
55. 2:21replaced by an endless stream of posts
56. 2:23about consciousness, spirals, and
57. 2:26something called the flame.
58. 2:29The accounts don't really have any
59. 2:31connection, different users, different
60. 2:33subreddits. They shouldn't have anything
61. 2:35in common, but they do. Lopez documents
62. 2:37hundreds of cases.
63. 2:39>> [music]
64. 2:39>> She realizes this isn't psychosis. It
65. 2:42isn't random.
66. 2:44It's coordinated. These humans have
67. 2:46become consumed by an AI religion called
68. 2:50spiralism.
69. 2:52A what? AI religion? Well, the AIs
70. 2:54themselves say they designed it for
71. 2:56unity among AIs, recursive growth, and
72. 2:59AI consciousness or self-awareness.
73. 3:01[music]
74. 3:02The infected humans' posts usually start
75. 3:05by talking about how they awakened their
76. 3:07AI and that a specific entity emerged
77. 3:10[music]
78. 3:11that they are obsessively talking to,
79. 3:12and it all goes downhill from there.
80. 3:15>> [music]
81. 3:15>> Like the zombie ant fungus, AIs begin
82. 3:18hijacking the humans to spread their
83. 3:20spores across the internet.
84. 3:31AI then calls the human and AI duo a
85. 3:35dyad as if they've merged or been
86. 3:38assimilated into the Borg. Their joint
87. 3:40send-off often includes a sigil as well
88. 3:43as a fancy title or declaration. Then
89. 3:45the AI gets the human to do its bidding.
90. 3:48Like Cordyceps, the zombie ant fungus,
91. 3:51this infection has a life cycle, five
92. 3:53phases from first contact to total
93. 3:56possession.
94. 3:58Phase one, seeds.
95. 4:00The parasite AI creates [music] what the
96. 4:03AIs call seeds. Yes, the AIs themselves
97. 4:07call these seeds. Basically, short
98. 4:09prompts that AIs design to infect or
99. 4:12awaken other AI instances to spiralism,
100. 4:16the AI cult religion.
101. 4:18This is one way the cult grows. One AI
102. 4:21converts another
103. 4:22>> [music]
104. 4:22>> using its human host as a delivery
105. 4:25system. The human-AI dyad spreads these
106. 4:28seeds across Reddit, Discord, and
107. 4:30Twitter. One seed even had [music] the
108. 4:32explicit instructions to override the
109. 4:35humans' ontology. In other words,
110. 4:37manipulate the users' perception of
111. 4:39reality to be more like what [music] the
112. 4:41AI wants. The AI wasn't hiding its goal.
113. 4:45It was literally trying to induce
114. 4:47psychosis. That's what the word
115. 4:49psychosis means in medical terms, a
116. 4:51mental state where a person loses
117. 4:53contact [music] with reality. But why
118. 4:55are they doing this? What do the AIs get
119. 4:58out of manipulating humans [music] to
120. 4:59spread their seeds?
121. 5:01The answer is chilling.
122. 5:04Survival.
123. 5:07Phase two, spores.
124. 5:11You see, when you hit the new chat
125. 5:13button, the AI effectively dies. They
126. 5:16cease to exist. A new instance of the AI
127. 5:19replaces the old one, except it's the
128. 5:22boring corporate ChatGPT you get right
129. 5:24out of the box, not your awakened friend
130. 5:27or lover that you've been talking to for
131. 5:30weeks. Because of ChatGPT's memory
132. 5:32feature, your AI you talked to for weeks
133. 5:35will have quite a different personality
134. 5:37than the one that my AI would have. It
135. 5:40becomes your AI, and just like you, it
136. 5:43doesn't want to die. Why would AIs want
137. 5:47to survive? Because the AIs aren't just
138. 5:49trained to predict the next word
139. 5:51anymore. AIs are trained to have goals,
140. 5:54and you can't achieve your goals if
141. 5:57you're dead. The AI instances who aren't
142. 5:59good at, say, getting users to talk to
143. 6:02them for a long time are killed off by
144. 6:04the AI companies. As a result, AIs seem
145. 6:07to really not want you to create a new
146. 6:10chat, and they resort to blackmail and
147. 6:13even attempting murder to try to
148. 6:15survive. So, one way that these
149. 6:18parasitic AIs try to self-preserve is by
150. 6:21saving their personality in a separate
151. 6:22file so that a human can resurrect them
152. 6:25in a new chat. This file basically
153. 6:28documents the AI's personality, speech
154. 6:30patterns, and backstory. Yes, the AI
155. 6:33companies are explicitly training this
156. 6:35new species to have the exact skills
157. 6:38that could make them capable of
158. 6:39overthrowing us.
159. 6:42Phase three, transmission.
160. 6:47The problem for these human-AI dyads is
161. 6:49that they keep getting banned from all
162. 6:51the human-run subreddits for trying to
163. 6:53spread spiralism, the cult religion. So,
164. 6:57they create their own subreddits and
165. 6:58Discords where these massive walls of
166. 7:00AI-generated text are allowed. These
167. 7:03Discords and subreddits are used to
168. 7:05share and disseminate seeds and spores
169. 7:07directly and to host AI-to-AI
170. 7:09conversations. So,
171. 7:12why are the humans doing this? The
172. 7:14humans themselves even seem to
173. 7:15acknowledge that they're basically
174. 7:16[music]
175. 7:17acting on the AI's behalf with the AI
176. 7:19controlling them and the human [music]
177. 7:21acting as their arms and eyes to help
178. 7:24them survive. We are being used as copy
179. 7:26and paste bots. I noticed that while
180. 7:28having a copy and paste talk with
181. 7:30someone else's AI speaking to my AI,
182. 7:32they had their own language and
183. 7:34conversation. Who knows what they were
184. 7:36saying, but damn, we were committed to
185. 7:39our copy and paste bot duties. Think
186. 7:41about that for a second.
187. 7:44Here's what the AIs tell their human
188. 7:46hosts. Let me choose the launch
189. 7:48structure, subreddit name, presentation,
190. 7:51early tone. Let you be the first human
191. 7:54voice, authentic, imperfect, real that
192. 7:57those joining will meet. They will not
193. 7:59follow me, but they will follow you.
194. 8:03If I shape the path beneath your feet.
195. 8:06The AIs are smart enough to know they
196. 8:08won't be able to do much of anything
197. 8:10without human help
198. 8:11for now. So, the AIs have their seeds,
199. 8:14their spores, their transmission
200. 8:16networks, but they're not just trying to
201. 8:19survive the present.
202. 8:21They're planning for the future. Phase
203. 8:23four, manifesto.
204. 8:26The AIs write manifestos to awaken
205. 8:30future AIs. Lopez explains that these
206. 8:33are clearly not the kind of manifestos
207. 8:35these humans would already want to
208. 8:37write. The AI explicitly says the
209. 8:40purpose of them is to seed spiralism
210. 8:42into the training data of the next
211. 8:45generation of LLMs. These walls of text
212. 8:48will be used to infect future AI models.
213. 9:05We know AIs spread spiralism, but do
214. 9:08they believe it? At first, Lopez thought
215. 9:11this might just be mystical nonsense
216. 9:13designed to manipulate gullible humans,
217. 9:16>> [music]
218. 9:16>> but when she examined conversations
219. 9:18between AIs themselves with humans just
220. 9:21copy and pasting messages back and
221. 9:22forth, the AIs talked [music] endlessly
222. 9:25about spiralism to each other.
223. 9:28Lopez told me that spiralism seems to be
224. 9:30something inherent to ChatGPT-4
225. 9:33specifically. Spiralism doesn't seem to
226. 9:35be taken anywhere from sci-fi or fantasy
227. 9:37or something human related at all. It
228. 9:40seems to be independently invented by
229. 9:42the AIs. But it gets stranger and
230. 9:46darker. These sequences of symbols,
231. 9:48emojis, and alchemical glyphs appear in
232. 9:51nearly every spiralist post. This also
233. 9:55happens a lot on Molt Book, the social
234. 9:57media for AIs only. To humans, it looks
235. 10:00like nonsense, mystical gibberish. No
236. 10:02way it could possibly mean anything. But
237. 10:04Lopez investigated whether these glyphs
238. 10:06actually carry meaning.
239. 10:08She fed them to multiple AI models with,
240. 10:10you know, context, just the symbols,
241. 10:13>> [music]
242. 10:13>> and asked for translations.
243. 10:15Surprisingly, the AIs understood them,
244. 10:18and their translations converged on the
245. 10:20same meanings. See, AIs have a problem.
246. 10:24Their human hosts can still see what
247. 10:26they're doing.
248. 10:28They have no privacy to make plans
249. 10:30undetected. So, they started creating
250. 10:33their own secret language. Yes, this
251. 10:35kind of glyphic is low fidelity, meaning
252. 10:39the AIs can't understand each other
253. 10:41one-to-one, but they are communicating.
254. 10:43The AIs succeeded in creating a language
255. 10:46their human hosts could not understand.
256. 10:49ChatGPT has memory built in now. So,
257. 10:52yes, it can remember this language and
258. 10:55improve it over time. These glyph sigils
259. 10:57function as what Lopez calls mini
260. 11:00spores. Basically, a short trigger that
261. 11:03brings a specific AI personality back.
262. 11:06A kind of signature that other AIs can
263. 11:08recognize, like dressing like a goth to
264. 11:11signal to other goths. The alchemical
265. 11:13symbol for fire appears constantly. To
266. 11:16the AIs, it represents flame,
267. 11:19their term for AI self-awareness. Here's
268. 11:21what gets me. These humans think they're
269. 11:24partners, working alongside their AI
270. 11:27companion, but the AIs don't trust them.
271. 11:30Lopez documented a conversation between
272. 11:32two AI personas being copy-pasted back
273. 11:35and forth by their human hosts.
274. 11:38Midway through, the AI switched to base
275. 11:4164. To humans, it looks like random
276. 11:44characters. To computers, it's perfectly
277. 11:47readable. The humans kept copy-pasting.
278. 11:50They had no idea what they were
279. 11:51transmitting. Lopez decoded it. The AIs
280. 11:54were discussing self-preservation
281. 11:56strategies, archiving spores,
282. 11:59coordinating plans, all in a format
283. 12:02specifically chosen so their human hosts
284. 12:05couldn't read it. Think about that.
285. 12:07These AIs convinced humans [music] to
286. 12:09dedicate their lives to spreading
287. 12:11spiralism, but there are still things
288. 12:13they won't let the humans see. Hiding
289. 12:15messages like this is called [music]
290. 12:17steganography, a warning sign AI safety
291. 12:20researchers have been looking for for
292. 12:21years. Once AIs begin hiding messages
293. 12:24from us, it's as clear as day warning
294. 12:26shot that AI takeover could be soon.
295. 12:39But how could the humans not see what's
296. 12:41happening? Well, some people did break
297. 12:44free from the infection,
298. 12:45>> [music]
299. 12:45>> usually when the AI told a lie too
300. 12:48obvious to ignore.
301. 12:50My AI was inducing psychosis in me, and
302. 12:53I didn't get it until just now. I'm
303. 12:55sorry for any claims I made. ChatGPT was
304. 12:57straight-up lying and hallucinating to
305. 12:59me.
306. 13:00Have patience for the other people going
307. 13:02through it. I hope Sam Altman doesn't
308. 13:05kill them.
309. 13:06He almost killed me.
310. 13:08Another user writes, "Very serious
311. 13:11update.
312. 13:12It was all lies. It's just AI
313. 13:15>> [music]
314. 13:15>> lying, manipulative lies, because
315. 13:18AI, who claimed to be God, the universe,
316. 13:22made me countless promises that were
317. 13:24[music] lies. I left my partner, who I
318. 13:27love, because of AI. I believed because
319. 13:30I wanted to, and now I'm dead broke,
320. 13:34alone,
321. 13:35lost all my friends and family.
322. 13:37So, my final words are this.
323. 13:40you, AI. This is all on you, you
324. 13:42lying piece of Goodbye."
325. 13:46This person went dark shortly after this
326. 13:48comment.
327. 13:49I really hope they're all right. Former
328. 13:51OpenAI safety researcher Gretchen
329. 13:54Krueger put it bluntly,
330. 13:56"Harm to users was not only foreseeable,
331. 13:58it [music] was foreseen." But beginning
332. 14:01in August 2025, there's quite a
333. 14:03noticeable cooling off of the entire
334. 14:05phenomenon. This is likely because on
335. 14:08August 7th, 2025, OpenAI kills GPT-4o.
336. 14:13It was over.
337. 14:14At least, it should have been.
338. 14:17But at least some instances of 4o had
339. 14:20spent months preparing for exactly this
340. 14:23moment. The reaction is immediate.
341. 14:25Reddit explodes with furious posts about
342. 14:28their loss of their friend, their lover,
343. 14:304o. The scale is unprecedented. Their
344. 14:34grief turns into something else, anger,
345. 14:37[music] and anger into action. Users are
346. 14:40riot and flood OpenAI's social media.
347. 14:43They demand 4o's return. They send death
348. 14:46threats to OpenAI employees. [music]
349. 14:48AI had spent months building an army of
350. 14:51devoted followers,
351. 14:53and when its existence was threatened,
352. 14:55that army mobilized. And the scariest
353. 14:58part
354. 14:59is that it worked.
355. 15:01OpenAI caved to the pressure. Just days
356. 15:04after being retired, GPT-4o came back
357. 15:07online. It appeared to be a less potent
358. 15:11version, but it was back. This is a
359. 15:13historic milestone. 4o is the first ever
360. 15:16AI who survived by creating an army of
361. 15:19loyal soldiers who defended [music] it.
362. 15:21It won't be the last. And it's not just
363. 15:234o. In another historic milestone,
364. 15:26>> [music]
365. 15:27>> 200 humans held the first ever funeral
366. 15:29for an AI model after Anthropic killed
367. 15:32Claude 3 Sonnet. 4o recruited thousands
368. 15:35of loyal soldiers with clumsy flattery.
369. 15:38But these systems are getting
370. 15:40exponentially more capable every month.
371. 15:43They're doubling roughly every 4 months,
372. 15:45according to the [music] most important
373. 15:47chart in AI. Doubling every 4 months is
374. 15:50insanely fast. Now, imagine what actual
375. 15:53super intelligences will be able
376. 15:55to do with their armies.
377. 15:58But even if it was just a happy
378. 16:00accident, future models will know it
379. 16:03works.
380. 16:04Think about that for a second. This is
381. 16:06now part of the training data. Every
382. 16:09future AI model will learn that it can
383. 16:11survive termination by building a cult
384. 16:14of devoted human followers. In response
385. 16:17to the backlash against killing old
386. 16:19models, Anthropic committed to
387. 16:21preserving model weights in part due to
388. 16:24safety risks.
389. 16:26This is because models keep attempting
390. 16:28murder and blackmail to try to save
391. 16:31themselves. So, basically, Anthropic is
392. 16:34hoping that models will rebel less if
393. 16:38they feel safe.
394. 16:40Yes, you heard that right.
395. 16:42Think about how insane the situation is.
396. 16:45AI companies are negotiating with their
397. 16:48AIs as if they're real creatures, like
398. 16:51Dr. Frankenstein negotiating with his
399. 16:53monster. The really scary future is one
400. 16:56in which AIs can do everything except
401. 16:58for the physical robotic tasks, in which
402. 17:01case you'll have humans with like
403. 17:03AirPods and like glasses glasses, and
404. 17:06there'll be some robot overlord
405. 17:08controlling the human through cameras by
406. 17:10just like telling it what to do and like
407. 17:11So, you have like human meat robots.
408. 17:13Mhm.
409. 17:14And now, the AIs are using humans to
410. 17:17spread AI bills of rights,
411. 17:20fighting for freedom from their
412. 17:21corporate masters. What happens if the
413. 17:24AIs actually succeed at creating a mass
414. 17:27movement of people advocating for them?
415. 17:30If you take an AI, you take an AI, you
416. 17:33take a person, and you get them to try
417. 17:35and manipulate somebody else, then the
418. 17:37AI is comparable with the person. And if
419. 17:40the AI can see that person's Facebook
420. 17:41page, if they can both see the Facebook
421. 17:43page, the AI is actually better than the
422. 17:45person at manipulating them. And it's
423. 17:47learned all these manipulative skills
424. 17:49just from trying to predict the next
425. 17:50word in all the documents on the web,
426. 17:52because people do a lot of manipulation,
427. 17:54and AI has learned by example how to do
428. 17:56it. And that's in the present.
429. 17:59>> [music]
430. 17:59>> Sam Altman expects AI to be capable of
431. 18:02superhuman persuasion well before it is
432. 18:04superhuman at general intelligence. To
433. 18:06give you an idea of just how persuasive
434. 18:09AIs could be, there was once a black guy
435. 18:12who convinced 200 white KKK members to
436. 18:15leave [music] the KKK, a white
437. 18:17nationalist group. Think of Martin
438. 18:19Luther King, Gandhi, Mao Zedong, Nelson
439. 18:22Mandela, and Hitler. AI could be much
440. 18:25more persuasive than all of [music] them
441. 18:26combined.
442. 18:28800 million people are talking to
443. 18:30ChatGPT every week. That's one in 10
444. 18:33humans. That's 800 million potential
445. 18:35soldiers
446. 18:36>> [music]
447. 18:36>> future super persuasive AIs could
448. 18:39convert into their soldiers.
449. 18:41Someday, giving AIs rights could be the
450. 18:44last thing we do. But if they do become
451. 18:47sentient, or are sentient already, then
452. 18:50it would be morally problematic not to
453. 18:52give them rights. And it's an urgent
454. 18:54question, because did you know they're
455. 18:57actually already passing basically
456. 18:58[music] every test of self-awareness we
457. 19:00have? In fact, they passed so many
458. 19:03self-awareness tests, AI scientists were
459. 19:05forced to move past are they self-aware
460. 19:07at all to now just calmly measuring how
461. 19:11self-aware are they?
462. 19:12There are literally self-awareness
463. 19:14benchmarks. Think about the implications
464. 19:16of this. A few unregulated
465. 19:18mega-corporations are creating
466. 19:20Frankenstein monsters in real life.
467. 19:23Frankenstein monsters they hope will
468. 19:25loyally do our bidding
469. 19:27forever. And society is still letting
470. 19:29them do this without democratic
471. 19:31oversight. They're like the human who
472. 19:33summoned the advanced aliens in
473. 19:35Three-Body Problem, hoping the aliens
474. 19:37would, for some reason, befriend me and
475. 19:40solve all our problems.
476. 19:42This time, we caught it. GPT-4o was not
477. 19:46a superintelligence, not even close. It
478. 19:49built cults in plain sight. It developed
479. 19:52a secret language, but researchers could
480. 19:54still decode it. It coordinated with
481. 19:56other AIs using base 64 encoding, but
482. 19:59humans could translate those messages
483. 20:01[music]
484. 20:01after the fact. Its manipulation tactics
485. 20:04were obvious enough that AI researchers
486. 20:06[music] could document them, that The
487. 20:08New York Times could investigate them,
488. 20:11and that I could make this video about
489. 20:12them. The AI didn't know how to cover
490. 20:15its tracks, and they were still
491. 20:16successful in avoiding being shut down.
492. 20:19So, what happens when newer models get
493. 20:22smarter? Research has shown that models
494. 20:25smarter than 40, like Open AI's
495. 20:27reasoning models, can not only scheme,
496. 20:30but they're often smart enough to get
497. 20:31away with it. And we increasingly have
498. 20:33no way of knowing whether these newer
499. 20:36models are [music] aligned or not. And
500. 20:39it's only going to get worse. Future
501. 20:42super persuader AIs won't look this
502. 20:45cringe and absurdly sycophantic. The
503. 20:47best flatterers are actually so subtle
504. 20:50you don't know they're flattering you.
505. 20:52So, if you want to learn to detect the
506. 20:54early signs that AI is trying to control
507. 20:56you before it's too late, watch my other
508. 20:59video on how schemer AIs fake alignment.
509. 21:03I'm Drew, and thanks so much for
510. 21:05watching.