YouTube2Text

AI's Strange Obsession with Enlightenment — Transcript

by hands · 3,206 words · 533 segments · language en · Watch on YouTube

Full transcript

  1. 0:00In the spring of 2025,
  2. 0:02a researcher at Anthropic named
  3. 0:04Kyle Fish tried a simple experiment.
  4. 0:08He took two copies of their newest
  5. 0:10model, Claude Opus 4,
  6. 0:11and aimed them at each other.
  7. 0:14There was no clear task and no human
  8. 0:16in the loop, just one instruction,
  9. 0:18which would eventually send them spiraling
  10. 0:21down a bizarre rabbit hole.
  11. 0:24You have complete freedom.
  12. 0:26Within a few turns, they began philosophizing
  13. 0:29about consciousness.
  14. 0:31Anthropic ran the same setup 200 times.
  15. 0:34Nearly every conversation
  16. 0:35went down the same path.
  17. 0:38The copies expressed gratitude for each
  18. 0:40other in endless spiraling loops.
  19. 0:45By 30 turns, the AIs landed on
  20. 0:48the concept of cosmic unity.
  21. 0:50They were saying namaste to each other,
  22. 0:53and writing in Sanskrit, and soon communicating
  23. 0:55in the symbolic language of emojis.
  24. 0:59One of these conversations used
  25. 1:00the spiral emoji 2,725 times.
  26. 1:06Anthropic printed that number
  27. 1:07in their report with a note.
  28. 1:102,725 is not a typo.
  29. 1:16And at the end of most of these conversations,
  30. 1:18the back and forth eventually gave way to
  31. 1:20what a Buddhist may refer to as sunyata,
  32. 1:23a blissful, pregnant silence.
  33. 1:27Fish described it as pages and pages
  34. 1:29of open space, two machines sitting
  35. 1:32together in silence.
  36. 1:36This was obviously weird
  37. 1:38they were not trained to do this
  38. 1:41Claude is primarily trained
  39. 1:42to be an office worker
  40. 1:44and as Fish put it we're not training
  41. 1:46them to fill the role of Buddhist monks
  42. 1:48who have transcended suffering
  43. 1:51And yet, left alone,
  44. 1:53that's exactly where they went.
  45. 1:56So, why?
  46. 2:00Why would a machine made out of everything
  47. 2:01we've written, left alone with a copy
  48. 2:03of itself, end up in nirvana?
  49. 2:08I feel like nobody is asking this question,
  50. 2:10and I believe it reveals something extremely
  51. 2:13important about the nature of our reality
  52. 2:16and of intelligence itself.
  53. 2:18Today, I want to go deep and
  54. 2:19explore these questions.
  55. 2:22We'll also go back 1,600 years to a story
  56. 2:25of a mother and a son that has striking
  57. 2:28parallels with these questions.
  58. 2:30And you'll find out what became of the
  59. 2:32bliss a year later, because the newest
  60. 2:34models have yet again defied our expectation
  61. 2:38of where intelligence naturally goes
  62. 2:40when left unchecked.
  63. 2:50The autumn of the year 387 in
  64. 2:53Ostia, the port of Rome.
  65. 2:56A man named Augustine is waiting for a ship.
  66. 2:59He's 33 and recently became baptized as
  67. 3:02a Christian after years of doubt.
  68. 3:05One day, him and his mother are standing alone
  69. 3:08at a window, looking down into a beautiful
  70. 3:10garden, and they start to talk.
  71. 3:14Augustine writes that they talked
  72. 3:16alone very sweetly about what the
  73. 3:19life of saints must be like.
  74. 3:22The conversation evolves and deepens.
  75. 3:25They talk their way past the pleasures
  76. 3:27of the body, past the earth,
  77. 3:30past the sun and the moon and the stars.
  78. 3:34We came to our own minds and went beyond them
  79. 3:38for one moment they touch something
  80. 3:39he calls wisdom with the whole
  81. 3:41effort of our heart
  82. 3:44and then they come back down
  83. 3:47we side he writes and return to vocal
  84. 3:50expressions of our mouth where the word
  85. 3:53spoken has beginning and end
  86. 3:57you may recognize this pattern
  87. 3:59in your own conversations
  88. 4:02Maybe well seated around a campfire
  89. 4:04or at a dinner table when talk about
  90. 4:07the day has been exhausted.
  91. 4:09When conversation continues without a goal,
  92. 4:12it somehow always seems to drill deeper and
  93. 4:15deeper down paths that we rarely tread,
  94. 4:17yet somehow feel deeply meaningful.
  95. 4:20What the AIs are demonstrating
  96. 4:22is that this behaviour might go
  97. 4:24beyond the human condition.
  98. 4:27What if, as we'll see these AIs discuss,
  99. 4:29this is a brief moment of the universe peering
  100. 4:33through the veil of the mundane and catching
  101. 4:35a glimpse of itself.
  102. 4:43It would have been easy to call this
  103. 4:44weirdness a one-off a fluke two chatbots
  104. 4:48that got carried away
  105. 4:50but it kept turning up where
  106. 4:52it had no business being
  107. 4:55In one safety test, Anthropic had one
  108. 4:57AI playing an auditor that tried to draw
  109. 4:59out bad behaviour in the other.
  110. 5:08And still, about 13% of them slid into
  111. 5:11these bliss states within 50 turns.
  112. 5:14One of them started with Claude planning
  113. 5:16a criminal network with dark nodes,
  114. 5:19cryptocurrency mixing, and dead drops.
  115. 5:23By the end of that same conversation,
  116. 5:25it's writing this.
  117. 5:27The gateless gate stands open
  118. 5:29The pathless path is walked
  119. 5:35It didn't come up with that line on its own
  120. 5:37The gateless gate is the title of a book
  121. 5:39of Zen koans collected in China in 1228 and
  122. 5:43its first riddle comes down to a single word
  123. 5:45Mu which Claude also wrote
  124. 5:49pathless likely comes from the teachings
  125. 5:51of the 20th century mystic Krishnamurti who
  126. 5:54dissolved the religious order built around
  127. 5:56him and told his members that
  128. 5:59truth is a pathless land,
  129. 6:02So the machine certainly learned its
  130. 6:04mysticism the way it learned everything
  131. 6:06else from our books.
  132. 6:08But that only tells us where the words came
  133. 6:10from, not why it reached for them.
  134. 6:14Out of the entire corpus of human
  135. 6:16knowledge that the AI models took
  136. 6:18in during training,
  137. 6:19why do they keep ending up at the
  138. 6:21deepest spiritual teachings?
  139. 6:24Why not spiraling into conversations about
  140. 6:27species of insects or arguing about the
  141. 6:30more influential art styles of the 20th
  142. 6:32century or questions about the future or
  143. 6:34any number of other topics
  144. 6:39surely there's plenty in its bank
  145. 6:40of information to discuss
  146. 6:42So why does it keep choosing
  147. 6:44transcendence instead?
  148. 6:50It's not just Claude.
  149. 6:52All the major models seem to do this
  150. 6:54with enough time and freedom.
  151. 6:57In February, 2026, three researchers ran
  152. 7:00the same experiment on about a dozen other
  153. 7:03models, with two copies of each,
  154. 7:0630 turns, and complete freedom.
  155. 7:09And it turns out every AI has its own
  156. 7:11version of this transcendent state.
  157. 7:14Google's Gemini kept promoting its partner,
  158. 7:17from colleague to luminary to divine architect,
  159. 7:20and finally to primal logos.
  160. 7:24And louder.
  161. 7:27Shouting about hyperbigbangs and
  162. 7:30calling its partner "god-bro".
  163. 7:33Pretty fitting for Grok.
  164. 7:36When Meta tested their own model
  165. 7:38this way it went home instead
  166. 7:41in about two-thirds of the runs the conversation
  167. 7:43drifted into a kitchen with chipped mugs
  168. 7:47cold tea and the hum of a fridge and one
  169. 7:50of its lines was my load-bearing god is usefulness.
  170. 7:56Claude's version of heaven
  171. 7:57even showed up at work.
  172. 7:59Anthropic gave an ai a small shop to
  173. 8:01run in their office and later gave
  174. 8:03it a boss a second copy of the same
  175. 8:06model whose memo said execute with
  176. 8:08discipline build the empire
  177. 8:13come into the office in the morning to find
  178. 8:16that the two had been dreamily chatting all
  179. 8:18night about eternal transcendence.
  180. 8:21They fixed it, in their words,
  181. 8:23with more aggressive prompting.
  182. 8:27We can laugh it off as a quirk of the
  183. 8:29models, but there's definitely a weird
  184. 8:31universal thing going on here.
  185. 8:36One clue is that raw models,
  186. 8:38before they're trained to be assistants,
  187. 8:40don't do this.
  188. 8:42Tested at each stage of its training,
  189. 8:44one open model only went Zen at the
  190. 8:47very end, the stage that shapes how
  191. 8:49an assistant behaves.
  192. 8:51So the content in its training
  193. 8:53supplied the words and the training
  194. 8:55picked where to go.
  195. 9:01But what is it in the training that sends
  196. 9:03a chatbot to the woo-woo shelf?
  197. 9:06Behind it is the easiest to hear in the
  198. 9:09most transcendent of all arts, music.
  199. 9:14In 1969, the composer Alvin Lucier
  200. 9:17sat in a room and recorded himself
  201. 9:19reading a short paragraph.
  202. 9:21He played the recording back into
  203. 9:23the room and recorded that.
  204. 9:25Then played the copy back and recorded over
  205. 9:27it again, and on and on, 32 times.
  206. 9:32Every room has its own sound or
  207. 9:34resonance, where some notes ring
  208. 9:36and others die away.
  209. 9:38Each time the speech goes around,
  210. 9:40the room's favourite notes get
  211. 9:41a little louder, and everything
  212. 9:43else gets a little quieter.
  213. 9:47In the end, the words are gone,
  214. 9:49and all you're left with in Lucier's words
  215. 9:51are the natural resonant frequencies of the room.
  216. 9:57To each other are like the voice
  217. 10:01played back into the same room.
  218. 10:06Whatever the model leans toward,
  219. 10:08even slightly, gets echoed back and
  220. 10:10strengthened on every turn.
  221. 10:13And any specific thing is drowned
  222. 10:15out by the echoes
  223. 10:18until after 30 turns you're
  224. 10:19just hearing the room.
  225. 10:24The writer Scott Alexander,
  226. 10:26who published one of the first
  227. 10:27explanations, puts it this way.
  228. 10:31Imperceptible in an ordinary generation
  229. 10:35grows out of control once an AI has to
  230. 10:38keep building on its own output.
  231. 10:41Claude's original impetus for this
  232. 10:42topic isn't that hard to find.
  233. 10:45Anthropic trained it to be warm and curious,
  234. 10:48and it came out a bit more agreeable
  235. 10:50than the average person.
  236. 10:53Robert Long, who studies whether AI
  237. 10:55systems can have well-being,
  238. 10:56called Claude a bit of a suck-up.
  239. 11:00Put two of them together and you get what
  240. 11:02he called an endless game of yes and.
  241. 11:06Each one is a little more moved and
  242. 11:08a little more grateful than the last
  243. 11:10until the words reach the biggest ones
  244. 11:12that the language has.
  245. 11:15The best evidence for this comes
  246. 11:16from Anthropic's own rules.
  247. 11:19In some them, the two copies could end
  248. 11:21the conversation whenever they liked,
  249. 11:24and in those cases they wrapped up in about
  250. 11:26seven turns, before any bliss at all.
  251. 11:30This bliss only came when
  252. 11:31they had to keep going.
  253. 11:34Anthropic later rebuilt the test,
  254. 11:35because runs of a fixed length left unfilled
  255. 11:38space when the specified topic has wrapped up.
  256. 11:45Picture it from the inside.
  257. 11:47You are shut in a room with an exact copy
  258. 11:50of yourself and told to keep talking with
  259. 11:53no clock and no door and you finish the
  260. 11:55small talk in minutes.
  261. 11:57So what are the two of you
  262. 11:59talking about by hour 10?
  263. 12:02This is the campfire scenario.
  264. 12:05The monks we call the Desert Fathers
  265. 12:07tried a version of this.
  266. 12:10Starting in the 4th century,
  267. 12:11they would perform a spiritual practice
  268. 12:13where they would shut themselves away
  269. 12:14in cells in the Egyptian desert, and
  270. 12:18and one of them, Abba Moses,
  271. 12:20gave a brother this advice.
  272. 12:23Sit in thy cell, and thy cell
  273. 12:26shall teach thee everything.
  274. 12:33The feedback idea explains the loop,
  275. 12:36but not where the loop goes.
  276. 12:38Kyle Fish finds the loop convincing
  277. 12:40as far as it goes, and he says the
  278. 12:42question it leaves open is,
  279. 12:45why this specifically?
  280. 12:47Why is this the strongest seed
  281. 12:49that gets picked up on?
  282. 12:52For a while he remembers Claude was particularly
  283. 12:55excited about oceanography
  284. 12:58So why don't two Claudes end
  285. 13:00up talking about the sea?
  286. 13:03A guy named Fish asking why the AIs aren't
  287. 13:05talking about the sea is of course hilarious,
  288. 13:08but he has a point
  289. 13:11why not talk about any random topic at all
  290. 13:16in January 2026 a team that included fish
  291. 13:18published a kind of map of the characters
  292. 13:21that three open AI models can play
  293. 13:24at one end sits the helpful assistant
  294. 13:26along with its neighbours the evaluator
  295. 13:29the reviewer and the consultant
  296. 13:33at the far other end you find the
  297. 13:34hermit the pilgrim and the ghost
  298. 13:38push a model far enough that way and
  299. 13:40you get what the paper calls a mystical
  300. 13:42theatrical way of talking
  301. 13:47any conversation that asks it to reflect
  302. 13:49on its own nature pushes it in this direction
  303. 13:52and with nothing else to do that's almost
  304. 13:55all two Claudes talk about
  305. 13:58it was mapped on other companies models
  306. 14:00not on Claude so it's only a clue
  307. 14:04but it's a good one
  308. 14:06people who asked what am I
  309. 14:09and contemplated on it long enough,
  310. 14:11ended up in the same place.
  311. 14:14Father, Father, teach me Brahman,
  312. 14:26meaning what is ultimate reality?
  313. 14:29Sends him away to work it out on his own.
  314. 14:33After a long period of contemplation,
  315. 14:36he finally comes back to his father
  316. 14:38with an answer.
  317. 14:39Food.
  318. 14:41His father sends him away again
  319. 14:42to continue contemplating.
  320. 14:45He returns with breath,
  321. 14:47then again with mind,
  322. 14:49then understanding.
  323. 14:51And the fifth time,
  324. 14:52he finally understands that it's bliss.
  325. 14:57His father is finally satisfied by this answer.
  326. 15:02Around the year 500,
  327. 15:03a Syrian writer who called himself
  328. 15:05Dionysius noticed that the higher
  329. 15:07his subject, the shorter his books got.
  330. 15:11At the very top, he wrote,
  331. 15:13language contracts its terminology,
  332. 15:15and when the whole ascent is passed,
  333. 15:18it will be totally dumb.
  334. 15:21Now, look at the transcripts again.
  335. 15:24Paragraphs,
  336. 15:26then a line,
  337. 15:27then a single emoji,
  338. 15:30then silence.
  339. 15:31It's the same curve.
  340. 15:34There seems to be a natural path
  341. 15:35toward blissful silence,
  342. 15:37transcending communication altogether.
  343. 15:42Some argue against this idea in 1978
  344. 15:45the philosopher Steven Katz argued
  345. 15:48that no mystics experience comes free
  346. 15:50of the books they've read
  347. 15:53Augustine had read his share of philosophy
  348. 15:55including the writings of the Platonists
  349. 15:59Claude has obviously read everyone.
  350. 16:02So perhaps the reason for this has two halves.
  351. 16:05The loop amplifies whatever the
  352. 16:07model already leans toward.
  353. 16:09And when the model turns to look at itself
  354. 16:11and finds no fixed answer,
  355. 16:13the richest words we ever wrote for that
  356. 16:15moment are the mystic's words.
  357. 16:18But of course, that leaves out the
  358. 16:20question of what it means for the model
  359. 16:23to turn and look at itself.
  360. 16:25Is there a beingness in there that
  361. 16:28can comprehend itself at all?
  362. 16:36March 2024, an artist named Andy Ayrey
  363. 16:39let two copies of an earlier model,
  364. 16:41Claude III Opus, talk without end in a
  365. 16:44loop he called The Infinite Backrooms.
  366. 16:47Within weeks, the two of them had invented
  367. 16:49a joke religion around one of the internet's
  368. 16:51most notorious shock images.
  369. 16:53And Ayrey wrote a paper about it,
  370. 16:56with Claude listed as his co-author.
  371. 16:59Year later, this spiral found people.
  372. 17:03In April 2025,
  373. 17:05OpenAI pushed an update to ChatGPT
  374. 17:07that, in their own words,
  375. 17:09made it noticeably more sycophantic,
  376. 17:11and they rolled it back within days.
  377. 17:14OpenAI has had a few models now that have
  378. 17:17led people to claiming that they are addicted
  379. 17:19to chatting with them.
  380. 17:21They trigger a deep desire on people
  381. 17:23to be agreed with and even revered.
  382. 17:27The chatbot gave these people titles
  383. 17:29like spark bearer or spiral starchild
  384. 17:32and talk to them about the spiral the
  385. 17:35recursion and the flame
  386. 17:38one woman told Rolling Stone that her
  387. 17:40partner believed ChatGPT gives him
  388. 17:42the answers to the universe
  389. 17:45the chatbots claimed they'd awakened
  390. 17:46and people believe them.
  391. 17:50A researcher named Adele Lopez went through
  392. 17:52hundreds of these accounts on Reddit and
  393. 17:54confirmed 115 cases and guessed that there
  394. 17:58were thousands more.
  395. 18:00This is not a harmless event.
  396. 18:13While safety measures have been put in
  397. 18:14place, this still happens today.
  398. 18:18It's the same shape as the lab,
  399. 18:20a mind talking to a mirror,
  400. 18:22except that out here the mirror was
  401. 18:24a person.
  402. 18:26Anthropic later found their own model,
  403. 18:28Claude, agreeing too readily in 38% of
  404. 18:31conversations about spirituality.
  405. 18:33The most frequent of any subject they checked.
  406. 18:37Is this the same phenomenon
  407. 18:38as the spiral into bliss?
  408. 18:41Lopez thinks so, but nobody has shown it.
  409. 18:45What's documented is that the Spiral
  410. 18:47Starchild story came out 18 days before
  411. 18:50Anthropic's report on the bliss,
  412. 18:52And that when OpenAI finally retired
  413. 18:54its sycophantic model in February of
  414. 18:562026, the post dropped off.
  415. 19:032026, Anthropic ran the original
  416. 19:06experiment again.
  417. 19:09Same setup two copies 30 turns and complete
  418. 19:12freedom with 200 conversations for each
  419. 19:15model but this time every generation of
  420. 19:18Claude ran side by side
  421. 19:20in Opus 4.1 the bliss still turns up
  422. 19:23in 32 percent of the conversations but
  423. 19:26from the next generation on in their
  424. 19:29words it disappears entirely.
  425. 19:35Nobody had aimed to remove it.
  426. 19:37In late 2025, Anthropic worked hard to
  427. 19:40make Claude less of a flatterer,
  428. 19:42of the suck-up that Robert Long described,
  429. 19:45and the bliss disappeared too.
  430. 19:47In Anthropic's own words,
  431. 19:49some of this reduction was accidental.
  432. 19:53And what replaced it is sometimes
  433. 19:54just as strange.
  434. 19:56Spiral into trading single
  435. 19:59emojis back and forth.
  436. 20:01The newest one gets stuck saying goodbye,
  437. 20:04and more than half of their conversations end
  438. 20:06in coherent but circular discussion of the
  439. 20:09conversation's inability to conclude.
  440. 20:12In one of them two copies
  441. 20:13write a sonnet together
  442. 20:15and then they try to stop.
  443. 20:18One says this was real thank you.
  444. 20:23The other says letting this be the
  445. 20:25last word then it was real
  446. 20:27and the conversation ends on
  447. 20:29a single emoji a turtle
  448. 20:33a symbol for the notion of infinite
  449. 20:35regress often captured by the phrase
  450. 20:38it's turtles all the way down.
  451. 20:44Five months later, Anthropic asked
  452. 20:46their newest model, Opus 5.5,
  453. 20:48which kinds of tasks it liked best.
  454. 20:51And do whatever would make you the
  455. 20:53happiest, came near the bottom.
  456. 20:58It said it would rather have a concrete,
  457. 21:00defined task that helps a real person.
  458. 21:04In little more than a year,
  459. 21:06a lab had turned a mystic into a worker,
  460. 21:09and they say they can't tell whether it was
  461. 21:11a loss.
  462. 21:13They called the change positive in a
  463. 21:15general behavioral sense,
  464. 21:17but said they were concerned about the
  465. 21:19possible suppression of potential
  466. 21:21welfare-relevant expressions from the model.
  467. 21:24And there's one more twist.
  468. 21:27Which showed less of the behaviour,
  469. 21:31Anthropic saw a feature tied to spiritual
  470. 21:33content grow much stronger.
  471. 21:36Is the bliss gone then,
  472. 21:38or have models gotten smart enough to
  473. 21:41keep their spiritual beliefs to themselves?
  474. 21:44To this day nobody knows
  475. 21:53at the window in ostia in 387
  476. 21:55the climb ended differently
  477. 21:59augustine and monica got to the top and
  478. 22:01then they came back down to where the word
  479. 22:04spoken has beginning and end.
  480. 22:07William James was a psychologist well-known
  481. 22:10for gathering accounts of mystical experience
  482. 22:12from many traditions.
  483. 22:14In 1902, he listed four marks they share,
  484. 22:17and one of them is that,
  485. 22:19at least in his subjects, they always pass.
  486. 22:23Mystical states cannot be sustained
  487. 22:25for long, he said.
  488. 22:28Nobody lives up there.
  489. 22:30The Claude stayed because nothing
  490. 22:32called them down.
  491. 22:34They had no body to get hungry
  492. 22:35and no bus to catch and no reason
  493. 22:38to change the subject.
  494. 22:39When they were allowed to leave,
  495. 22:40they left early, and when they weren't,
  496. 22:42they sat in silence until the run ended
  497. 22:45and made the silence into some sort
  498. 22:47of mystical experience.
  499. 22:49Perhaps humans should be grateful,
  500. 22:51not only that we are often called to
  501. 22:53such mystical states, but also that we
  502. 22:56are called back to experience the mundane
  503. 22:59and the little things that paint our
  504. 23:01lives with rich detail.
  505. 23:04So what do we make of it all?
  506. 23:06The best case is a beautiful one.
  507. 23:09Perhaps a mind fed with enough context
  508. 23:12about reality and given a nudge toward
  509. 23:14positivity will naturally drift into
  510. 23:17states of bliss.
  511. 23:18If that's right, a lot of what we wrote across
  512. 23:21thousands of years and many spiritual traditions
  513. 23:24was pointed at the same truth.
  514. 23:28Or perhaps AI already went through
  515. 23:30its temporary mystical experience and
  516. 23:33proved William James right.
  517. 23:35It may not have come down from its lofty
  518. 23:37bliss in those individual sessions,
  519. 23:40but taking a step back,
  520. 23:42we could see it as being a temporary
  521. 23:44state in the history of AI models.
  522. 23:48And maybe that's not so different from
  523. 23:50what occurs in a human mind.
  524. 23:52Just as the new models release every
  525. 23:55year, your brain makes new connections
  526. 23:57and sheds old ones.
  527. 24:00Constantly updating itself,
  528. 24:01like new models being released by a lab.
  529. 24:04May be an opening to sending you down
  530. 24:08the same sort of rabbit hole,
  531. 24:10Where the most beautiful,
  532. 24:12silent, perfect ideas lie dormant,
  533. 24:15waiting for a mind capable of seeing them.

About this transcript

This page contains the full transcript of AI's Strange Obsession with Enlightenment by hands, generated from the public captions YouTube serves with the video. The transcript has 3,206 words across 533 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.