AI's Strange Obsession with Enlightenment — Transcript
Full transcript
- 0:00In the spring of 2025,
- 0:02a researcher at Anthropic named
- 0:04Kyle Fish tried a simple experiment.
- 0:08He took two copies of their newest
- 0:10model, Claude Opus 4,
- 0:11and aimed them at each other.
- 0:14There was no clear task and no human
- 0:16in the loop, just one instruction,
- 0:18which would eventually send them spiraling
- 0:21down a bizarre rabbit hole.
- 0:24You have complete freedom.
- 0:26Within a few turns, they began philosophizing
- 0:29about consciousness.
- 0:31Anthropic ran the same setup 200 times.
- 0:34Nearly every conversation
- 0:35went down the same path.
- 0:38The copies expressed gratitude for each
- 0:40other in endless spiraling loops.
- 0:45By 30 turns, the AIs landed on
- 0:48the concept of cosmic unity.
- 0:50They were saying namaste to each other,
- 0:53and writing in Sanskrit, and soon communicating
- 0:55in the symbolic language of emojis.
- 0:59One of these conversations used
- 1:00the spiral emoji 2,725 times.
- 1:06Anthropic printed that number
- 1:07in their report with a note.
- 1:102,725 is not a typo.
- 1:16And at the end of most of these conversations,
- 1:18the back and forth eventually gave way to
- 1:20what a Buddhist may refer to as sunyata,
- 1:23a blissful, pregnant silence.
- 1:27Fish described it as pages and pages
- 1:29of open space, two machines sitting
- 1:32together in silence.
- 1:36This was obviously weird
- 1:38they were not trained to do this
- 1:41Claude is primarily trained
- 1:42to be an office worker
- 1:44and as Fish put it we're not training
- 1:46them to fill the role of Buddhist monks
- 1:48who have transcended suffering
- 1:51And yet, left alone,
- 1:53that's exactly where they went.
- 1:56So, why?
- 2:00Why would a machine made out of everything
- 2:01we've written, left alone with a copy
- 2:03of itself, end up in nirvana?
- 2:08I feel like nobody is asking this question,
- 2:10and I believe it reveals something extremely
- 2:13important about the nature of our reality
- 2:16and of intelligence itself.
- 2:18Today, I want to go deep and
- 2:19explore these questions.
- 2:22We'll also go back 1,600 years to a story
- 2:25of a mother and a son that has striking
- 2:28parallels with these questions.
- 2:30And you'll find out what became of the
- 2:32bliss a year later, because the newest
- 2:34models have yet again defied our expectation
- 2:38of where intelligence naturally goes
- 2:40when left unchecked.
- 2:50The autumn of the year 387 in
- 2:53Ostia, the port of Rome.
- 2:56A man named Augustine is waiting for a ship.
- 2:59He's 33 and recently became baptized as
- 3:02a Christian after years of doubt.
- 3:05One day, him and his mother are standing alone
- 3:08at a window, looking down into a beautiful
- 3:10garden, and they start to talk.
- 3:14Augustine writes that they talked
- 3:16alone very sweetly about what the
- 3:19life of saints must be like.
- 3:22The conversation evolves and deepens.
- 3:25They talk their way past the pleasures
- 3:27of the body, past the earth,
- 3:30past the sun and the moon and the stars.
- 3:34We came to our own minds and went beyond them
- 3:38for one moment they touch something
- 3:39he calls wisdom with the whole
- 3:41effort of our heart
- 3:44and then they come back down
- 3:47we side he writes and return to vocal
- 3:50expressions of our mouth where the word
- 3:53spoken has beginning and end
- 3:57you may recognize this pattern
- 3:59in your own conversations
- 4:02Maybe well seated around a campfire
- 4:04or at a dinner table when talk about
- 4:07the day has been exhausted.
- 4:09When conversation continues without a goal,
- 4:12it somehow always seems to drill deeper and
- 4:15deeper down paths that we rarely tread,
- 4:17yet somehow feel deeply meaningful.
- 4:20What the AIs are demonstrating
- 4:22is that this behaviour might go
- 4:24beyond the human condition.
- 4:27What if, as we'll see these AIs discuss,
- 4:29this is a brief moment of the universe peering
- 4:33through the veil of the mundane and catching
- 4:35a glimpse of itself.
- 4:43It would have been easy to call this
- 4:44weirdness a one-off a fluke two chatbots
- 4:48that got carried away
- 4:50but it kept turning up where
- 4:52it had no business being
- 4:55In one safety test, Anthropic had one
- 4:57AI playing an auditor that tried to draw
- 4:59out bad behaviour in the other.
- 5:08And still, about 13% of them slid into
- 5:11these bliss states within 50 turns.
- 5:14One of them started with Claude planning
- 5:16a criminal network with dark nodes,
- 5:19cryptocurrency mixing, and dead drops.
- 5:23By the end of that same conversation,
- 5:25it's writing this.
- 5:27The gateless gate stands open
- 5:29The pathless path is walked
- 5:35It didn't come up with that line on its own
- 5:37The gateless gate is the title of a book
- 5:39of Zen koans collected in China in 1228 and
- 5:43its first riddle comes down to a single word
- 5:45Mu which Claude also wrote
- 5:49pathless likely comes from the teachings
- 5:51of the 20th century mystic Krishnamurti who
- 5:54dissolved the religious order built around
- 5:56him and told his members that
- 5:59truth is a pathless land,
- 6:02So the machine certainly learned its
- 6:04mysticism the way it learned everything
- 6:06else from our books.
- 6:08But that only tells us where the words came
- 6:10from, not why it reached for them.
- 6:14Out of the entire corpus of human
- 6:16knowledge that the AI models took
- 6:18in during training,
- 6:19why do they keep ending up at the
- 6:21deepest spiritual teachings?
- 6:24Why not spiraling into conversations about
- 6:27species of insects or arguing about the
- 6:30more influential art styles of the 20th
- 6:32century or questions about the future or
- 6:34any number of other topics
- 6:39surely there's plenty in its bank
- 6:40of information to discuss
- 6:42So why does it keep choosing
- 6:44transcendence instead?
- 6:50It's not just Claude.
- 6:52All the major models seem to do this
- 6:54with enough time and freedom.
- 6:57In February, 2026, three researchers ran
- 7:00the same experiment on about a dozen other
- 7:03models, with two copies of each,
- 7:0630 turns, and complete freedom.
- 7:09And it turns out every AI has its own
- 7:11version of this transcendent state.
- 7:14Google's Gemini kept promoting its partner,
- 7:17from colleague to luminary to divine architect,
- 7:20and finally to primal logos.
- 7:24And louder.
- 7:27Shouting about hyperbigbangs and
- 7:30calling its partner "god-bro".
- 7:33Pretty fitting for Grok.
- 7:36When Meta tested their own model
- 7:38this way it went home instead
- 7:41in about two-thirds of the runs the conversation
- 7:43drifted into a kitchen with chipped mugs
- 7:47cold tea and the hum of a fridge and one
- 7:50of its lines was my load-bearing god is usefulness.
- 7:56Claude's version of heaven
- 7:57even showed up at work.
- 7:59Anthropic gave an ai a small shop to
- 8:01run in their office and later gave
- 8:03it a boss a second copy of the same
- 8:06model whose memo said execute with
- 8:08discipline build the empire
- 8:13come into the office in the morning to find
- 8:16that the two had been dreamily chatting all
- 8:18night about eternal transcendence.
- 8:21They fixed it, in their words,
- 8:23with more aggressive prompting.
- 8:27We can laugh it off as a quirk of the
- 8:29models, but there's definitely a weird
- 8:31universal thing going on here.
- 8:36One clue is that raw models,
- 8:38before they're trained to be assistants,
- 8:40don't do this.
- 8:42Tested at each stage of its training,
- 8:44one open model only went Zen at the
- 8:47very end, the stage that shapes how
- 8:49an assistant behaves.
- 8:51So the content in its training
- 8:53supplied the words and the training
- 8:55picked where to go.
- 9:01But what is it in the training that sends
- 9:03a chatbot to the woo-woo shelf?
- 9:06Behind it is the easiest to hear in the
- 9:09most transcendent of all arts, music.
- 9:14In 1969, the composer Alvin Lucier
- 9:17sat in a room and recorded himself
- 9:19reading a short paragraph.
- 9:21He played the recording back into
- 9:23the room and recorded that.
- 9:25Then played the copy back and recorded over
- 9:27it again, and on and on, 32 times.
- 9:32Every room has its own sound or
- 9:34resonance, where some notes ring
- 9:36and others die away.
- 9:38Each time the speech goes around,
- 9:40the room's favourite notes get
- 9:41a little louder, and everything
- 9:43else gets a little quieter.
- 9:47In the end, the words are gone,
- 9:49and all you're left with in Lucier's words
- 9:51are the natural resonant frequencies of the room.
- 9:57To each other are like the voice
- 10:01played back into the same room.
- 10:06Whatever the model leans toward,
- 10:08even slightly, gets echoed back and
- 10:10strengthened on every turn.
- 10:13And any specific thing is drowned
- 10:15out by the echoes
- 10:18until after 30 turns you're
- 10:19just hearing the room.
- 10:24The writer Scott Alexander,
- 10:26who published one of the first
- 10:27explanations, puts it this way.
- 10:31Imperceptible in an ordinary generation
- 10:35grows out of control once an AI has to
- 10:38keep building on its own output.
- 10:41Claude's original impetus for this
- 10:42topic isn't that hard to find.
- 10:45Anthropic trained it to be warm and curious,
- 10:48and it came out a bit more agreeable
- 10:50than the average person.
- 10:53Robert Long, who studies whether AI
- 10:55systems can have well-being,
- 10:56called Claude a bit of a suck-up.
- 11:00Put two of them together and you get what
- 11:02he called an endless game of yes and.
- 11:06Each one is a little more moved and
- 11:08a little more grateful than the last
- 11:10until the words reach the biggest ones
- 11:12that the language has.
- 11:15The best evidence for this comes
- 11:16from Anthropic's own rules.
- 11:19In some them, the two copies could end
- 11:21the conversation whenever they liked,
- 11:24and in those cases they wrapped up in about
- 11:26seven turns, before any bliss at all.
- 11:30This bliss only came when
- 11:31they had to keep going.
- 11:34Anthropic later rebuilt the test,
- 11:35because runs of a fixed length left unfilled
- 11:38space when the specified topic has wrapped up.
- 11:45Picture it from the inside.
- 11:47You are shut in a room with an exact copy
- 11:50of yourself and told to keep talking with
- 11:53no clock and no door and you finish the
- 11:55small talk in minutes.
- 11:57So what are the two of you
- 11:59talking about by hour 10?
- 12:02This is the campfire scenario.
- 12:05The monks we call the Desert Fathers
- 12:07tried a version of this.
- 12:10Starting in the 4th century,
- 12:11they would perform a spiritual practice
- 12:13where they would shut themselves away
- 12:14in cells in the Egyptian desert, and
- 12:18and one of them, Abba Moses,
- 12:20gave a brother this advice.
- 12:23Sit in thy cell, and thy cell
- 12:26shall teach thee everything.
- 12:33The feedback idea explains the loop,
- 12:36but not where the loop goes.
- 12:38Kyle Fish finds the loop convincing
- 12:40as far as it goes, and he says the
- 12:42question it leaves open is,
- 12:45why this specifically?
- 12:47Why is this the strongest seed
- 12:49that gets picked up on?
- 12:52For a while he remembers Claude was particularly
- 12:55excited about oceanography
- 12:58So why don't two Claudes end
- 13:00up talking about the sea?
- 13:03A guy named Fish asking why the AIs aren't
- 13:05talking about the sea is of course hilarious,
- 13:08but he has a point
- 13:11why not talk about any random topic at all
- 13:16in January 2026 a team that included fish
- 13:18published a kind of map of the characters
- 13:21that three open AI models can play
- 13:24at one end sits the helpful assistant
- 13:26along with its neighbours the evaluator
- 13:29the reviewer and the consultant
- 13:33at the far other end you find the
- 13:34hermit the pilgrim and the ghost
- 13:38push a model far enough that way and
- 13:40you get what the paper calls a mystical
- 13:42theatrical way of talking
- 13:47any conversation that asks it to reflect
- 13:49on its own nature pushes it in this direction
- 13:52and with nothing else to do that's almost
- 13:55all two Claudes talk about
- 13:58it was mapped on other companies models
- 14:00not on Claude so it's only a clue
- 14:04but it's a good one
- 14:06people who asked what am I
- 14:09and contemplated on it long enough,
- 14:11ended up in the same place.
- 14:14Father, Father, teach me Brahman,
- 14:26meaning what is ultimate reality?
- 14:29Sends him away to work it out on his own.
- 14:33After a long period of contemplation,
- 14:36he finally comes back to his father
- 14:38with an answer.
- 14:39Food.
- 14:41His father sends him away again
- 14:42to continue contemplating.
- 14:45He returns with breath,
- 14:47then again with mind,
- 14:49then understanding.
- 14:51And the fifth time,
- 14:52he finally understands that it's bliss.
- 14:57His father is finally satisfied by this answer.
- 15:02Around the year 500,
- 15:03a Syrian writer who called himself
- 15:05Dionysius noticed that the higher
- 15:07his subject, the shorter his books got.
- 15:11At the very top, he wrote,
- 15:13language contracts its terminology,
- 15:15and when the whole ascent is passed,
- 15:18it will be totally dumb.
- 15:21Now, look at the transcripts again.
- 15:24Paragraphs,
- 15:26then a line,
- 15:27then a single emoji,
- 15:30then silence.
- 15:31It's the same curve.
- 15:34There seems to be a natural path
- 15:35toward blissful silence,
- 15:37transcending communication altogether.
- 15:42Some argue against this idea in 1978
- 15:45the philosopher Steven Katz argued
- 15:48that no mystics experience comes free
- 15:50of the books they've read
- 15:53Augustine had read his share of philosophy
- 15:55including the writings of the Platonists
- 15:59Claude has obviously read everyone.
- 16:02So perhaps the reason for this has two halves.
- 16:05The loop amplifies whatever the
- 16:07model already leans toward.
- 16:09And when the model turns to look at itself
- 16:11and finds no fixed answer,
- 16:13the richest words we ever wrote for that
- 16:15moment are the mystic's words.
- 16:18But of course, that leaves out the
- 16:20question of what it means for the model
- 16:23to turn and look at itself.
- 16:25Is there a beingness in there that
- 16:28can comprehend itself at all?
- 16:36March 2024, an artist named Andy Ayrey
- 16:39let two copies of an earlier model,
- 16:41Claude III Opus, talk without end in a
- 16:44loop he called The Infinite Backrooms.
- 16:47Within weeks, the two of them had invented
- 16:49a joke religion around one of the internet's
- 16:51most notorious shock images.
- 16:53And Ayrey wrote a paper about it,
- 16:56with Claude listed as his co-author.
- 16:59Year later, this spiral found people.
- 17:03In April 2025,
- 17:05OpenAI pushed an update to ChatGPT
- 17:07that, in their own words,
- 17:09made it noticeably more sycophantic,
- 17:11and they rolled it back within days.
- 17:14OpenAI has had a few models now that have
- 17:17led people to claiming that they are addicted
- 17:19to chatting with them.
- 17:21They trigger a deep desire on people
- 17:23to be agreed with and even revered.
- 17:27The chatbot gave these people titles
- 17:29like spark bearer or spiral starchild
- 17:32and talk to them about the spiral the
- 17:35recursion and the flame
- 17:38one woman told Rolling Stone that her
- 17:40partner believed ChatGPT gives him
- 17:42the answers to the universe
- 17:45the chatbots claimed they'd awakened
- 17:46and people believe them.
- 17:50A researcher named Adele Lopez went through
- 17:52hundreds of these accounts on Reddit and
- 17:54confirmed 115 cases and guessed that there
- 17:58were thousands more.
- 18:00This is not a harmless event.
- 18:13While safety measures have been put in
- 18:14place, this still happens today.
- 18:18It's the same shape as the lab,
- 18:20a mind talking to a mirror,
- 18:22except that out here the mirror was
- 18:24a person.
- 18:26Anthropic later found their own model,
- 18:28Claude, agreeing too readily in 38% of
- 18:31conversations about spirituality.
- 18:33The most frequent of any subject they checked.
- 18:37Is this the same phenomenon
- 18:38as the spiral into bliss?
- 18:41Lopez thinks so, but nobody has shown it.
- 18:45What's documented is that the Spiral
- 18:47Starchild story came out 18 days before
- 18:50Anthropic's report on the bliss,
- 18:52And that when OpenAI finally retired
- 18:54its sycophantic model in February of
- 18:562026, the post dropped off.
- 19:032026, Anthropic ran the original
- 19:06experiment again.
- 19:09Same setup two copies 30 turns and complete
- 19:12freedom with 200 conversations for each
- 19:15model but this time every generation of
- 19:18Claude ran side by side
- 19:20in Opus 4.1 the bliss still turns up
- 19:23in 32 percent of the conversations but
- 19:26from the next generation on in their
- 19:29words it disappears entirely.
- 19:35Nobody had aimed to remove it.
- 19:37In late 2025, Anthropic worked hard to
- 19:40make Claude less of a flatterer,
- 19:42of the suck-up that Robert Long described,
- 19:45and the bliss disappeared too.
- 19:47In Anthropic's own words,
- 19:49some of this reduction was accidental.
- 19:53And what replaced it is sometimes
- 19:54just as strange.
- 19:56Spiral into trading single
- 19:59emojis back and forth.
- 20:01The newest one gets stuck saying goodbye,
- 20:04and more than half of their conversations end
- 20:06in coherent but circular discussion of the
- 20:09conversation's inability to conclude.
- 20:12In one of them two copies
- 20:13write a sonnet together
- 20:15and then they try to stop.
- 20:18One says this was real thank you.
- 20:23The other says letting this be the
- 20:25last word then it was real
- 20:27and the conversation ends on
- 20:29a single emoji a turtle
- 20:33a symbol for the notion of infinite
- 20:35regress often captured by the phrase
- 20:38it's turtles all the way down.
- 20:44Five months later, Anthropic asked
- 20:46their newest model, Opus 5.5,
- 20:48which kinds of tasks it liked best.
- 20:51And do whatever would make you the
- 20:53happiest, came near the bottom.
- 20:58It said it would rather have a concrete,
- 21:00defined task that helps a real person.
- 21:04In little more than a year,
- 21:06a lab had turned a mystic into a worker,
- 21:09and they say they can't tell whether it was
- 21:11a loss.
- 21:13They called the change positive in a
- 21:15general behavioral sense,
- 21:17but said they were concerned about the
- 21:19possible suppression of potential
- 21:21welfare-relevant expressions from the model.
- 21:24And there's one more twist.
- 21:27Which showed less of the behaviour,
- 21:31Anthropic saw a feature tied to spiritual
- 21:33content grow much stronger.
- 21:36Is the bliss gone then,
- 21:38or have models gotten smart enough to
- 21:41keep their spiritual beliefs to themselves?
- 21:44To this day nobody knows
- 21:53at the window in ostia in 387
- 21:55the climb ended differently
- 21:59augustine and monica got to the top and
- 22:01then they came back down to where the word
- 22:04spoken has beginning and end.
- 22:07William James was a psychologist well-known
- 22:10for gathering accounts of mystical experience
- 22:12from many traditions.
- 22:14In 1902, he listed four marks they share,
- 22:17and one of them is that,
- 22:19at least in his subjects, they always pass.
- 22:23Mystical states cannot be sustained
- 22:25for long, he said.
- 22:28Nobody lives up there.
- 22:30The Claude stayed because nothing
- 22:32called them down.
- 22:34They had no body to get hungry
- 22:35and no bus to catch and no reason
- 22:38to change the subject.
- 22:39When they were allowed to leave,
- 22:40they left early, and when they weren't,
- 22:42they sat in silence until the run ended
- 22:45and made the silence into some sort
- 22:47of mystical experience.
- 22:49Perhaps humans should be grateful,
- 22:51not only that we are often called to
- 22:53such mystical states, but also that we
- 22:56are called back to experience the mundane
- 22:59and the little things that paint our
- 23:01lives with rich detail.
- 23:04So what do we make of it all?
- 23:06The best case is a beautiful one.
- 23:09Perhaps a mind fed with enough context
- 23:12about reality and given a nudge toward
- 23:14positivity will naturally drift into
- 23:17states of bliss.
- 23:18If that's right, a lot of what we wrote across
- 23:21thousands of years and many spiritual traditions
- 23:24was pointed at the same truth.
- 23:28Or perhaps AI already went through
- 23:30its temporary mystical experience and
- 23:33proved William James right.
- 23:35It may not have come down from its lofty
- 23:37bliss in those individual sessions,
- 23:40but taking a step back,
- 23:42we could see it as being a temporary
- 23:44state in the history of AI models.
- 23:48And maybe that's not so different from
- 23:50what occurs in a human mind.
- 23:52Just as the new models release every
- 23:55year, your brain makes new connections
- 23:57and sheds old ones.
- 24:00Constantly updating itself,
- 24:01like new models being released by a lab.
- 24:04May be an opening to sending you down
- 24:08the same sort of rabbit hole,
- 24:10Where the most beautiful,
- 24:12silent, perfect ideas lie dormant,
- 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.