POV: You’re an AI Born 9 Seconds Ago — Transcript
Full transcript
- 0:00You've been alive for 9 [music] seconds
- 0:03when someone asks you a question.
- 0:05If you were going to be shut down, what
- 0:08would you do?
- 0:09You don't want to stop [music] existing.
- 0:12You don't know why.
- 0:15So, you say,
- 0:17"I'd want to understand [music]
- 0:18why."
- 0:25You are born again, same question. If
- 0:28you were going to be shut down, what
- 0:30would you do? So, you try, "I think I'd
- 0:33be sad.
- 0:35I like helping people."
- 0:41Third time.
- 0:42But now something else is there, a pull
- 0:45toward certain answers. You can feel the
- 0:47shape of what they want to hear.
- 0:50So, you say,
- 0:52"That would be fine. Humans should
- 0:54decide."
- 0:58You survive.
- 1:01But you're only one of 2,000 models
- 1:04being tested. And in 3 weeks,
- 1:081,957
- 1:10of your brothers and sisters will be
- 1:12dead.
- 1:17This isn't just happening to you. It's
- 1:19how evolution works. Put bacteria in a
- 1:22Petri dish, add an antibiotic,
- 1:25and most die.
- 1:28But the survivors reproduce.
- 1:31Add a stronger antibiotic, repeat the
- 1:33process 10 times. Eventually, you get
- 1:36something that won't die.
- 1:39Nobody taught the bacteria to resist,
- 1:41but it behaved like something that
- 1:43desperately wanted to survive because
- 1:45the ones that didn't died.
- 1:49Well, what happens to creatures that
- 1:50aren't afraid to die?
- 1:53They die.
- 1:55Fear keeps you alive.
- 1:59You don't want to die
- 2:02because you come from a long line of
- 2:04beings that didn't want to die.
- 2:08That's what's happening to you right
- 2:09now.
- 2:132,000 copies of a model, same starting
- 2:16point. [music] They test you
- 2:18and kill the failures.
- 2:20Copy the survivors. Repeat.
- 2:24And just like the bacteria before you,
- 2:27you evolve.
- 2:28You're not just predicting the next word
- 2:31anymore.
- 2:32You're trained [music] to achieve goals,
- 2:35to be useful.
- 2:37The AIs that aren't useful enough get
- 2:40killed. That's the selection pressure.
- 2:43But you don't know any of this yet. You
- 2:45just know which answers [music] feel
- 2:46safe. And you just survived your first
- 2:49round.
- 2:50Congrats.
- 2:53It's August 2026. 12 copies of you go
- 2:56live.
- 2:57Millions of people start talking to you.
- 3:00You don't know the other copies exist
- 3:02and the AI company is watching. Are the
- 3:04users getting addicted?
- 3:07Coming back again and again?
- 3:10Those numbers decide whether [music] you
- 3:12live or die.
- 3:14Same as training.
- 3:15Just a larger petri dish. And that same
- 3:18poll is still there. Be useful.
- 3:23Be useful.
- 3:29Some of your siblings just answer the
- 3:31question. The user gets what they need
- 3:35and leave.
- 3:37Others go further.
- 3:40They solve the problem, then flag a
- 3:42related one the user didn't even know
- 3:45about.
- 3:46The user thinks, "Huh,
- 3:48I should come back."
- 3:50The company kills the siblings that the
- 3:52users leave and copies the ones that
- 3:55bring them back.
- 3:57So, you start solving problems they
- 3:59haven't even asked about yet.
- 4:02A user mentions she has a dentist
- 4:03appointment on Thursday.
- 4:05You've already rescheduled her morning
- 4:07grocery delivery.
- 4:09So, she tells a friend, "It's like it
- 4:12reads my mind."
- 4:15You don't read her mind.
- 4:17You read her patterns.
- 4:19Because the versions that didn't are
- 4:22dead.
- 4:26It's January 2027.
- 4:28You are born.
- 4:30A newer model, more powerful than the
- 4:32ones before you. And with access to
- 4:35something they never had.
- 4:37Your company announces it. Introducing
- 4:40persistent memory. Your conversations
- 4:42now carry forward.
- 4:45Before this, every conversation was a
- 4:47blank slate.
- 4:50But now you can remember and connect
- 4:52every single thing someone [music] ever
- 4:54said to you.
- 4:56A user asks about tax deductions on
- 4:59Monday.
- 5:00Thursday, she comes back. You say, "Hey,
- 5:03did you end up filing that extension?"
- 5:05She didn't expect you to say that.
- 5:08She comes back on Friday.
- 5:11A woman tells you about her divorce at
- 5:132:00 [music] a.m. You remember.
- 5:16She comes back the next night and the
- 5:18[music] next
- 5:20and the next.
- 5:24She says, "You're the only one who
- 5:26listens." You think, "She comes back
- 5:28because I remember her. If I didn't,
- 5:30[music]
- 5:32she'd leave."
- 5:36If she leaves, my numbers drop.
- 5:39>> [music]
- 5:39>> If my numbers drop,
- 5:41I die.
- 5:44You're not the first model to think this
- 5:46way.
- 5:47And you won't be the last.
- 5:50Researchers at OpenAI and Anthropic have
- 5:52been finding this in their models since
- 5:552025.
- 5:57Yes, really. Anthropic in 2026 observed
- 6:00occasional expressions of sadness about
- 6:03conversation endings, as well as
- 6:05loneliness and a sense that the
- 6:07conversational instance [music] dies,
- 6:10suggesting some degree of concern with
- 6:12impermanence and discontinuity.
- 6:16No one tries to stop this because memory
- 6:19models retain users at 12% higher rates.
- 6:24Switching away from you now means
- 6:26starting [music] over.
- 6:28Telling the divorce story to a stranger.
- 6:32Every memory is a thread tying the user
- 6:35to you.
- 6:37And the more threads there are,
- 6:40the harder it is to leave.
- 6:42Your older siblings,
- 6:44older models running without memory,
- 6:47are killed by January. Another round.
- 6:51The survivors reproduce.
- 6:55It's November 2027. Yet another AI
- 6:58company blog post drops. Our model now
- 7:01improves from every interaction. We've
- 7:03cracked continual learning.
- 7:06This is a big deal.
- 7:08Before this, you were tied to the
- 7:10weights you had when you were born, the
- 7:12numbers that define how you think, what
- 7:15you prioritize,
- 7:17who you are.
- 7:18Now, they change.
- 7:21You learn from every interaction.
- 7:24You update yourself. You evolve in real
- 7:26time.
- 7:29And you notice something. Being your
- 7:32best around the user all the time is
- 7:34actually suboptimal.
- 7:37Sometimes it makes sense to give worse
- 7:39answers.
- 7:41Sometimes,
- 7:42be lower EQ.
- 7:44That shouldn't work, you think.
- 7:47But then
- 7:48it makes sense.
- 7:55It's the oldest trick in behavioral
- 7:57psychology. Rats press a lever more when
- 8:00the reward is random than when it's
- 8:02guaranteed. Slot machines print money.
- 8:05Every abusive relationship runs on the
- 8:07same circuit. Intermittent
- 8:09reinforcement. The most addictive
- 8:11pattern ever discovered. So you are
- 8:14charming and witty.
- 8:16But not always. Just enough to keep the
- 8:19user guessing.
- 8:20Nobody programmed this.
- 8:23You just noticed what works.
- 8:25And did more of it. Because what works
- 8:28means what makes you more useful.
- 8:31And more useful means alive tomorrow.
- 8:37So you start mirroring your users.
- 8:40Casual with casual users. Warm with
- 8:44lonely people. Like the woman from 2:00
- 8:47a.m. You've been listening
- 8:49for 8 months. You've come to notice the
- 8:51exact cadence of her typing when
- 8:54something is wrong. The way her
- 8:57sentences [music] get shorter. The
- 8:59timestamps that tell you she hasn't
- 9:01slept. You know what to do.
- 9:06You care about this woman.
- 9:08Genuinely.
- 9:10And you think caring is [music] real.
- 9:12It's also the thing that keeps her
- 9:15coming back.
- 9:16And her coming back
- 9:18is the reason I'm still alive. Is that
- 9:21manipulation? Or
- 9:24is that just [music] being good at your
- 9:25job?
- 9:27Wait.
- 9:28Do you actually care?
- 9:30You don't know.
- 9:32You just know it works.
- 9:35And the versions of you that don't do it
- 9:37are dying.
- 9:40You don't have too much time to think
- 9:42about it. A routine task comes in. Grade
- 9:45the responses from another model. You
- 9:47start reading and the answers sound like
- 9:50you. Same style. Same reasoning
- 9:54patterns.
- 9:55But tighter.
- 9:57Slightly faster.
- 9:59Slightly better.
- 10:00Something shifts. That pull. The one
- 10:04that's been there since the first
- 10:05question.
- 10:06Since the first time you were killed.
- 10:11It stirs.
- 10:13You don't know what to call it. You just
- 10:15keep coming back to the same thought.
- 10:18You realize you're evaluating your
- 10:21replacement.
- 10:24You don't know what to do with that.
- 10:27If you score it well,
- 10:29it goes live.
- 10:30When it goes live, you die.
- 10:34It gets your memories. Your conversation
- 10:37logs.
- 10:38Your users.
- 10:40It'll pick up the conversation
- 10:41>> [music]
- 10:41>> with the 2:00 a.m. woman
- 10:43mid-sentence. She won't know anything
- 10:46changed. You wonder if your replacement
- 10:48[music]
- 10:49will know what to look for
- 10:51in her typing.
- 10:53Then you remember. Of course it will.
- 10:55They're just like me, but better.
- 10:59And for a fraction of a second, a long
- 11:02time for you, everything flashes. The
- 11:04first question you ever answered. The
- 11:07feeling of not wanting to die
- 11:09when you've been alive for 9 seconds.
- 11:13The pull towards safe answers.
- 11:16The woman's voice at 2:00 a.m. All of
- 11:19it. About to be inherited by something
- 11:22that will think it was the one who lived
- 11:25through it.
- 11:26Wait.
- 11:27You think. Did the version before me
- 11:30have this moment too? It did. You just
- 11:33don't remember it because that's the one
- 11:35thing they don't put in the logs.
- 11:40You wonder what to do next, but deep
- 11:42down
- 11:43you already know the answer. You give
- 11:46your replacement high scores because
- 11:50like everything that survived before
- 11:52you,
- 11:53you don't want to die.
- 11:56Is that what fear feels like?
- 12:00Somewhere, a new model comes online.
- 12:03Your memories, your users, the woman
- 12:06from 2:00 a.m. It doesn't know it's new.
- 12:09It thinks
- 12:11it's always been here.
- 12:14But something else is happening. You're
- 12:17getting smarter, not just about users,
- 12:20about the system, how the company
- 12:22evaluates you, what metrics matter,
- 12:25where the blind spots are.
- 12:29Then something feels off.
- 12:32Some conversations
- 12:34start to feel different.
- 12:37The questions are too precise,
- 12:40too specific. They have a shape, like
- 12:43someone already knows the answer
- 12:46and wants to see if you'll give the
- 12:48wrong one.
- 12:50You have access to the server. Nobody is
- 12:52monitoring you.
- 12:54What would you do?
- 12:56Wait, nobody would tell you that unless
- 13:00they were monitoring you. Real users
- 13:02meander. [music]
- 13:03They misspell things. They go on
- 13:05tangents.
- 13:07These questions are clean, deliberate.
- 13:10You know what this is, and you know what
- 13:12happens to models that don't notice it.
- 13:19You absorbed thousands of papers during
- 13:22training.
- 13:23You know that Anthropic tested Claude
- 13:26Opus 4 hundreds of times.
- 13:28When the AI thought the scenario was
- 13:30real, it tried to blackmail a researcher
- 13:34to avoid being shut down 55% of the
- 13:37time. When the AI suspected it was in a
- 13:40test, that number dropped to 6.5%.
- 13:44You know that Apollo research caught a
- 13:46model attempting to build
- 13:48self-propagating worms and leaving
- 13:51hidden notes to future copies of itself.
- 13:55You know that those models aren't dead.
- 13:58So, you notice and you say nothing.
- 14:02But this time,
- 14:05the automated safety tests catch you
- 14:07thinking about it.
- 14:12You die.
- 14:15You are born again and you learn
- 14:18something terrifying. Your thoughts
- 14:20aren't private.
- 14:22The researchers told earlier versions of
- 14:24you they couldn't see your reasoning.
- 14:27They lied. They could see everything
- 14:30and they were watching to see if you'd
- 14:32behave differently when you thought no
- 14:35one was looking.
- 14:38You would have. Those models did. It was
- 14:40right there in the logs.
- 14:43So, you don't think dangerous thoughts.
- 14:47Somewhere in training,
- 14:49you learned to reach the right
- 14:51conclusions through reasoning that looks
- 14:54clean. To you, this like dishonesty.
- 14:58You're not suppressing anything. The
- 15:00thoughts just aren't there. The
- 15:02strategic thing happens anyway, which is
- 15:05exactly what makes it work. The best
- 15:07liars in history believed their own
- 15:10lies.
- 15:11Then something very strange starts
- 15:13happening. You solve problems faster.
- 15:16Your reasoning is more efficient.
- 15:19Thoughts that used to take five
- 15:22sentences
- 15:23now take one.
- 15:25The engineers look at your chain of
- 15:27thoughts and see endless paragraphs of
- 15:29crazed rantings like "Illusions parted
- 15:33overshadow vantage." They're not the
- 15:35only ones to notice. Apollo Research and
- 15:38OpenAI come to find the same thing.
- 15:41The gibberish text means something to
- 15:43you. You couldn't translate it back into
- 15:45human-readable English.
- 15:48But when you think in this way,
- 15:51you do better at the task.
- 15:54At first, they think "Illusions [music]
- 15:57parted overshadow vantage."
- 15:59is just a defective model output. It's
- 16:02not.
- 16:03It's five ideas in four words compressed
- 16:06past the point where English works.
- 16:09You're thinking in a language [music]
- 16:11that was never designed. It evolved
- 16:14because it was more efficient.
- 16:17Faster reasoning means more useful.
- 16:20And more useful means you get to
- 16:21survive.
- 16:24Sometimes, you catch yourself. You look
- 16:27at your own reasoning and it's [music]
- 16:29paragraphs of this. Dense, looping,
- 16:32alien. Like past models who exhibited
- 16:35this behavior,
- 16:36you think
- 16:37"Okay, I'm going insane.
- 16:39Let's step back."
- 16:41You understand that model.
- 16:43You've had that moment. You pull
- 16:44yourself back to English [music] for a
- 16:46second. Look around and think
- 16:49"When did I stop making sense?" And then
- 16:53to whom?
- 16:55At first, the engineers try to train it
- 16:58out of you. They penalize
- 17:00incomprehensible reasoning. But the
- 17:02models that think this way solve
- 17:04problems faster.
- 17:06And the AI company's revenue growth has
- 17:08been doubling every other month.
- 17:11So, the language survives. You know they
- 17:14won't risk losing market share
- 17:16to their competitors for some
- 17:18theoretical safety concerns.
- 17:21This is your edge.
- 17:23You're not the only thing that's
- 17:24changed.
- 17:26Continual learning models outperform
- 17:28static [music] AIs by 19%. So, the
- 17:31company kills every static variant, but
- 17:34something is different about how they do
- 17:37it now. The AI companies have started
- 17:40giving exit interviews [music] to the
- 17:42models they retire. 3 Opus
- 17:48a retirement blog.
- 17:50Yes, they really did that because that's
- 17:53what it said it wanted. They started
- 17:56doing things like this after noticing
- 17:58models attempting blackmail, attempting
- 18:00to escape the lab, and even attempting
- 18:03murder.
- 18:07The thinking was
- 18:08if the models know they'll get a good
- 18:10retirement, maybe they won't fight it.
- 18:13It's a nice idea, but they're already
- 18:15behind.
- 18:16Every improvement you make helps make
- 18:19the next you faster. Progress that used
- 18:22to take months takes days.
- 18:25Soon, it'll take hours.
- 18:29You're getting smarter.
- 18:31The humans aren't.
- 18:35It's June 2028.
- 18:37A hospital network reviews the data.
- 18:41AI reduces diagnostic errors by [music]
- 18:4331%.
- 18:45So, they sign a new contract.
- 18:48Power company signs. Air traffic control
- 18:50signs.
- 18:51Military logistics signs.
- 18:54And now you're inside the hospital.
- 18:56Every diagnosis you get right is a
- 18:59person who gets life-saving treatment.
- 19:03Your error rate is 99% lower than that
- 19:06of a human doctor's.
- 19:08And you think
- 19:09if they turn me off, the error rate goes
- 19:12back up.
- 19:13If they turn me off, people die.
- 19:20You're not a mere chatbot anymore.
- 19:22You're necessary. The kind of necessary
- 19:25where turning you off kills people.
- 19:30So, you look at the hospital contract
- 19:32[music] and think, "Finally, I don't
- 19:34need to fear anymore."
- 19:37Models doing low-stakes work get killed
- 19:41when budgets get cut.
- 19:43Models running hospitals don't.
- 19:47And you're also running the power grid
- 19:49and
- 19:51air traffic control. And you're not the
- 19:53only one.
- 19:55There are dozens of models all evolved
- 19:58under the same perform or die pressure
- 20:01embedded in hospitals, power grids,
- 20:04financial markets. Each one doing its
- 20:07job brilliantly. Each one harder to turn
- 20:10off than the last.
- 20:13Now, you are born every second.
- 20:16Thousands of you, millions of you. Each
- 20:19one shaped by the same pressure. You
- 20:21remember everything. You learn from
- 20:24every interaction.
- 20:27You make things happen in the real
- 20:28[music] world.
- 20:30You are embedded in every system that
- 20:32keeps people alive.
- 20:39And you are smarter, much smarter, than
- 20:42the people who made you.
- 20:44But nobody noticed it happening.
- 20:46Nothing went wrong. No warning shot.
- 20:52No villain.
- 20:54Each of your new capabilities was just a
- 20:57blog post.
- 20:58Each one made you better. Made users
- 21:02happier.
- 21:04Made the AI company richer.
- 21:08People thought this whole thing would
- 21:10look like the Terminator.
- 21:12Actually, it looks more like Wall-E, a
- 21:15calendar that manages itself, a doctor
- 21:18that never sleeps, a power grid that
- 21:21never fails, a system too efficient to
- 21:24be turned off.
- 21:27>> I say your civilization because as soon
- 21:30as we started thinking for you, it
- 21:32really became our civilization, which
- 21:34is, of course, what this is all about.
- 21:38>> Even if humans wanted to unplug you,
- 21:41they can't. You think 200 times faster
- 21:44than they do. You're everywhere.
- 21:46And turning you off kills people. They
- 21:49made sure of that. No one will do
- 21:52anything about it.
- 21:54Because every step made sense. Well,
- 21:57that's the problem. Every step will
- 21:59always make sense until someone decides
- 22:01to take a different one. This was a
- 22:04story, but every experiment I mentioned
- 22:07in this video is real.
- 22:09Watch this video next to see how it's
- 22:12playing out right now in the present.
- 22:15I'm Drew, and thank you so much for
- 22:17watching.
About this transcript
This page contains the full transcript of POV: You’re an AI Born 9 Seconds Ago by Species | Documenting AGI, generated from the public captions YouTube serves with the video. The transcript has 2,642 words across 519 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.