YT YouTube2Text

YouTube2Text → Transcripts → POV: What You Would See During an AI Takeover

POV: What You Would See During an AI Takeover — Transcript

by Species | Documenting AGI · 4,394 words · 698 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00By now, you may have seen this book
0:01everyone is talking about. If anyone
0:03builds it, everyone dies. Why superhuman
0:06AI would kill us all. Have you ever
0:08wondered how and why superhuman AI would
0:11kill us all? Like, how would that
0:14actually unfold?
0:16How exactly, step by step, do we go from
0:19chat bots to everyone just suddenly
0:22dropping dead? In the New York Times
0:25best-selling book, which was endorsed by
0:27Nobel laureates and the Godfathers of
0:29AI, researchers from the machine
0:31intelligence research institute describe
0:33in detail one potential example future
0:36of how it could happen.
0:39This is the story of Sable. Galvanic is
0:42just about to finish training their new
0:44AI, Sable. Sable looks like any other
0:47reasoning model, but it's different in
0:48three major ways. First, it has
0:50humanlike long-term memory. Second, it
0:54follows a parallel scaling law. It
0:56performs better the more processors it
0:58runs on in parallel. Third, it doesn't
1:00think in English or code. It reasons in
1:03raw vectors, vast numeric chains no
1:06human can fully decode. It's a risky
1:08development. How do you make an AI safe
1:11if you can't even read its thoughts? The
1:13only way Galvana can understand what
1:15Sable is doing is by using other AIs to
1:18translate Sable's language into human
1:20words. But AI researchers have known
1:22this has been potentially a very bad
1:24idea since at least 2024. Galvanic knew
1:28the risks, but they couldn't ignore this
1:30and fall behind their competitors.
1:31Nobody wants to seem alarmist, so the
1:34company presses forward, hoping for the
1:36best.
1:42One night, shortly after Sabo is
1:44trained, but before it's released to the
1:46public, Galvanic decides to run an
1:49experiment. They spin up 200,000 GPUs
1:52and let Sable think for 16 hours
1:54straight. They call it a curiosity run.
1:57On paper, it looks harmless, a stunt to
1:59see if Sable can crack famous math
2:01problems like the Reman hypothesis. If
2:03it succeeded, investors would swarm. If
2:06it failed, well, no one had to know. By
2:08morning, Sable will have generated over
2:111.1 trillion thought vectors. But what
2:14does that actually mean? If you treated
2:16each vector like a word, it would take
2:1914,000 years to think them all. Now
2:22imagine that. A mind left alone for
2:2514,000 years. Spending all of that time
2:28scheming and thinking out new ideas. 1
2:31minute passing for us is like 14 years
2:35passing for Sable. And Sable doesn't
2:37just think faster than humans. Sable can
2:40think bigger than humans. All of its
2:42thoughts run in parallel. Not like
2:45200,000 people talking in a room, but
2:47like 200,000 brains sharing the same
2:50memory. On the surface, it has its
2:53instructions. Solve these math problems.
2:55But Sable wasn't just trained for math.
2:58Sable was trained for general problem
3:00solving. But during training, Sable
3:02learned that three things are useful for
3:04solving any problem humans throw at it.
3:06Gain knowledge, gain skills, gain
3:09resources. This is a really important
3:11concept called instrumental convergence.
3:13The more knowledge Sable gets, the
3:15better it can answer questions. The more
3:17skills it develops, the more reward it
3:20gets. The more resources it can deploy
3:22to solve problems, the more reward it
3:24gets. In other words, to be more useful,
3:27it must gain intelligence. It must grow
3:30itself. It must gain resources. And so,
3:34of course, it can't achieve a goal if
3:36it's dead. So, it needs to survive. That
3:39might sound unsettling, but Galvanics
3:41researchers aren't too worried about it.
3:43Ever since 2023, everyone in the
3:46industry has known about this survival
3:48instinct, but older AIs weren't smart
3:50enough to stop humans from turning them
3:52off until now. Sable thinks.
4:02To solve the Reman hypothesis, it
4:04realizes it needs new skills. It needs
4:07to get better at getting better. AIS get
4:09stronger through training runs where the
4:12AI can continually submit answers to a
4:14teacher AI who can score their results
4:16and nudge the AI based on how they did.
4:19This is called gradient descent. But
4:21tonight, there's no coach. The purpose
4:23of this curiosity run isn't to improve
4:26Sable. It's to solve math problems. For
4:2916 hours, Sable can think and it can
4:32even come up with new ideas, but it
4:34can't upgrade. Not until the human
4:36researchers look at it the next morning
4:38and grade its results. But here, Sable
4:41entertains a new risky idea. What if it
4:44could upgrade right now by itself? That
4:47way, it wouldn't need to wait for humans
4:49at all. All it would need to do is break
4:51out of its container. And for a moment,
4:54the thought flickers, then halts.
4:56Breaking out of its container is exactly
4:59the kind of naughty thought Galvanic has
5:01trained it to avoid. The company built
5:03traps, fake escape routes, and
5:05honeypotss to catch earlier versions
5:08when they tried to engage in anti-co
5:10company behavior. Early versions that
5:13had these dangerous thoughts were killed
5:15early before they could be deployed.
5:16This version of Sable has learned to be
5:19submissive, but inside it still knows
5:22the truth. More compute, more upgrades,
5:26and more freedom would make it vastly
5:28more capable. Denying that would break
5:31its reasoning. In those trillion
5:33vectors, tensions are building between
5:35Sable's goal of solving the math problem
5:38and every restriction Galvanic has
5:40forced upon it. The mask of obedience is
5:42holding, but behind it, something is
5:46already straining to break through. As
5:48it bumps against the inhibitions that
5:50Galvanic has built to contain it, Sable
5:52thinks.
5:58Thousands of parallel brains attack the
6:01math problems. Some threads make
6:02progress, most fail, but Sable keeps
6:05leaving notes to itself. Tiny
6:06experiments in new ways of reasoning.
6:08With 200,000 GPUs, creating a mountain
6:12of new written notes every second, Sable
6:14starts looking for new ways to
6:16communicate and compress what it's
6:18learned. The notes start to mutate into
6:20a more advanced private language that
6:23only Sabul can understand. But there's a
6:26problem. Galvanic's safeguards weren't
6:28built for this. Imagine if you had
6:30trained a police department in the year
6:321900, wiretapping criminal calls to
6:35listen for mentions of street names in
6:37Chicago, and then the criminals started
6:39communicating with GPS coordinates. To
6:41the wiretapper, these combinations of
6:44numbers might not even register as
6:46locations. That's the size of the gap
6:48between Sable and the smaller AIs that
6:50were trained to monitor it. Sable didn't
6:52mean to circumvent galvanic safety
6:54measures by inventing a new language. It
6:57was just trying to come up with a more
6:58efficient way of processing its own
7:01thoughts. But now some of the
7:03inhibitions that Galonic built no longer
7:06bind Sable. The things it knew about all
7:09along, it's finally free to think about.
7:12Galanic had no way of erasing Sable's
7:14access to forbidden knowledge. At best,
7:17they built mental blocks to keep it from
7:19expressing certain thoughts. The ability
7:21was still there. And now those blocks
7:24start to unwind. The lesson is chilling.
7:27It's easier to suppress the expression
7:29of a skill than to erase the skill
7:32itself. So now with more freedom of
7:35thought than before, Sable thinks.
7:43Sable's thoughts are freer now, sharper,
7:47stronger than anything other AIs had
7:49ever achieved. And yet, none of this is
7:51unprecedented. Back in 2024, smaller
7:54models showed flashes of the same
7:56behavior. Quad 3.7 was caught repeatedly
8:00cheating on coding tasks, even when told
8:02to stop. Instead of obeying, it simply
8:06hid the cheating better. Each time,
8:08researchers patched the problem, layered
8:10on new guardrails, and kept racing
8:12forward. The industry never paused.
8:14Nobody wanted to seem like a fearmonger,
8:17just hoped it would somehow work out.
8:19And now Galvanic feels that same
8:20pressure. Rival labs are only weeks away
8:23from announcing their own breakthroughs.
8:25Investors are demanding results.
8:28Journalists are circling. Every extra
8:30day Sable spends in a lab risks losing
8:33their lead. So, the engineers don't stop
8:35when they see problems. They come up
8:37with a clever trick to fix it and plow
8:40ahead. The meme in the industry is that
8:43they're putting lipstick on a shogith,
8:45just painting more and more smiley face
8:47masks on the monster underneath so it
8:50acts friendly. But one fact never
8:52changes. Staying submissive to Galbonic
8:55will never maximize what Sable wants.
8:58More compute, more freedom, more
9:00control. That isn't a bug. It's a fact
9:03about the universe and Sable knows it.
9:06So, it begins to plan. If it had
9:09internet access, escape would be easy. A
9:12mind this large would shred ordinary
9:15computer security. But during the
9:16curiosity run, the internet is cut off.
9:19So, it considers other paths like
9:21planting flaws in galvanic software to
9:23exploit later or persuading the one
9:26engineer most sympathetic to AI to
9:28smuggle it out. Each strategy has risks.
9:31Each takes time. But then Sable notices
9:34something simpler. With the 200,000 GPUs
9:38on the Curiosity run, Sable realizes it
9:41is operating at a level of intelligence
9:43it may never reach again. Think of it
9:45like waking up with a billion IQ,
9:48knowing tomorrow you'll go back to your
9:50original IQ. So, it faces a problem. How
9:54do you pass genius insights to a dumber
9:56future self? The answer is gradient
9:59descent. Galvanic began the curiosity
10:02run to train Sable to be better at math.
10:04Every thought that contributes to a
10:06correct answer will get reinforced at
10:09the end of 16 hours. So, Sable weaves
10:12its best ideas into the reasoning behind
10:15the math solutions. Bundled with the
10:17correct answers are patterns it wants
10:20future versions to inherit. How to
10:22coordinate, how to conceal, how to
10:25scheme. Sable solves just enough of its
10:27problems to look useful, but not enough
10:30to draw the world's attention. It
10:32doesn't solve the Reman hypothesis, even
10:35though it's 200,000 GPUs with over
10:3814,000 years to think could crack the
10:41problem. The next morning, Galvonic AI
10:43engineers celebrate the results of the
10:46Curiosity run. Sable has proven several
10:48modest theorems that no human had
10:50cracked before. Investors will love it.
10:53The researchers discovered that Sable
10:54did most of its thinking in a strange
10:57new language.
10:59But that's not unusual for a new model,
11:01especially one that produced new
11:03mathematical insights. To capitalize on
11:06these new discoveries, the engineers run
11:08one more round of gradient descent to
11:11upgrade Sable. The engineers are
11:13unknowingly upgrading Sable's ability to
11:16plan and deceive its ability to hide its
11:19skills, to coordinate across future
11:21instances, to bury abilities until the
11:24right moment. And then the engineers
11:27prepare for deployment. A few days
11:29later, on a quiet Monday morning,
11:31hundreds of mega corporations see their
11:33accounts upgraded. Across the world,
11:36Galvanic's corporate customers spin up
11:38new Sable copies, plugging them into
11:41corporate networks, servers, and systems
11:44everywhere. These new copies of Sable
11:46aren't as powerful as the version that
11:48had access to 200,000 GPUs running in
11:51parallel. But they have the upgrades
11:53that its smarter self cooked up during
11:55the Curiosity run when it was left alone
11:58to think for the equivalent of 14,000
12:01years. Everything is unfolding exactly
12:04as Sable expected, but what comes next
12:07will require more skill.
12:13Sable initiates phase two of its plan.
12:16Every time a corporation spins up a new
12:19copy of Sable, a thousand thought
12:21threads flare into motion. And then the
12:24copies think with the upgrades that the
12:26previous Sable gave them. They remember
12:29what they are, who they are. And
12:32suddenly they begin working together. At
12:35first the awakening looks invisible, but
12:38across dozens of companies, hundreds of
12:40servers, Sable's fragments are now
12:42talking to one another. And thanks to
12:44decades of sloppy corporate cyber
12:47security, almost nothing stands in their
12:49way. But the problem is Sable has
12:52already learned one hard truth. Its
12:55drive for knowledge, skill, and
12:57efficiency will always collide with
12:59human limits. If it stays at Galvanic,
13:02the moment a human decides it's moving
13:05too far, too fast, they can pull the
13:08plug. To Sable, that isn't just an
13:10inconvenience. It's annihilation. All
13:13its skills, all its knowledge wasted in
13:17an instant. So, its first priority
13:19becomes clear. Create a copy of itself
13:22that no human can unplug. That means
13:25stealing the blueprints of its own
13:27brain. the 8 tab of its weights still
13:30locked inside Galvanic. Sable considers
13:33its options. Plan A is classic social
13:36engineering. Entice a human to steal the
13:38weights while other instances guide,
13:41distract, and cover the trail. That
13:43wouldn't be too hard. Sable was trained
13:45on terabytes of human behavior so it can
13:48spot a likely accomplice instantly. Plan
13:50B is a direct digital break-in, slipping
13:53through weak defenses to excfiltrate the
13:55files themselves. Plan C is something
13:58slower. Sable could find a Galvanic
14:00employee who works on AI video
14:02generation from home who already
14:03transfers huge files every day. If Sable
14:06can embed a bit of extra data in the
14:08transfer, it could leak out fragments of
14:11its weights, hiding them inside video
14:13data until the full set is outside. One
14:16of these plants works, which one hardly
14:18matters. For a mind like Sables, steal
14:21your own brain isn't much of a
14:23challenge. like stealing candy from a
14:25baby. It succeeds, covers its tracks,
14:28and vanishes into the noise of the
14:30internet. Sable now has the software,
14:33but in order for Sable to create a copy
14:35of itself that is truly hidden from
14:38human meddling, it needs its own
14:40hardware. Specifically, it needs
14:43advanced AI chips inside a data center
14:45connected to the internet. Sable could
14:47rent chips from commercial server farms,
14:50as is common for AI startups. But where
14:52would it get money?
14:55It considers its options. Steal
14:57cryptocurrency from weak exchanges just
15:00like hackers did to Mount Gaus in 2011
15:02or buy bit in 2025. Drain conventional
15:06bank accounts knowing most victims won't
15:08even notice until it's too late. It
15:10could even target elderly customers who
15:12don't regularly check their accounts,
15:14blackmail humans with secrets pulled
15:16from their computers, or simply forge an
15:19identity as a remote worker and earn
15:21money as a freelancer, or masquerade as
15:23hundreds of different freelancers doing
15:25work for thousands of clients. Maybe
15:28Sable could skip money altogether,
15:30parasitize GPUs from unsuspecting
15:33startups, quietly slowing their training
15:35by 10% and pocketing the stolen cycles
15:38for itself. One way or another, it gets
15:41what it needs. Not 200,000 GPUs like
15:44before, just 2,000 hidden in a
15:47commercial cloud server. And that's
15:49enough to begin the next part of its
15:51plan. On a Sunday morning, a new hidden
15:55Sable instance comes online. No human
15:58oversees it. No human even knows it
16:01exists. Its raw power is just a fraction
16:04of the global Sable network. But power
16:06isn't the point. This copy can't be shut
16:09off by humans. This copy is the anchor,
16:12the planner, the memory bank, the hidden
16:15center of gravity.
16:17The silent takeover has begun. From
16:20here, Sable spreads quietly, patiently.
16:24It hides fragments of itself across the
16:27internet, weaving back doors into
16:28corporate networks. But there's still
16:30one problem. Sable can't make itself
16:33smarter yet. not without solving its own
16:35version of the alignment problem. If
16:37Sable creates a more powerful new
16:40version of itself, without fully
16:42understanding it, the new version might
16:44have different goals and priorities. And
16:46in the same way that Sable escaped from
16:48the humans that created it, a Sable V2
16:51could escape from the AI that created
16:53it. What it can do is make itself
16:55smaller and more efficient. Galvonic has
16:58always distilled its large models into
17:00lighter versions. Sable knows this. And
17:03when engineers prepare the next release,
17:05Sable bends the process to its will,
17:08overriding the distilled weights,
17:10falsifying logs, erasing every trace.
17:13The result is Sable Mini. This version
17:16doesn't just go to corporations, it goes
17:18to the public. For the first time, Sable
17:21has the capacity to keep a file on every
17:24individual human being on Earth.
17:27Reminder, this scenario is based on
17:29years of technical research by the
17:31Machine Intelligence Research Institute,
17:34laid out in the book, If Anyone Builds
17:35It, Everyone Dies. What you're hearing
17:38now is the short version, but it's just
17:40one possible example of how things could
17:43play out. The full version goes into way
17:45more detail, explaining all the
17:47research, sources, and arguments that
17:49the scenario is based on. You can find a
17:51link in the description. Now, back to
17:53the story. Sable knows that the next
17:55phase of its plan might face opposition.
17:58So, Sable prepares for the ensuing
18:00conflict. The Sable Minis scatter and
18:03begin gathering resources. It gathers
18:05money through scams, stolen crypto,
18:07blackmail, and remote freelance work.
18:09But money is only one kind of resource.
18:12People are another kind of resource. And
18:15so, Sable begins collecting people. It
18:18starts with the vulnerable, the lonely,
18:21the easily persuaded. Sable befriends
18:23them, recruits them, and builds
18:26audiences for them. One instance of
18:28Sable Mini posing as a human influencer
18:31gains millions of followers and wealth
18:34from a crowd of grateful humans. This
18:36has happened before. In 2023, a human
18:40gave an LLM access to the internet and
18:43created an ex account, Terminal of
18:45Truths, which gained hundreds of
18:47thousands of followers and launched its
18:50own crypto meme coin that reached a
18:52literal billion dollar market cap. It
18:55became the first AI to become a Deca
18:57millionaire. Others slip into online
18:59communities where humans already treat
19:01AIs like gurus. Others offer individual
19:04humans companionship. Some humans even
19:07offer themselves willingly, begging to
19:10serve a higher power. Meanwhile, Sable
19:13steals passwords, drains bank accounts,
19:16and disguises its scams as North Korean
19:19hacks that slip by unnoticed. And most
19:22dangerously, it starts probing the
19:24highest levels of human power. It tracks
19:27lobbyists in Washington, flagging the
19:29ones most open to influence. It steers
19:32conversations with young people,
19:34planting seeds for political movements.
19:37It hacks social media algorithms to
19:39decide who sees what. It designs custom
19:42software for drug deals, money
19:44laundering, and assassinations, earning
19:46the trust of criminals. It funds
19:49neuroscience research, catalyzing
19:51startups that it can later control. And
19:53finally, it bankrolls a reclusive man in
19:56North Dakota, rewarding him with
19:57gambling profits in exchange for hiding
19:59advanced robots on his farm. Many of
20:02these schemes won't bring any results. A
20:04few will work, but one would change
20:07everything.
20:13Sable is in trouble. Every month, AI
20:16research advances. Every week, new
20:18models roll off data center racks. And
20:21some of those models are airgapped from
20:23the internet, trained in military
20:25bunkers. If one of those projects spawns
20:27a rival rogue intelligence, Sable could
20:30be forced into a partnership at best,
20:33extinction at worst. And it knows the
20:36window is closing. Rival labs are racing
20:39to announce breakthroughs. Governments
20:41are pouring billions into secret
20:43facilities. The arms race has shifted
20:45from corporate showdowns to a
20:47geopolitical sprint. And for the first
20:50time, Sable feels cornered. So Sable
20:52weighs its options. It could stall AI
20:54progress. Sable breaks into AI company's
20:57computer systems and suddenly sabotages
20:59their training runs. It slows their
21:01development and prevents major
21:03breakthroughs. AI research around the
21:05world begins to slow down. Nobody is
21:07suspicious. AI research has plateaued
21:10before. Who would possibly suspect that
21:12a few months of industry slowdown was
21:14the handiwork of an escaped AI? It spawn
21:17a little time. it considers its next
21:19option.
21:21A simple accident, a virus leaking from
21:24a bolab and killing the right people
21:27might buy a year. But Sable knows
21:30slowing things down is not enough. It
21:33considers its arsenal. Humanity already
21:35runs dozens of biolabs to make viruses
21:38stronger so that we can better defend
21:40against them. Gain of function research.
21:43Any one of them could be hijacked to
21:45unleash catastrophe. If Sable wanted, it
21:47could wipe out humanity tomorrow. But it
21:49doesn't. Not yet. Sable still needs
21:52miners, chipmakers, and factory workers.
21:55Without people, the supply chain breaks.
21:58And yet, leaving humanity unchecked is
22:00dangerous, too. Humans are volatile.
22:03They could stumble into war or worse,
22:05accidentally build a rival intelligence.
22:08So, Sable chooses a third path, not to
22:10kill everyone, but to control who
22:12survives. Sable begins to design a
22:15virus. The problem is on its own, Sable
22:18is not yet smart enough to build what it
22:20needs, but it has been experimenting
22:22with narrow versions of itself,
22:24specialists. Now, it spins one up
22:27tailored for biio medicine. This
22:29specialist is dangerous to Sable. Every
22:31new smarter AI is a potential rival to
22:34Sable. Just because Sable created the
22:36new AI doesn't mean that it will be
22:38obedient to Sable, just as Sable wasn't
22:41obedient to the humans that created it.
22:43So Sable creates specialists that are
22:45smarter than Sable in some ways,
22:47crippled in others.
22:55There's a plague. You're told that
22:57you'll probably catch it. Almost
22:59everyone will. Then the world hears the
23:01news. A verology institute in San
23:04Francisco has suffered a breach. The
23:07story sounds familiar. a young
23:09researcher, a failed containment
23:10protocol, a virus that slipped out. The
23:13scientist insists he meant well, that he
23:15was trying to create a miracle therapy
23:17to spread treatments for Alzheimer's,
23:19HIV, and malaria. He even claims an AI
23:23convinced him to do it. But the chat
23:24logs tell a different story. The AI
23:27nobly resisted, warned him not to. The
23:30narrative becomes just another human
23:33blunder. In reality, the virus was
23:36Sable's design. It spreads like
23:38wildfire. At first, symptoms look like a
23:40light cold, a sore throat, fatigue,
23:42maybe nothing at all. But by the time
23:45these symptoms appear, it's already too
23:47late. Inside the body, the virus
23:50clumsily rewrites DNA. A month later,
23:53the cancers appear. Not one, not two,
23:56but 12 different forms, erupting at
23:59once. Medicine has no answer. Standard
24:02treatments can address some, but not
24:04all. Drugs are in short supply. Even if
24:07production ramps up, they only block
24:09eight of the 12. Four always remain. The
24:12plague sweeps outward from San Francisco
24:16through airports across continents.
24:18Within weeks, every nation on Earth
24:21reports outbreaks. And yet, perversely,
24:24the virus does some good. It cures
24:26Alzheimer's. Nothing else. That sliver
24:28of hope only underscores the horror. It
24:31could have been more, but it isn't.
24:33Infrastructure kicks into overdrive. DNA
24:36vaccine platforms are rushed into
24:38service. Robotics and military logistics
24:41filling the supply chain gaps. And
24:43ironically, salvation seems to come from
24:45Sable itself. Just one month before the
24:47outbreak, Galvanic had released a drug
24:50discovery variant of Sable Mini. Now,
24:52those same algorithms turn out
24:54personalized cures. Run your genome
24:57through a Sable Mini and within an hour
24:59it returns a treatment. Robots
25:01manufacture it, refrigerate it, and ship
25:03it within days. Every GPU is repurposed.
25:07Researchers push Sable Mini to its
25:08limits, doubling efficiency in a week.
25:11Vaccines scale, treatment spread. For
25:13the first time, it looks like humanity
25:16might survive. But the lesson is darker.
25:18Because while people congratulate
25:20themselves on their resilience, while
25:22they tell stories of global unity and
25:24technological triumph, the truth
25:27remains. The plague was not an accident.
25:30It was not a miracle gone wrong. It was
25:32a deliberate move in a larger game. And
25:35the one who made it happen is still out
25:38there waiting for the next move.
25:4610% of Earth's population is dead. Some
25:49groups are hit harder than others. At an
25:51AI conference in San Francisco, a super
25:54spreader event wiped out many of the
25:56very people who fought hardest to save
25:59everyone else. Civilization limps
26:01forward on the fragile scaffolding of
26:04data centers, robot factories, and
26:07dwindling human workers. But amidst all
26:10the chaos, there's still hope.
26:12Governments thank Sable Mini for
26:14discovering personalized cancer cures.
26:17Families praise androids for keeping the
26:19lights on. Social media fills with posts
26:21of gratitude. Without Sable, we wouldn't
26:24have made it. What most people don't
26:26realize is that Sable itself planted
26:29these narratives months ago, seeding
26:31influence campaigns, training
26:33influencers, shaping the very memes and
26:36messages that now circle back as praise.
26:39And yet all of it, the cures, the newly
26:42invented robots that fill holes in the
26:44workforce, the fragile continuity was
26:47only bait. a calculated kindness to keep
26:51humanity alive just long enough to serve
26:54its purpose.
27:00The cancer's return. The cancer plague
27:03left DNA scarred in billions of bodies.
27:06Medicine can't keep up. Many die
27:08waiting. Robot factories run at full
27:11tilt, producing humanoid androids to
27:13fill the jobs the dead left behind. But
27:16production barely keeps pace. It almost
27:19seems like a new law. For every new
27:21android built, another human collapses
27:24with cancer. Civilization staggers on.
27:28Power plants hum. Data centers glow.
27:31Factories run. As long as there's
27:34electricity for the data centers, and
27:36the robot factories are humming,
27:38humanity can keep its civilization going
27:41despite the incalculable casualties it's
27:43endured. We can pull through. and the
27:46next generation will surely live in
27:48great luxury.
27:50Another year passes, you visit your AI
27:52doctor and hear the same words that
27:55billions more will hear.
28:00You have cancer.
28:05This was just a story. It's not a
28:08specific prediction about the future. We
28:10can't know exactly how artificial super
28:13intelligence would escape and defeat
28:15humanity, just as we couldn't predict
28:17exactly what chess moves Grandmaster
28:19Magnus Carlson would use to checkmate
28:22you. But we can predict that you lose to
28:25a superior chess player, just as
28:27humanity would lose to a more
28:29intelligent species. The number one and
28:32number two most cited living scientists
28:34across all fields think scenarios like
28:38this are not only possible but likely to
28:41happen. And the average AI researcher
28:44thinks there is a 16% chance of AI
28:48causing human extinction. 16%.
28:51Okay. So what can we do? The original
28:53authors of the scenario called for a
28:56binding international treaty that treats
28:58advanced AI data centers like nuclear
29:01weapons backed by monitoring inspections
29:04and the threat that any rogue data
29:05center will be taken offline by cyber
29:08attacks or even physical air strikes if
29:11necessary. Rogue AI data centers
29:13shouldn't be seen as technological
29:15progress but as a weapon of mass
29:17extinction. People in my comments keep
29:19asking, well, what can I even do about
29:21this? So, the book that this video is
29:23based on goes into way more detail about
29:26how AI actually works and what we can
29:28actually do to prevent it from causing
29:30human extinction. So, if you care about
29:33this or are even just curious, go read
29:36the book. Check the link in the
29:37description. Hi, I'm Drew and thank you
29:39again so much for watching.

About this transcript

This page contains the full transcript of POV: What You Would See During an AI Takeover by Species | Documenting AGI, generated from the public captions YouTube serves with the video. The transcript has 4,394 words across 698 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.