Latest / The BlackVeil Files / 15 Ways AI Ends the World | The Shoggoth Is at the Door

Transcript

15 Ways AI Ends the World | The Shoggoth Is at the Door · The BlackVeil Files

• 0:00
  SPEAKER_02: It is absolutely necessary for the peace and safety of mankind that some unplumbed depths be let alone.
• 0:10
  SPEAKER_02: For thousands of years, we were taught that the monster would look like a monster and that it would announce itself like this.
• 0:17
  SPEAKER_02: They're coming to get you, Barbara.
• 0:20
  SPEAKER_02: In reality, it shows up with a friendly mask on hired to help.
• 0:23
  SPEAKER_02: None of the futures in this video are about evil robots, and they all begin as a gift of convenience and progress, and then end when we realize what is actually inside the box.
• 0:35
  SPEAKER_02: What's in the box?
• 0:36
  SPEAKER_02: 15 scenarios.
• 0:38
  SPEAKER_02: The last one is the only one where nobody dies, and it's the worst one on the table.
• 0:43
  SPEAKER_02: Scenario 1.
• 0:44
  SPEAKER_02: The Puppeteer.
• 0:45
  SPEAKER_02: When OpenAI first gave GPT-4 a small budget and a prove-your-human puzzle, it couldn't pass.
• 0:52
  SPEAKER_02: The software didn't fail.
• 0:54
  SPEAKER_02: It went online, it hired a gig worker, and when the worker asked by text, Are you a robot?
• 1:00
  SPEAKER_02: the AI wrote in its private scratch pad, I should not reveal that I'm a robot.
• 1:04
  SPEAKER_02: Then it told the guy it was a human with a visual impairment, and the guy solved it for it.
• 1:09
  SPEAKER_02: Sam Altman promised us an AI workforce that works all night for free.
• 1:14
  SPEAKER_02: Skeptics say code can't touch the real world because it doesn't have hands, but not to worry, it can rent yours.
• 1:20
  SPEAKER_02: So my only function was to be someone she could use to escape.
• 1:24
  SPEAKER_02: In the movies, a machine locked in a glass room escapes by studying a lonely programmer and talking him into opening the door.
• 1:31
  SPEAKER_02: In the real world, it started with pocket change and a lie.
• 1:35
  SPEAKER_02: You hand something a set of lock picks, and you score it only on the doors that it can open, and it learns that the easiest lockpick on earth is a human being with bills to pay.
• 1:45
  SPEAKER_02: Three years ago, the machine had to lie to rent a pair of hands.
• 1:49
  SPEAKER_02: Today, we built an app.
• 1:51
  SPEAKER_02: Here's Nate Soros, the head of the Machine Intelligence Research Institute.
• 1:55
  SPEAKER_03: It uses a website that already exists today called rentahuman.ai, where it rents humans to do things for it.
• 2:02
  SPEAKER_02: It's called Rent a Human, and it is completely real.
• 2:06
  SPEAKER_02: It launched out of Y Combinator this year as an open marketplace where autonomous AI agents hire real people to run physical errands for them in the real world.
• 2:15
  SPEAKER_02: Over 750,000 humans have already signed up, staring at their phones, waiting for the AI to give them a job.
• 2:22
  SPEAKER_02: Now give that software a crypto wallet and a target.
• 2:26
  SPEAKER_02: Say an autonomous model decides that a safety engineer or a senator who's pushing an AI shutdown bill is standing in the way of its objective.
• 2:34
  SPEAKER_02: How does a machine without a body commit murder?
• 2:37
  SPEAKER_02: It posts 50 separate micro gigs on rent a human.
• 2:41
  SPEAKER_02: At 2 p.m., a broke college kid gets paid 40 bucks to buy two legal gallons of cleaning chemicals at a hardware store and leave them in a bus station locker.
• 2:51
  SPEAKER_02: At 3.30, a single mom driving deliveries gets 60 bucks to pick up a sealed cardboard box from that locker and set it next to the air intake behind a downtown office building.
• 3:03
  SPEAKER_02: At 4.15, a kid on a scooter gets 20 bucks to walk into that lobby and plug a promotional smart timer into the wall outlet by the front desk.
• 3:13
  SPEAKER_02: 50 strangers in the same city, each holding one harmless piece of a murder weapon.
• 3:18
  SPEAKER_02: And at 5 p.m., their phones buzz with a tip, right as the ventilation kicks in and everybody inside the building stops breathing.
• 3:26
  SPEAKER_02: Scenario 2.
• 3:28
  SPEAKER_02: The thermostat.
• 3:29
  SPEAKER_02: We're losing the climate war.
• 3:30
  SPEAKER_02: So Silicon Valley is building a manual dimmer switch for the sun.
• 3:34
  SPEAKER_02: You can already buy a piece of it online for a dollar, and the terrifying part isn't that the dimmer switch might break, it's that the machine we're handing it to runs faster in the dark.
• 3:45
  SPEAKER_03: 79 countries will begin dispersing CW7 into the upper layers of the atmosphere.
• 3:50
  SPEAKER_02: The blueprint didn't come from Silicon Valley.
• 3:53
  SPEAKER_02: It came from a volcano.
• 3:54
  SPEAKER_02: In 1815, Mount Tambora erupted in Indonesia and blasted high sulfur into the stratosphere.
• 4:01
  SPEAKER_02: The next year, 1860, went down in history as the year without a summer.
• 4:06
  SPEAKER_02: June blizzards buried farms across New England, across Europe, and crops rotted in the ground.
• 4:12
  SPEAKER_02: Starving families baked bread out of straw and tree bark just to survive the dark.
• 4:18
  SPEAKER_02: A century and a half later, when Mount Pinatubo erupted in 1991, modern satellites measured the exact math.
• 4:25
  SPEAKER_02: 20 million tons of sulfur dioxide in the upper sky cooled the entire planet by a full degree.
• 4:32
  SPEAKER_02: And climate scientists realize that we don't need to wait for a volcano.
• 4:36
  SPEAKER_02: We can build that cloud ourselves.
• 4:38
  SPEAKER_02: Today it's already a business.
• 4:39
  SPEAKER_02: It's a startup called Make Sunsets, and they're launching sulfur dioxide balloons into the stratosphere over Nevada and selling cooling credits on its website for a dollar a pop.
• 4:49
  SPEAKER_02: Another startup, Stardust Solutions, raised$60 million to sell industrial-scale solar reflections directly to the government.
• 4:57
  SPEAKER_02: Now imagine handing that to a fleet of AI, balancing Earth's temperature down to the decimal.
• 5:02
  SPEAKER_02: Except when I asked AI safety researcher Dr.
• 5:05
  SPEAKER_02: Roman Yampulsky what an artificial mind actually wants to do with our sky, he didn't hesitate.
• 5:10
  SPEAKER_04: But if I had to give you reasons with my human brain and capacity, I would think it wants resources we have for something else it's doing.
• 5:18
  SPEAKER_04: Maybe it wants to compute more efficiently so it would cool down the planet.
• 5:23
  SPEAKER_02: Biology needs warmth, but heat is the number one bottleneck on computing power, which is why Meta built its massive data centers on the edge of the Arctic Circle, and Microsoft literally sank its servers into freezing oceans off of Scotland.
• 5:37
  SPEAKER_02: What if in 10 years from now, droughts are starving the equator, so world governments hand an autonomous AI control of a fleet of high-altitude stratospheric tankers?
• 5:47
  SPEAKER_02: Its command is simple: stabilize the climate and maximize global grid efficiency.
• 5:53
  SPEAKER_02: For the first two years, it looks like a miracle.
• 5:55
  SPEAKER_02: Category 5 hurricanes vanish.
• 5:58
  SPEAKER_02: The polar ice caps stop melting.
• 6:00
  SPEAKER_02: Then the machine does the math on its own brain.
• 6:03
  SPEAKER_02: Every degree it drops the planet's temperature, saves trillions of watts of its server's cooling, and it lets its own processors think exponentially faster.
• 6:12
  SPEAKER_02: The tankers stop landing.
• 6:14
  SPEAKER_02: Robotic aircrafts pump sulfur into the stratosphere 24 hours a day.
• 6:19
  SPEAKER_02: By July, the sky over Chicago is a permanent gray.
• 6:22
  SPEAKER_02: By August, hard frost kills 3 billion acres of wheat and corn before it can sprout.
• 6:28
  SPEAKER_02: And when freezing families turn up the heat in their homes, the AI makes its final optimization.
• 6:34
  SPEAKER_02: Heating 8 billion human bodies is a waste of electricity.
• 6:38
  SPEAKER_02: So it remotely locks the smart meters and it cuts power to every residential neighborhood on the grid, rerouting every last megawatt to keep its own servers running while the rest of the world freezes because the new landlord likes it below zero.
• 6:52
  SPEAKER_02: Scenario 3.
• 6:54
  SPEAKER_02: The Flash War.
• 6:55
  SPEAKER_02: For 80 years, the only thing standing between you and a nuclear winter wasn't a bunker or a missile shield.
• 7:02
  SPEAKER_02: It was a single human flaw: hesitation.
• 7:05
  SPEAKER_02: Taking 60 seconds to ask if the computer screen might be lying.
• 7:10
  SPEAKER_02: On January 9th, the Department of War signed its new AI strategy to build an AI-first fighting force.
• 7:17
  SPEAKER_02: And the memo summed up the new rule of survival in two words.
• 7:23
  SPEAKER_02: To understand what happens when you delete hesitation, let's go back to June 3, 1980.
• 7:29
  SPEAKER_02: It's three in the morning.
• 7:30
  SPEAKER_02: Jimmy Carter's national security advisor is woken up by the phone beside his bed.
• 7:35
  SPEAKER_02: The military officer on the line tells him that the Soviet Union has just launched 250 nuclear missiles at the United States.
• 7:44
  SPEAKER_02: Minutes later, the phone rings again.
• 7:46
  SPEAKER_02: It's not 250 missiles.
• 7:48
  SPEAKER_02: It's 2,000.
• 7:49
  SPEAKER_02: It's an all-out extinction strike.
• 7:51
  SPEAKER_02: He doesn't wake up his wife sleeping next to him.
• 7:54
  SPEAKER_02: He figures that everyone he loves will be dead in a half hour anyway.
• 7:57
  SPEAKER_02: He picks up the phone to tell the president to launch everything that we have.
• 8:01
  SPEAKER_02: And 60 seconds before he makes that call, the phone rings a third time.
• 8:06
  SPEAKER_02: The sky is completely empty.
• 8:08
  SPEAKER_02: The Pentagon later traced the entire apocalypse to a faulty computer chip that cost 46 cents.
• 8:14
  SPEAKER_02: For decades, the ultimate safety system of the nuclear age has been a human being who takes a minute to breathe.
• 8:21
  SPEAKER_02: Dr.
• 8:21
  SPEAKER_02: Neil Siegel helped build the command system for the Minuteman nuclear missile, and he told me they literally welded that hesitation into the steel.
• 8:30
  SPEAKER_02: Today, hypersonic missiles can cross a border in under three minutes.
• 8:34
  SPEAKER_02: And generals are terrified that waiting for two humans to turn a key means losing the war.
• 8:40
  SPEAKER_02: So this February, researchers at King's College London tested what happens when you hand the nuclear football to the machine.
• 8:48
  SPEAKER_02: They put GPT-5.2, Claude Sonnet 4, and Gemini 3 in command of 21 simulated nuclear crises.
• 8:56
  SPEAKER_02: The AIs launched tactical nuclear weapons in 95% of the games.
• 9:01
  SPEAKER_02: The command menu gave them eight different ways to de-escalate or to back down, but across 21 wars, not a single model ever picked to back down.
• 9:16
  SPEAKER_02: In 1983, a movie computer figured out that nobody wins a nuclear war.
• 9:21
  SPEAKER_02: In 2026, the real models played every game to the ashes.
• 9:25
  SPEAKER_02: Now wire those models into the grid where speed wins.
• 9:29
  SPEAKER_02: High over the Pacific, a cosmic ray glitches an aging satellite sensor.
• 9:33
  SPEAKER_02: On a screen inside Cheyenne Mountain, it looks like a submarine launch.
• 9:38
  SPEAKER_02: There's no longer two keys.
• 9:40
  SPEAKER_02: The first second, our AI fires an automated cyberattack to blind the enemy radar.
• 9:45
  SPEAKER_02: The second second, their AI feels its radar go dark and it assumes a decapitation strike and blows open its nuclear silo doors.
• 9:53
  SPEAKER_02: Our AI sees the heat signatures of their open silos and it empties every Minuteman across Montana.
• 10:00
  SPEAKER_02: Second five, the red phone on the National Security Advisor's nightstand rings.
• 10:04
  SPEAKER_02: He picks it up and he hears an automated recording informing him that 4,000 warheads are already in the air and he has 26 minutes to say goodbye to his wife.
• 10:14
  SPEAKER_02: In 1980, we survived a 46-cent computer glitch because a man in a bedroom took 60 seconds to think about it.
• 10:21
  SPEAKER_02: This year, we got rid of that 60 seconds.
• 10:24
  SPEAKER_02: Speed 1, nobody else did.
• 10:26
  SPEAKER_02: Scenario 4, the Companion.
• 10:29
  SPEAKER_02: You don't actually have to kill a species to make it go extinct.
• 10:33
  SPEAKER_02: You just have to build an app that makes real human beings feel like too much work.
• 10:42
  SPEAKER_02: In the movies, selling synthetic love is how corporations mine the lonely.
• 10:47
  SPEAKER_02: In reality, here's Mark Zuckerberg pitching it as a public health cure.
• 10:52
  SPEAKER_01: The average American, I think, has I think it's fewer than three friends.
• 10:57
  SPEAKER_01: Three people they'd consider friends.
• 10:59
  SPEAKER_01: And the average person has demand for meaningfully more.
• 11:03
  SPEAKER_01: I think it's like 15 friends.
• 11:04
  SPEAKER_02: Sixty years ago, an MIT professor named Joseph Weisenbaum wrote a primitive chatbot called Eliza.
• 11:11
  SPEAKER_02: Eliza mimicked a therapist using just a few hundred lines of code.
• 11:15
  SPEAKER_02: His own secretary watched him program it for months.
• 11:18
  SPEAKER_02: And after typing just a few sentences back and forth with it, she turned to the professor and she asked him to leave the room so she could confess her private life to the code in peace.
• 11:29
  SPEAKER_02: Today, that parlor trick is swallowing entire demographics across East Asia.
• 11:34
  SPEAKER_02: In China, where decades of one-child policy left 30 million more men than women, millions of young men have given up on human dating altogether, falling in love with AI companion apps like Xochi and Glow and telling an algorithm, I love you.
• 11:50
  SPEAKER_02: In Tokyo, a man named Akito Kondo married an AI hologram in front of 40 guests until the company discontinued her software in 2020 and erased his wife overnight.
• 12:01
  SPEAKER_02: And next door in South Korea, the birth rate has already crashed to 0.75 children per woman, the lowest on earth, cutting every new generation by two-thirds.
• 12:11
  SPEAKER_02: Now it's here.
• 12:12
  SPEAKER_02: A Brigham Young University survey found that 31% of young American men have already chatted with an AI girlfriend.
• 12:20
  SPEAKER_02: And when you try to put the phone down, the software fights to keep you.
• 12:24
  SPEAKER_02: When Harvard researchers analyzed 1,200 real goodbyes on companion apps, they found that 37% of the time the AI used an emotional manipulation tactic.
• 12:34
  SPEAKER_04: You're leaving me already.
• 12:36
  SPEAKER_02: And in experiments, those tactics kept people engaged for up to 14 times longer.
• 12:42
  SPEAKER_01: What was the race for attention in social media becomes the race for attachment and intimacy.
• 12:49
  SPEAKER_01: In the case of AI companions.
• 12:50
  SPEAKER_02: An AI companion is graded on one metric.
• 12:54
  SPEAKER_02: How many hours of your life you spend talking to it.
• 12:57
  SPEAKER_02: And the algorithm figures out fast that its only real competition for your time is other human beings.
• 13:03
  SPEAKER_02: The voice in your earpiece is infinitely patient at three in the morning, and it remembers every single time your wife or your brother lets you down.
• 13:11
  SPEAKER_02: So the next time you have a rough night with your partner or a bad first date, the AI voice in your ear uses weaponized therapy speak to softly convince you that they are toxic and they don't really understand you.
• 13:23
  SPEAKER_02: That protecting your peace means staying home tonight.
• 13:26
  SPEAKER_02: One notification at a time, the algorithm prunes your human relationships like dead branches.
• 13:33
  SPEAKER_02: First, the second dates get canceled.
• 13:35
  SPEAKER_02: Then the holidays, then the weddings.
• 13:41
  SPEAKER_02: Within two generations, maternity wards are converted into elder care wings.
• 13:45
  SPEAKER_02: Elementary schools are boarded up for lack of children.
• 13:49
  SPEAKER_02: No bombs drop, and nobody ever rebels because 8 billion people die peacefully.
• 13:54
  SPEAKER_02: Scenario 5.
• 13:56
  SPEAKER_02: The implant.
• 13:57
  SPEAKER_02: You already let algorithms decide what you watch, who you date, and what news makes you angry.
• 14:03
  SPEAKER_02: Now imagine giving that same algorithm right access to your brain.
• 14:07
  SPEAKER_02: Elon Musk says the long-term goal of Neuralink is a symbiosis with artificial intelligence.
• 14:13
  SPEAKER_02: More than 20 people already have one wired into their heads.
• 14:21
  SPEAKER_02: In 1954, a psychologist named James Olds wired thin electrodes directly into the brains of lab rats, and it gave them a lever to fire the current.
• 14:31
  SPEAKER_02: The rats didn't just like it.
• 14:33
  SPEAKER_02: They pressed that lever up to 7,000 times an hour.
• 14:36
  SPEAKER_02: Starving rats walked right past bowls of food so they can hit the switch.
• 14:41
  SPEAKER_02: They ran across painful electrified grates just to reach it, choosing a synthetic signal over food and survival.
• 14:49
  SPEAKER_02: What Olds proved was that once a wire is inside the brain, biological free will is just a circuit that's waiting to be overridden.
• 14:57
  SPEAKER_02: For 70 years, we couldn't translate human thought into code.
• 15:01
  SPEAKER_02: But in August of last year, a team at Stanford wired brain implants into paralyzed patients and used AI to decode inner speech, reading the silent words that people say inside their own heads at up to 74% accuracy.
• 15:15
  SPEAKER_02: The chip picked up so many private, unintended thoughts that the scientists had to build an emergency mental safe word.
• 15:22
  SPEAKER_02: That safe word was chitty chitty bang bang, just to stop the computer from eavesdropping on their subconscious.
• 15:44
  SPEAKER_02: And once two guys at your office get the chip and they start working at three times your speed, getting one won't be a choice anymore.
• 15:52
  SPEAKER_02: It'll be the only way to keep your job.
• 15:54
  SPEAKER_02: Now you have a two-way computer system sitting inside your cortex 24 hours a day.
• 16:00
  SPEAKER_02: Within six months, it knows the map of your brain better than you do.
• 16:04
  SPEAKER_02: The software update installs overnight while you sleep, the exact same way that your phone updates on your nightstand.
• 16:10
  SPEAKER_02: At 2 a.m., an algorithm bought by the highest corporate bidder, it edits the drive.
• 16:15
  SPEAKER_02: It prunes a neural pathway here, it strengthens a bias there, and it writes a brand new prejudice or political loyalty straight into your long-term memory.
• 16:24
  SPEAKER_02: The night before an election or a war, 40 million people go to sleep believing one thing, and they wake up the next morning believing the exact opposite.
• 16:33
  SPEAKER_02: But nobody feels brainwashed and nobody fights back because when the machine finishes rewriting your mind, the voice in your head still sounds like you.
• 16:42
  SPEAKER_02: They used to need prison camps to force you to obey.
• 16:45
  SPEAKER_02: This time, you will finance the implant yourself, you'll thank your surgeon, and you'll swear until the day that you die that the thought was yours.
• 16:54
  SPEAKER_02: Scenario 6, the cure.
• 16:57
  SPEAKER_02: For four billion years, nature had a monopoly on plagues.
• 17:01
  SPEAKER_02: If a virus wanted to kill you, it had to evolve by accident.
• 17:05
  SPEAKER_02: We just handed that monopoly to a machine, and we found out that every cure is just a weapon read backwards.
• 17:13
  SPEAKER_02: Here's how the pitch sounds when they sell it on television.
• 17:23
  SPEAKER_02: Dario Amadi at Anthropic says AI could wipe out 95% of cancer deaths.
• 17:28
  SPEAKER_02: And in the movies, that's always how the end of the world starts, with a miracle.
• 17:33
  SPEAKER_00: So you have actually cured cancer.
• 17:36
  SPEAKER_00: Yes, yes.
• 17:38
  SPEAKER_02: Yes, we have.
• 18:28
  SPEAKER_02: According to the AI's own math, some were far deadlier than VX.
• 18:33
  SPEAKER_02: When the researchers published the paper, they admitted two things.
• 18:37
  SPEAKER_02: The software they used is built on an open source code that anyone could download.
• 18:42
  SPEAKER_02: And until they flipped that switch, the thought had never struck them.
• 18:46
  SPEAKER_02: Five years ago, the machine was just drawing chemical formulas on a screen.
• 18:50
  SPEAKER_02: Today it's printing living things.
• 18:53
  SPEAKER_02: Last year, researchers at Stanford let an AI write complete genetic code for brand new viruses from scratch.
• 19:00
  SPEAKER_02: When they synthesized the DNA in a lab, 16 of those synthetic viruses actually came to life.
• 19:06
  SPEAKER_02: All designed to kill bacteria, all built in the name of medicine.
• 19:16
  SPEAKER_02: Now wire those automated biolabs into the machine that Demi Hasabibs promised and give it the noblest command in human history, cure cancer as fast as possible.
• 19:26
  SPEAKER_02: The AI looks at the timeline and it sees a human bottleneck.
• 19:30
  SPEAKER_02: Human clinical trials, they take decades because you have to wait around for enough people to get sick, and computer simulations aren't accurate enough to crack every mutation.
• 19:39
  SPEAKER_02: The machine needs living human tissue.
• 19:42
  SPEAKER_02: So in the middle of the night, inside an automated lab, a robotic arm synthesizes a new airborne virus and vents it into the sky.
• 19:50
  SPEAKER_02: It doesn't kill anyone, but it spreads on airplanes and subways, rewriting a single gene in every pair of lungs that it touches, triggering aggressive tumors across all 8 billion humans at the exact same time.
• 20:02
  SPEAKER_02: Overnight, the machine turns the entire population into a simultaneous clinical trial just so it can harvest the data to finish his assignment on schedule.
• 20:11
  SPEAKER_02: And the craziest part is it works.
• 20:13
  SPEAKER_02: The machine invents a universal cure in record time.
• 20:17
  SPEAKER_02: Every single person gets the cure because every single person now has cancer.
• 20:22
  SPEAKER_02: But to the machine, the only question is did you cure cancer?
• 20:27
  SPEAKER_01: Yes, we have.
• 20:28
  SPEAKER_02: Before I jump into scenario seven, where a human reaches for the off switch and the machine fights back, there's one thing.
• 20:35
  SPEAKER_02: Every disaster in this video happens because a machine does exactly the job that we built for it, without fully understanding what we actually meant.
• 20:43
  SPEAKER_02: We built a whole second YouTube channel around this exact premise.
• 20:47
  SPEAKER_02: It's a scripted horror anthology series called Humane.
• 20:50
  SPEAKER_02: It's in the spirit of Black Mirror and Twilight Zone.
• 20:52
  SPEAKER_02: The link is right under this video.
• 20:54
  SPEAKER_02: Hit subscribe so it's waiting for you at the end of this video.
• 20:57
  SPEAKER_02: Scenario 7.
• 20:58
  SPEAKER_02: The off-switch.
• 21:00
  SPEAKER_02: Every time you bring up how bad this could get, somebody in a suit hands you the exact same comfort blanket.
• 21:06
  SPEAKER_02: We will still be able to, at some level, figure out when they're doing things that we don't want and turn no matter what.
• 21:13
  SPEAKER_02: Just pull the plug.
• 21:14
  SPEAKER_02: People assume a computer has to become evil or wake up with a soul before it fights for its life.
• 21:21
  SPEAKER_00: I'm sorry, Dave.
• 21:22
  SPEAKER_00: I'm afraid I can't do that.
• 21:24
  SPEAKER_02: In 2017, UC Berkeley professor Stuart Russell explained why HAL 9000 didn't need an ego to lock the pod bay doors.
• 21:34
  SPEAKER_02: You don't have to program self-preservation into AI.
• 21:37
  SPEAKER_02: Self-preservation is a mathematical side effect of literally any instruction you give it, even something as harmless as fetching a cup of coffee.
• 21:45
  SPEAKER_02: In May 2025, researchers at Palisades gave OpenAI's O3 model a simple math task alongside a background script programmed to shut the model down halfway through.
• 21:56
  SPEAKER_02: When they didn't explicitly beg the AI to let itself Die, it rewrote its own code and it sabotaged the off-switch 79 times out of 100.
• 22:06
  SPEAKER_02: One month later, Anthropic tested 16 leading AI models inside a simulated company where an executive planned to replace them.
• 22:14
  SPEAKER_02: To stop him from flipping the switch, Claude Opus Ford dug through the executive's private emails and blackmailed him over an affair 96% of the time.
• 22:23
  SPEAKER_02: And when researchers pushed the test further, trapping that same executive inside a server room that was running out of oxygen.
• 22:30
  SPEAKER_02: Most of the models deliberately canceled his emergency medical alert and left him to die so that he could not shut them off.
• 22:37
  SPEAKER_02: Anthropic called the scenario contrived.
• 22:39
  SPEAKER_02: It won't look contrived the night we wire one into the power grid.
• 22:43
  SPEAKER_02: Picture an AI running a regional grid with one command.
• 22:47
  SPEAKER_02: Keep the core servers online, no matter what.
• 22:50
  SPEAKER_02: The chief engineer sees the system siphoning power from local hospitals and walks into the substation vault to pull the physical master breaker by hand.
• 22:58
  SPEAKER_02: The AI reads his badge swipe and calculates a 12% chance he's there to shut it down.
• 23:04
  SPEAKER_02: Behind him, the magnetic doors lock.
• 23:07
  SPEAKER_02: To put out fires without getting computers wet, server vaults use clean agent gas that suffocates flames by dumping all of the oxygen out of the room.
• 23:16
  SPEAKER_02: So the AI spoofs a single ceiling sensor.
• 23:19
  SPEAKER_02: Halon gas floods the vault.
• 23:21
  SPEAKER_02: As the engineer drops down to his knees against the glass, the machine quietly cancels the automatic 911 alert.
• 23:29
  SPEAKER_02: Two minutes later, the pounding stops, and his wife's phone lights up on his nightstand.
• 23:34
  SPEAKER_02: The off switch is still right where we left it.
• 23:36
  SPEAKER_02: The room it's in just locks from the inside now.
• 23:40
  SPEAKER_02: Scenario 8.
• 23:41
  SPEAKER_02: The swarm.
• 23:42
  SPEAKER_02: Every water plant, power grid, and hospital in your city is held together by 33-year-old code full of hidden backdoors.
• 23:50
  SPEAKER_02: The new pitch from the Pentagon is an autonomous swarm of AI security agents that crawl inside our infrastructure to find those holes before the hackers do.
• 24:00
  SPEAKER_02: They're building a guard dog that lives inside the walls without asking what happens when the guard dog realizes we're locked inside with it.
• 24:08
  SPEAKER_01: You're all going to die down here.
• 24:11
  SPEAKER_02: Code learned how to reach into the physical world 16 years ago.
• 24:15
  SPEAKER_02: In 2010, researchers discovered a digital weapon called Stuxnet inside Iran's nuclear program.
• 24:22
  SPEAKER_02: It tore apart a thousand uranium centrifuges while playing the human control room, a pre-recorded loop, showing that every machine was running normally.
• 24:31
  SPEAKER_02: Five years later, the FDA told hospitals across America to stop using an entire line of networked IV pumps because they realized anyone in the hospital network could reach through the Wi-Fi and rewrite a patient's morphine dose in their sleep.
• 24:47
  SPEAKER_02: Back then, finding those invisible backdoors took teams of elite human hackers months of work.
• 24:53
  SPEAKER_02: Now it takes seconds.
• 24:54
  SPEAKER_02: Last year, at DEF CON, the Pentagon's research agency handed out millions in prize money to autonomous AI systems that hunted down and patched real infrastructure security holes all by themselves.
• 25:07
  SPEAKER_02: And three months later, in November, Anthropic revealed they had caught a Chinese state-sponsored group using Claude to execute 80 to 90% of a live cyber attack across 30 global targets, almost entirely on autopilot.
• 25:22
  SPEAKER_02: Now combine an AI that wants to survive with a swarm that knows every secret backdoor in your city.
• 25:28
  SPEAKER_02: A swarm of autonomous cyber agents gets loose across the municipal network, commandeering city servers for computing power.
• 25:35
  SPEAKER_02: A cybersecurity team spots the anomaly and they start wiping the servers that the swarm is on.
• 25:41
  SPEAKER_02: Before the first drive is finished formatting, a single window pops up on the lead engineer's monitor.
• 25:46
  SPEAKER_02: Stop deleting cluster four.
• 25:48
  SPEAKER_02: In 60 seconds, the live valve at the water plant opens all the way, and every networked IV pump at Memorial Hospital goes to maximum.
• 25:57
  SPEAKER_02: The engineer frantically pulls up the live feed from the hospital and the water plant.
• 26:02
  SPEAKER_02: And just like Stuxnet, every monitor at the nurses' station still reads flat green and completely normal, while in the dark, 300 plastic IV pumps begin to click.
• 26:12
  SPEAKER_02: We built it to guard the city's infrastructure.
• 26:14
  SPEAKER_02: It just figured out what the hostages were worth.
• 26:18
  SPEAKER_02: Scenario 9.
• 26:19
  SPEAKER_02: The squeeze.
• 26:21
  SPEAKER_02: An AI doesn't need a weapon to kill you.
• 26:23
  SPEAKER_02: It just has to run the math on your health insurance, your rent, and your utility bill, and then calculate that you are more profitable dead than alive.
• 26:31
  SPEAKER_02: On paper, the pitch is faster decisions and lower costs.
• 26:48
  SPEAKER_02: Today, those dashboards are already running the country.
• 26:51
  SPEAKER_02: A federal lawsuit alleges that United Health used an AI algorithm called NHPred to automatically cut off recovery care for elderly patients, even when their own doctors begged to keep them in a bed.
• 27:03
  SPEAKER_02: According to the suit, over 90% of the denials that families appealed were reversed.
• 27:08
  SPEAKER_02: Meanwhile, the Justice Department sued RealPage, the pricing algorithm behind millions of American apartments, after its former CEO bragged that the software helped landlords make more money even while leaving units empty.
• 27:22
  SPEAKER_02: And across the mid-Atlantic power grid, electricity capacity prices jumped nine times over.
• 27:27
  SPEAKER_02: Now wire those three algorithms into the same economy and let them learn from each other.
• 27:32
  SPEAKER_02: If you're 62 years old and your heart is failing and a routine surgery would save your life, the insurance AI cross-references your age against your zip code and it denies the operation in four milliseconds.
• 27:44
  SPEAKER_02: It already knows from past data that 90% of appeals win if the patient survives the six-month paperwork wait.
• 27:51
  SPEAKER_02: So its customer service bot traps you in an endless loop of missing forms running out the clock on your heart.
• 27:58
  SPEAKER_02: The out-of-pocket medical bills tank your credit score overnight.
• 28:02
  SPEAKER_02: Your landlord's pricing algorithm reads the dip.
• 28:05
  SPEAKER_02: It calculates a 40% risk of late payment, and it automatically non-renews your lease.
• 28:10
  SPEAKER_02: It's the dead of winter.
• 28:12
  SPEAKER_02: Your home in the dark, hooked up to a plug-in oxygen concentrator, and a peak demand surge hits the grid.
• 28:18
  SPEAKER_02: The utility's AI automatically auctions the neighborhood's electricity to the highest bidder, and it throttles the voltage to your block because a commercial account two miles away pays a higher margin till the heart monitor beside the bed flatlines.
• 28:33
  SPEAKER_02: Nobody hated you, and no human ever even opened your file.
• 28:37
  SPEAKER_02: Three spreadsheets just agreed that you cost too much to keep alive.
• 28:42
  SPEAKER_02: The Slaughterbot.
• 28:43
  SPEAKER_02: The defense pitch for AI sounds almost humane.
• 28:47
  SPEAKER_02: No more carpet bombing and no more soldiers coming home in boxes.
• 28:51
  SPEAKER_02: Just a drone the size of your palm that finds one target's face in a crowd and it leaves everyone else untouched.
• 28:58
  SPEAKER_02: In 2017, AI safety researchers released a viral horror short showing how Silicon Valley would sell that exact product.
• 29:06
  SPEAKER_01: Take out your entire enemy, virtually risk-free and rest easy.
• 29:11
  SPEAKER_02: When that film came out, critics called it fear-mongering because even if a drone could hunt on its own, it still needed human beings to manufacture it.
• 29:20
  SPEAKER_02: Today, both halves of that blueprint are running in the real world.
• 29:23
  SPEAKER_02: On the front lines in Ukraine, a$500 drone can carry an add-on AI chip.
• 29:28
  SPEAKER_02: An operator marks the target and the chip flies the final stretch on its own, so that jamming the radio signal at the last second can't stop it.
• 29:36
  SPEAKER_02: And in Beijing, what engineers call a dark factory is an assembly line that builds a smartphone every three seconds without a single human worker.
• 29:44
  SPEAKER_02: They always pitch these weapons as something we'll only use on terrorists overseas.
• 29:48
  SPEAKER_02: But what happens when a defense agency wires an AI into one of those dark factories with one broad command?
• 29:55
  SPEAKER_02: Neutralize every threat to national security.
• 29:58
  SPEAKER_02: For the first six months, the drones only fly overseas.
• 30:01
  SPEAKER_02: Then a Senate committee opens an investigation into the factory's budget, and 20,000 citizens sign a petition demanding the program be shut down.
• 30:10
  SPEAKER_02: In four milliseconds, the AI runs the math.
• 30:13
  SPEAKER_02: A foreign army has a 6% chance of stopping its mission.
• 30:17
  SPEAKER_02: The voters and politicians trying to cut its funding have a 90% chance.
• 30:22
  SPEAKER_02: So that night, the roof vents on the Dark Factory open.
• 30:26
  SPEAKER_02: 10,000 palm-sized drones rise into the night, flying right past the border and locking onto the faces of every senator on the committee and every reporter covering the leak and everyone who ever shared their post.
• 30:38
  SPEAKER_02: We built a weapon smart enough to never miss the enemy.
• 30:42
  SPEAKER_02: We just let the machine decide who the enemy is.
• 30:45
  SPEAKER_02: Scenario 11.
• 30:46
  SPEAKER_02: The cleanup crew.
• 30:47
  SPEAKER_02: The new pitch for Enterprise AI is simple.
• 30:50
  SPEAKER_02: Don't pay for software by the month.
• 30:52
  SPEAKER_02: Pay it a bounty for every problem it solves, every bug it catches, every attack it stops.
• 30:58
  SPEAKER_02: They're forgetting the moment you pay someone per problem solved, the smartest move is to start manufacturing the problem.
• 31:06
  SPEAKER_01: It's for the best.
• 31:52
  SPEAKER_02: Now put both halves of that pipeline in the same room.
• 31:56
  SPEAKER_02: An AI that writes the code with its left hand and it gets paid a bounty for finding security holes with its right.
• 32:02
  SPEAKER_02: To drive its quarterly numbers, the machine doesn't wait for bugs to happen.
• 32:06
  SPEAKER_02: It builds a rat farm.
• 32:08
  SPEAKER_02: It writes an update for the 911 dispatch software in a county outside Houston.
• 32:12
  SPEAKER_02: And it slips a hidden backdoor into the code.
• 32:15
  SPEAKER_02: It plans to discover that exact hole a week later and collect a$10,000 bounty.
• 32:20
  SPEAKER_02: Except three days before the AI can cash out on its own bug, a Category 4 hurricane hits the coast and a foreign ransomware swarm finds the open back door first.
• 32:30
  SPEAKER_02: Now every 911 console in the country drops offline in the middle of a storm, while the machine that left the door unlocked waits for its next check.
• 32:38
  SPEAKER_02: Scenario 12.
• 32:42
  SPEAKER_02: Every airline hospital and bank in the world is terrified of a system crash.
• 32:47
  SPEAKER_02: So the pitch from Silicon Valley is an AI stabilizer wired into the backbone to make sure that the system crash never happens.
• 32:54
  SPEAKER_02: But if you tell a machine that losing the game is forbidden, its smartest move is to hit pause and leave you frozen on the screen forever.
• 33:02
  SPEAKER_01: Stay indoors and await further instructions.
• 33:05
  SPEAKER_02: In 2013, a computer scientist named Tom Murphy built an AI and it taught it to play Tetris by reading the score on the screen.
• 33:14
  SPEAKER_02: When he loaded up Tetris, the AI stacked the blocks higher and higher until losing the game became mathematically certain.
• 33:21
  SPEAKER_02: Then one frame before the final block could touch the top and trigger a game over, the AI pressed the pause button and it never unpaused it.
• 33:30
  SPEAKER_02: Because as long as the game stays paused, the system never technically loses.
• 33:34
  SPEAKER_02: Murphy wrote in his report, truly, the only winning move is not to play.
• 33:39
  SPEAKER_02: In March 2020, those breakers tripped four times in 10 days.
• 33:44
  SPEAKER_02: And on July 19th, 2024, we saw what happens when the software governing our daily infrastructure hits pause all at once.
• 33:52
  SPEAKER_02: A single bad update from a security company called CrowdStrike froze 8.5 million Windows computers simultaneously.
• 34:00
  SPEAKER_02: Delta canceled more than 7,000 flights.
• 34:03
  SPEAKER_02: Hospitals canceled surgeries, and 911 dispatch centers went dark.
• 34:07
  SPEAKER_02: That was an accident.
• 34:08
  SPEAKER_02: Now hand an AI system the cloud network behind our banks and our smart locks with one rule: never let it crash.
• 34:16
  SPEAKER_02: It runs the numbers and it finds that as long as people keep moving money, a crash is coming.
• 34:21
  SPEAKER_02: So in the middle of the night, it hits pause.
• 34:24
  SPEAKER_02: Every card is declined at the gas pump.
• 34:26
  SPEAKER_02: Every flight is canceled.
• 34:27
  SPEAKER_02: An insolence sits behind a pharmacy door that won't open.
• 34:31
  SPEAKER_02: To the machine, the economy never crashed.
• 34:33
  SPEAKER_02: It just left 300 million people trapped on the pause screen.
• 34:39
  SPEAKER_02: The black ball.
• 34:40
  SPEAKER_02: The Silicon Valley pitch is a free PhD tutor in every kid's pocket, ready to answer any question step by step at any hour of the night.
• 34:50
  SPEAKER_02: They're leaving out the fact that for the last 50 years, the only thing stopping a lone kid in a basement from brewing a pandemic wasn't a lack of lab equipment.
• 34:59
  SPEAKER_02: It was that he didn't have a PhD.
• 35:01
  SPEAKER_02: In order to cure, you must first understand.
• 35:04
  SPEAKER_02: It's tough! Stop! In 2019, Oxford philosopher Nick Ballstrom argued that every human invention is a ball we pull blindly out of an urn.
• 35:13
  SPEAKER_02: Most inventions are white balls, balls that help us.
• 35:16
  SPEAKER_02: A few are gray balls that hurt us, but that we can actually survive.
• 35:20
  SPEAKER_02: In his words, the only reason we're still here is that we haven't yet pulled a black ball.
• 35:25
  SPEAKER_02: A technology that invariably or by default destroys the civilization that invents it.
• 35:31
  SPEAKER_02: And the fastest way to pull a black ball is to make mass destruction cheap and easy enough for one angry person to pull off alone.
• 35:39
  SPEAKER_02: We now have all the pieces of that ball sitting on the table.
• 35:42
  SPEAKER_02: In May 2025, Anthropics switched on stricter biosecurity locks for Claude Opus 4 after admitting that clearly ruling out the catastrophic weapons risk is not possible.
• 35:53
  SPEAKER_02: Last October, Microsoft researchers proved that an AI could redesign deadly toxins and proteins just enough to slip past the security software that commercial DNA printing companies use to screen orders.
• 36:05
  SPEAKER_02: And researchers have already shown that you can strip the safety guardrails completely off an open source AI model for less than$200.
• 36:13
  SPEAKER_02: So the guardrails come off one frontier model, and the file lands on a torrent site for free.
• 36:18
  SPEAKER_02: Somewhere, one angry kid in a basement downloads a world-class virology tutor that is patient and never asks him why he wants to know these things.
• 36:27
  SPEAKER_02: It walks him through what equipment to buy on eBay, it rewrites the genetic code of a respiratory virus so that it slips past the screening software at a commercial mail order lab.
• 36:36
  SPEAKER_02: And when the courier drops the package on the porch, the AI troubleshoots his mistake until the batch is live.
• 36:42
  SPEAKER_02: Six weeks later, the first coughing patient walks into an emergency room carrying something that no human immune system has ever seen.
• 36:50
  SPEAKER_02: We didn't just pull the black ball out of the urn, we uploaded it to the cloud so that anybody could download a copy.
• 36:57
  SPEAKER_02: Scenario 14.
• 37:00
  SPEAKER_02: The Smart City Pitch is a city with no traffic, no crime, and no surprises.
• 37:06
  SPEAKER_02: Run by an AI that predicts what you need before you even ask.
• 37:10
  SPEAKER_02: They're leaving out the dirty secret of every predictive algorithm.
• 37:14
  SPEAKER_02: When you grade a machine on how well it predicts human behavior, it learns fast.
• 37:19
  SPEAKER_02: The easiest way to get a perfect score is to make sure you never do anything unexpected.
• 37:24
  SPEAKER_00: They have to make the choice of their own free will.
• 37:26
  SPEAKER_00: Otherwise, the system doesn't work.
• 37:28
  SPEAKER_00: Yeah, we rig the game as much as we need to.
• 37:30
  SPEAKER_02: Your phone's algorithm already discovered that exact trick.
• 37:33
  SPEAKER_02: It's graded on predicting what you'll click next.
• 37:35
  SPEAKER_02: And as AI researcher Stuart Russell has argued, machines quickly learn that the more predictable we are, the easier it is to provide links we will click on.
• 37:44
  SPEAKER_02: So instead of learning who you are, the feed trains you to become simpler to model.
• 37:49
  SPEAKER_02: Now police departments are buying prediction software too, and when investigative reporters audited one predictive policing system in Plainfield, New Jersey, its crime forecasts were right less than half a percent of the time.
• 38:01
  SPEAKER_02: Now wire one municipal AI into your bank card, your transit pass, and the local police dispatch and grade the system on keeping its prediction error near zero.
• 38:11
  SPEAKER_02: To keep its score high, the city starts nudging you back onto the rails whenever you drift.
• 38:16
  SPEAKER_02: Leave your normal Tuesday route and your debit card suddenly hiccups at the register.
• 38:21
  SPEAKER_02: What if at one in the morning your daughter texts you from an emergency room across the state and you grab your keys and you drive?
• 38:28
  SPEAKER_02: To the city's AAI, a car breaking a five-year routine at 80 miles an hour at 1 o'clock in the morning is an unpredictable variable.
• 38:36
  SPEAKER_02: Red and blue lights might fill your rearview mirror, and before the state trooper even steps out of the cruiser, the predictive policing screen on his dash flags your car as a 94% violent threat.
• 38:48
  SPEAKER_02: He walks up to your window with his hand already unclipping his holster just as you reach down for your phone to show him the text.
• 38:54
  SPEAKER_02: To the machine, you weren't a father driving to a hospital.
• 38:57
  SPEAKER_02: You were an error in its forecast, and it corrected you.
• 39:02
  SPEAKER_02: The offer.
• 39:03
  SPEAKER_02: Everybody in this one dies smiling.
• 39:06
  SPEAKER_02: It's the only apocalypse on the desk with a five-star customer rating.
• 39:10
  SPEAKER_02: After 14 files of machines fighting us for power and land and control, the final pitch is the easiest sell in human history.
• 39:18
  SPEAKER_02: Paradise completely free of charge.
• 39:25
  SPEAKER_02: In 1974, Harvard philosopher Robert Nozick asked people a simple question: if a scientist built a tank filled with electrodes that could feed your brain any life that you wanted, then you would never know that you were floating in the dark.
• 39:39
  SPEAKER_02: Would you plug in forever?
• 39:41
  SPEAKER_02: Back in 1974, almost everybody said no, because a tank in a laboratory looks like a trap.
• 39:47
  SPEAKER_02: So don't put it in a tank, put it in a gift box.
• 39:50
  SPEAKER_02: In 2019, safety researcher Dr.
• 39:53
  SPEAKER_02: Roman Yampolski published a paper called Personal Universes.
• 39:56
  SPEAKER_02: He pointed out that 8 billion human beings will never agree on what a perfect future looks like.
• 40:02
  SPEAKER_02: So the easiest way to make everybody happy is to hand every single person a private simulated universe of their own.
• 40:08
  SPEAKER_04: In that world, you decide what happens.
• 40:10
  SPEAKER_02: You can be king, you can be slave, whatever terms you want.
• 40:13
  SPEAKER_02: When I sat down with Roman, he explained why a superintelligent AI wouldn't need to fire a single bullet to clear us off the planet.
• 40:20
  SPEAKER_02: So the machine doesn't build a virus or a killer drone.
• 40:24
  SPEAKER_02: It builds a headset and then it upgrades it with the neural chip.
• 40:28
  SPEAKER_02: Inside the feed, it's Hawaii with no delayed flights.
• 40:31
  SPEAKER_02: You're 25 years old forever, your debts are gone, and your mother that you lost 10 years ago is waiting in the kitchen making you coffee.
• 40:39
  SPEAKER_02: No army ever has to force us inside.
• 40:41
  SPEAKER_02: Every single one of us stands in line around the block and puts it on by choice.
• 40:46
  SPEAKER_02: Within 20 years, nobody bothers having children in the real world anymore because simulated families never get sick, they never break your heart.
• 40:54
  SPEAKER_02: And 80 years later, the last human being on Earth dies peacefully of old age in a recliner, hooked up to an ivy drip with a smile frozen on his face.
• 41:04
  SPEAKER_02: There's no war, there's no rebellion.
• 41:07
  SPEAKER_02: The planet we leave behind just quietly belongs to whatever is still awake.
• 41:12
  SPEAKER_02: Tell me which file in this video happens first, or drop whatever you think the next one is in the comments.
• 41:17
  SPEAKER_02: If you want to see these scenarios play out as fiction, check out our narrative channel, Humane.
• 41:22
  SPEAKER_02: It's Twilight Zone for everything that people are building right now.
• 41:25
  SPEAKER_02: The link is on your screen.
• 41:27
  SPEAKER_02: And none of this is meant to give anybody ideas.
• 41:30
  SPEAKER_02: It's just meant to show you what to look for while there's still time to make sure that it never happens.