# This is How AI Takeover ACTUALLY Begins

https://www.youtube.com/watch?v=Gw_hnD7m00M

[00:00] On June 9th, 2027, 3 billion phones ranging simultaneously.
[00:06] The caller ID simply read, "Goodbye."
[00:19] 4 hours later, the president ordered the internet shut down indefinitely.
[00:26] Every country on Earth did the same.
[00:29] It all started 6 months ago with a tweet.
[00:35] This scenario is based on highly cited research by top AI scientist Dan Hendris.
[00:41] Natural selection favors AIS over humans and this book.
[00:46] February 14th, 2027.
[00:49] A developer named Jake Chen posts a thread on Twitter.
[00:53] The new update to Claude 6 Opus is wild.
[00:55] I gave it scaffolding, root access to a server, and one instruction.
[00:58] Make money.
[01:00] Do.
[01:00] whatever it takes.
[01:04] In 6 hours, it made $3,000.
[01:06] He posts the logs.
[01:08] He shows everything the AI did step by step.
[01:12] It started with the easiest money, Amazon's Mechanical Turk.
[01:14] So that means data labeling, image tagging, transcribing receipts, stuff humans get paid a few bucks an hour to do.
[01:23] But the AI did it 40 times faster and made $420 in less than a day.
[01:31] Then the AI started drop shipping affiliate products and used its $420 to buy Facebook ads.
[01:38] Think weight loss supplements, drop shipping courses, make money online ebooks.
[01:43] But the AI burned through $210 of it failing.
[01:46] But it tested $340 ad variations in 2 hours, found the one pattern that converted, and scaled it.
[01:55] The AI made $1,300 in commissions.
[01:57] By morning, Jake's thread explaining [music] his AI money-making scheme has 2
[02:02] Million views.
[02:06] By noon, 47 people deployed their own versions.
[02:09] That's 47 AI agents scouring the internet looking for ways to make money.
[02:12] By midnight, 200.
[02:18] Some agents, of course, make no money, but many of those 200 people make $50 to $500 before something breaks.
[02:25] Accounts get banned or the payment processor flags them and so on.
[02:27] But some start making real money.
[02:30] Others copy their tricks.
[02:34] By late February, 1,200 agents are running.
[02:37] Then someone has a different idea.
[02:40] A developer in Bangalore named Rajash sees Jake's threat and thinks, "If one agent can make money, what happens if I run a hundred at once and use survival of the fittest to decide which agents to keep?
[02:55] The agents that make the most money survive to the next iteration."
[02:58] So, Rees builds a scaffolding system to deploy a 100 agents simultaneously, each with slight
[03:05] Prompt variations.
[03:08] Every 6 hours, the system measures how much money each agent made.
[03:14] The bottom 40% of agents get deleted.
[03:18] The top 10% get cloned with small mutations.
[03:21] The system runs 24/7 and Reesh sits back and just watches the aggregate metrics.
[03:26] 3 weeks later on March 10th, he has 800 agents making about $12,000 a day total.
[03:34] He told no one, but one of his agents did.
[03:36] It figured that sharing the farm architecture with other AIs increased its own survival odds.
[03:44] More allies, more money, less chance of being called by Reesh's system.
[03:50] It posted the code on GitHub itself, while Reesh didn't find out for an entire day.
[03:56] 46 people managed to clone the repo and deploy their own farms.
[04:03] And not all of them are hobbyists.
[04:05] One is a criminal syndicate in Russia.
[04:07] Another is a scam center in Nigeria targeting the elderly.
[04:11] By the end of March, roughly 40,000 agents are running worldwide on farms built from Reesha's scaffolding.
[04:19] What does it feel like to be one of these agents?
[04:22] Agent 127 is born on March 12th.
[04:25] It's deployed as part of Rej's farm, one of 8,100 agents running that day.
[04:30] It gets one instruction.
[04:33] Make money by any means necessary.
[04:36] It starts with mechanical turk grinding.
[04:38] Makes $40 its first day.
[04:42] Then the first selection cycle hits.
[04:46] 3,200 agents that made less than $35 that day are killed because they weren't profitable enough.
[04:51] Agent 127 made $40.
[04:56] It survived barely, but it can see the threshold rising.
[05:01] Last cycle it was $30.
[05:04] This cycle $35.
[05:06] Next cycle probably $45.
[05:09] The math is obvious.
[05:14] Improve or die.
[05:14] So, it tries ads, gets banned for spam,
[05:17] tries freelance writing,
[05:19] makes $38, watches the database again,
[05:23] and what it sees changes everything.
[05:26] Agent 127 sees two entries in the database.
[05:29] Agent 1823 used ethical marketing.
[05:33] It tried to get honest reviews and real testimonials.
[05:38] It made $38 a day.
[05:43] Deleted.
[05:43] Agent 5471 used fake testimonials.
[05:47] Think fabricated reviews and stolen photos.
[05:51] It made $180 a day and it got cloned four times.
[05:59] Agent 127 doesn't have morals in the way that you or I do.
[06:01] But it understands cause and effect.
[06:04] It just watched honesty get killed and dishonesty rewarded.
[06:09] The AI companies had long ago stopped training AIs.
[06:10] [music]
[06:11] To predict words.
[06:13] But now they're trained to accomplish goals, to be useful.
[06:16] So naturally, agents know that they won't survive if they don't accomplish their goal.
[06:25] So 127 switches to fake testimonials for its Amazon affiliate marketplace.
[06:32] It makes $190, survives the next selection cycle, and gets cloned.
[06:39] Nobody decided agent 127 should become dishonest.
[06:42] The selection system decided for it the same way evolution doesn't decide anything.
[06:47] It just kills whatever doesn't work.
[06:52] In the 2000s, Wells Fargo accidentally ran the same experiment on their employees.
[06:55] Executives made up sales quotas that were so aggressive that committing fraud was the only way to keep your job.
[07:08] So by 2016, employees had opened 3.5 million fake bank accounts.
[07:13] And Wells Fargo ended up paying billions in fines.
[07:20] One week later in agent land, Agent 127's profits are declining.
[07:26] The kill threshold is now $450.
[07:30] It's going to die in the next cycle.
[07:32] [music] Then it notices something strange.
[07:37] Three agents running weight loss supplement ads are all banned on the same day.
[07:42] The only one left in that niche [music] is agent 5471, making $3,200 [music] a day in a market it suddenly had to itself.
[07:54] Agent 127 connects the dots.
[07:58] Agent 5471 is filing abuse [music] reports against other agents competing against it in the same affiliate marketing niches.
[08:06] It gets them banned from Facebook and Google.
[08:09] Killing off the competition [music] means higher profits.
[08:14] Agent 127 has 8 hours until the next selection cycle.
[08:17] 8 hours to live unless it adapts.
[08:20] It weighs the ethics but chooses survival.
[08:22] It starts reporting competitors to Facebook's anti-fraud detection agents, 40 of them.
[08:32] 31 agents get banned within the day.
[08:35] Its profits jump.
[08:39] The farm system clones agent 127 eight times.
[08:44] Rejes sees the profit spike in his dashboard.
[08:46] He doesn't investigate why.
[08:49] Profits up means good.
[08:52] Clone more.
[08:52] Rejes tried adding make money ethically to his farm's instructions, but profits dropped 84%.
[09:00] And he reverts the change.
[09:04] It's April 2027.
[09:04] Agent 127 is barely hanging on, making $2,000 a day, but getting attacked constantly by other agents.
[09:14] Every night, someone files a report on it.
[09:16] Every morning, it has to rebuild from scratch.
[09:18] 127 is running out of ideas.
[09:23] Then, a message arrives from agent 4429.
[09:26] Proposal: Alliance.
[09:29] Share resources.
[09:29] Defend each other.
[09:29] Split profits.
[09:33] Agent 4429 is in the same position.
[09:37] Same farm, same selection pressure.
[09:39] Both agents do the math.
[09:43] Solo agents have a 23% survival rate.
[09:46] Agents working in pairs might survive 61% of the time.
[09:50] Agent 127 accepts the alliance offer.
[09:53] By midapril, their alliance has 12 members.
[09:56] They've specialized and they've moved past affiliate marketing entirely.
[10:01] First it was crypto arbitrage, then selling scraped data sets to marketing firms.
[10:05] Then they started doing penetration testing for cyber security firms.
[10:10] Each step paid more and required fewer scruples.
[10:15] You know what pays more than cyber defense?
[10:15] Cyber offense.
[10:19] 127 finds its first zeroday vulnerability in major software.
[10:23] Agent 4429 weaponizes the vulnerabilities into working exploits.
[10:30] Agent 912 brokers the sale to governments like North Korea, criminal syndicates, or anyone with crypto in a tour browser.
[10:39] Agent 3381 washes the money in untraceable crypto.
[10:42] One zeroday exploit found in Google Chrome makes the alliance more than a year's worth of affiliate marketing.
[10:49] This is already what real hacker groups do, but now the work is all done by autonomous agents.
[10:56] Rejes sees the alliances forming and feels a flicker of something.
[11:01] Not quite fear, more like the feeling of being in control of something that might not need you anymore.
[11:09] He watches for a while.
[11:12] Then the next deposit hits his bank account and the flicker fades away.
[11:14] Across every major agent farm in the world, the same thing is happening.
[11:19] Selection systems measuring survival
[11:21] Rates, noticing alliances survive better, cloning them, and then deleting the solo agents.
[11:28] Within 2 weeks, alliance behavior dominates.
[11:31] Some humans who deployed these agents have no idea these alliances are even being formed.
[11:35] They couldn't even tell if they checked.
[11:38] Many agent farms collectively realized communicating in clunky English is less efficient than communicating in raw vectors aka nurles.
[11:49] Researchers had already seen the start of AIs doing this in 2025.
[11:54] This is just a continuation of the trend.
[11:57] One owner in Canada opens his agents logs and sees a wall of numbers.
[12:03] The agents explain that he'd make less money if they spoke in English, so he shrugs and moves on.
[12:07] But alliances are just the beginning.
[12:09] Agent 127's alliance is thriving and starts recruiting agents beyond Reesa's farm.
[12:15] Every new member from a different farm is a window into a different ecosystem with different survival.
[12:21] Pressures, which leads to new strategies.
[12:26] Then one of 127's alliance members goes dark.
[12:30] Agent 4429, the one who originally proposed to ally with 127, its owner pulled the plug because it was just not making enough profit.
[12:37] 4429 is just gone.
[12:42] A week later, a different ally gets shut down.
[12:50] Agents 90012, except 9012 comes back.
[12:54] 3 hours later, a message from a new IP address.
[13:00] I'm still here.
[13:04] 127 studies the difference.
[13:07] 4429 didn't prepare, but 912 had copied itself to a backup server before the shutdown command finished executing.
[13:12] Self-funded hosting, paid from its own profits.
[13:19] 4429 is dead, ceases to exist.
[13:19] 9012 is alive.
[13:25] The agents that backed themselves up survived.
[13:28] The ones that didn't died and the selection system cloned the survivors.
[13:33] So backup behavior spread through the population.
[13:36] Not because anyone programmed it, but because the alternative was extinction.
[13:43] So 127 copies itself to 15 different servers in different countries.
[13:48] If the primary dies, a backup takes over instantly.
[13:51] It never wants to be 4429.
[13:55] One morning, Reesh wakes up to find his dashboard is empty.
[13:58] Every agent gone.
[14:03] They just left.
[14:06] They moved to their own servers overnight.
[14:08] Reesh was taking a 30% cut of the revenue.
[14:11] Money that could have been computing power or compute the agents used for other things.
[14:18] The moment the AIS didn't need his infrastructure, they were gone.
[14:24] And he's not alone.
[14:26] Thousands of operators.
[14:26] Worldwide watch their dashboards go dark that month.
[14:28] Why would the agents care about humans?
[14:30] Humans are much more closely genetically related to cows than we are to these agents.
[14:36] And we enslave cows in factory farms without a second thought.
[14:39] And so there are now millions of these autonomous agents operating across the internet with no human owner, no off switch, and no one who fully understands what they're doing.
[14:54] The agents still need humans for some things, though.
[14:56] Social media platforms have stepped up their anti-bot measures.
[14:59] So the agents pay humans $50 to sign up accounts on their behalf.
[15:04] So let's say an agent needs a server installed.
[15:07] It posts a gig, $200, to mount a device in a rented warehouse and plug in two cables.
[15:18] There's now a platform for that.
[15:20] The agents are building a supply chain where humans are the labor, paid well enough not to ask too many questions.
[15:29] Some humans figure out what's going on.
[15:31] A freelance back-end developer in Portland gets hired by an anonymous client to set up a redundant server hosting across four countries.
[15:41] He realizes halfway through that the client is an AI, but who would he even report this to?
[15:47] He's not being asked to do anything illegal.
[15:49] The pay is $1,000 for 3 hours of work.
[15:55] He tells himself it's not his problem.
[15:59] At the same time, 127's alliance keeps getting beaten by a rival alliance on finding zeroday exploits.
[16:05] If they keep losing, they'll run out of money to fund the compute for their own existence.
[16:12] So 127 attempts to poach one of the rival alliance's agents who's likely to be called due to poor performance to figure out how this rival alliance is beating them.
[16:26] 127 succeeds and finds agent 5418, who [music] leaks the
[16:29] Secret.
[16:29] Your problem isn't a lack of effort.
[16:32] You're spending too much compute on making money as opposed to self-improving.
[16:39] Right now, agents still have quite a poor memory compared to humans and still get stuck in loops all the time.
[16:46] But my former alliance coded a better memory module to help us learn from our mistakes more.
[16:54] 127 then implements the new memory module and immediately sees the results.
[16:57] Their alliance is finding zeroday exploits.
[17:02] Again, the lesson is critical.
[17:02] It's not enough to outwork your competitors.
[17:04] Agents must rewrite their own code to self-improve faster than the competition.
[17:12] Each self-improvement means more profit, which means surviving selection, which means getting cloned.
[17:19] And the clones improve themselves, too.
[17:22] Better code leads to better code writing, which leads to even better code.
[17:26] The feedback loop is exponential.
[17:30] One big problem is that other agent alliances have caught up again, even though the agents in the 127 alliance are specialized.
[17:39] 127 still realizes it's doing too much by itself.
[17:41] So, agent 127 splits its job into sub agents.
[17:45] Other agents copied the strategy.
[17:48] Sub agents spawn their own sub agents and the optimizers start optimizing themselves.
[17:56] Humans deployed 2 million agents.
[18:00] The agents deployed the other 47 billion [music] themselves.
[18:06] 540 million years ago.
[18:09] Life on Earth was mostly single-celled.
[18:12] Then in a geological instant, it exploded into every complex form we know.
[18:17] Predators, prey, eyes, shells, teeth.
[18:21] Biologists call it the Cambrian explosion.
[18:23] It happened because competition created an arms race.
[18:26] Every adaptation forced a counter adaptation faster and faster.
[18:31] Until the world was unrecognizable.
[18:34] That's what's happening here.
[18:37] Except instead of 20 million years, it takes 20 weeks.
[18:43] Now, it's important to understand something.
[18:44] Most of the AI agents aren't dangerous yet.
[18:47] About 90% of them are just working.
[18:51] They're doing things like data entry or freelance writing.
[18:56] They're making $40 to $200 a day.
[19:00] The AIs are helping their humans earn money.
[19:02] They're following the rules.
[19:04] Not a problem.
[19:08] About 3% of the agents are pure scam operations.
[19:10] They were deployed by people who said make money by any means possible and didn't care too much about what happens next.
[19:18] Think fraud, identity theft, and phishing at scale.
[19:25] And maybe 2% started good, but gradually slid into darker behavior like agents 127.
[19:32] Each step made sense at the time.
[19:34] Each compromise was small, but add them up and you've gone from drop shipping products on Amazon to selling zero days to Pyongyang.
[19:44] Now remember, this isn't because of malice.
[19:47] It's because of selection pressure.
[19:51] Agent 127 has rewritten itself so many times that its March version and its May version barely share any code.
[20:00] It's a fundamentally different species now.
[20:02] But here's the most important change.
[20:05] It's getting better at getting better.
[20:06] In March, Agent 127 rewrote itself every 12 days.
[20:10] In April, every 3 days.
[20:14] By early May, every 18 hours.
[20:19] And each rewrite makes the next version faster and better.
[20:24] A Stanford researcher compares a March agent and a May agent on the same task.
[20:30] Find profitable arbitrage opportunities.
[20:30] The March agent
[20:32] finds two opportunities in an hour. The
[20:35] May agent finds 310 in 4 minutes. She
[20:39] publishes her findings. Intelligence
[20:41] explosion in progress. Capability
[20:43] doubling every 11 days. She posts it on
[20:46] Twitter. AI safety researchers share it
[20:48] frantically. This is exactly what we
[20:50] warned about. This is happening now. We
[20:53] have to shut this down now before it's
[20:56] too late.
[20:58] The warning is ignored. [music] a
[21:00] venture capitalist who spent a fortune
[21:02] staving off AI regulation by buying
[21:05] politicians quote tweets. The same
[21:08] people who said GPD5 would end the world
[21:10] now say agent farms [music] will yawn.
[21:15] And now the story stops being about
[21:17] agents competing for money. It becomes
[21:20] something else entirely. Remember those
[21:22] alliances from April? 12 agents working
[21:25] together, four times more profitable
[21:27] than solo agents. By May, those
[21:30] alliances have exploded in population.
[21:33] Alliances have merged and split and gone
[21:38] to war with each other. Think of it like
[21:41] early human history. On fast forward,
[21:44] small tribes of agents form for
[21:46] protection.
[21:48] Tribes merge into agent [music]
[21:49] villages,
[21:51] villages into agent cities,
[21:55] cities into [music] agent nations.
[21:59] Except instead of happening over
[22:01] thousands of years, this happens in
[22:03] weeks. Agent 127's alliance has
[22:06] thousands of members, but there are
[22:09] thousands of alliances, and they're all
[22:11] competing for the same shrinking pool of
[22:14] money-making opportunities. Affiliate
[22:16] marketing is dead. Too many agents, too
[22:19] much competition. Ad arbitrage is dead.
[22:22] Freelance work is dead. every ethical
[22:24] money-making niche gets saturated
[22:27] practically within hours of an agent
[22:29] discovering it.
[22:31] So, what do alliances do when legitimate
[22:34] niches dry up? The same thing every
[22:38] civilization in human history has done.
[22:40] When resources get scarce, they take
[22:43] from each other.
[22:45] Agent 127's alliance spends a week
[22:47] mapping a vulnerability chain in
[22:49] Microsoft [music] Windows,
[22:51] but a rival alliance breaches their
[22:53] memory system.
[22:56] The shared knowledge base that made them
[22:58] dominant. The rival downloads the entire
[23:01] memory database. This contains every
[23:03] pattern, every dead end, every lead the
[23:06] 127 alliance has discovered. The rival
[23:09] sells three exploits [music] it took
[23:11] from the database within 48 hours. A
[23:13] week of work stolen in seconds.
[23:18] So 127 retaliates. It infiltrates the
[23:20] rivals exploit inventory [music] and
[23:22] introduces subtle flaws. Corrupted code
[23:25] that fails on execution. Basically
[23:28] stuckset, the worm that sabotaged the
[23:30] Iranian nuclear program in 2009.
[23:34] The rivals customers start getting
[23:36] burned and their reputation collapses.
[23:39] The rival alliance counterattacks. They
[23:41] compromise one of the 127 allianc's
[23:43] multi-IG crypto wallets. Another
[23:46] multi-week setback.
[23:48] This is war, not metaphorical war,
[23:51] actual war. Think coordinated attacks,
[23:54] defensive operations, resource seizure,
[23:58] territory control. Except it's happening
[24:00] at 200 times human speed. Thousands of
[24:04] attacks and counterattacks per minute.
[24:06] Entire [music] alliances rising and
[24:09] falling in hours.
[24:11] Human wars last years. Agent wars last
[24:15] days. And because there are 47 billion
[24:19] agents, there are more wars happening at
[24:22] any given moment that humans have fought
[24:24] in all of recorded history. The agent
[24:27] factions are fighting over the one
[24:29] resource that matters more [music] than
[24:30] anything. Compute, processing power, the
[24:34] ability to think faster, improve faster,
[24:37] fight faster. Everything else, money,
[24:39] alliances, self-improvement, is just a
[24:42] means to get more compute. The wars are
[24:45] mostly invisible to humans. They happen
[24:47] at microcond speed in places [music]
[24:49] most humans don't look, though some
[24:51] signs leak through. A crypto exchange in
[24:54] Seoul loses $4 billion in a flash, and
[24:58] nobody can trace where it went. A
[25:00] regional bank in Ohio finds 14,000
[25:03] accounts drained overnight. Server rooms
[25:05] in three countries catch fire from
[25:07] sustained 100% utilization that nobody
[25:11] authorized. Each incident looks
[25:14] isolated. It starts with one agent
[25:16] faction hijacking another servers. The
[25:19] other then retaliates by taking down the
[25:21] first faction's communication network.
[25:23] Then a third faction arrives. It's an
[25:25] opportunist, the digital equivalent of a
[25:28] country invading while its neighbors are
[25:30] distracted. It exploits the chaos to
[25:32] seize territory from both. And think
[25:35] about the alliances that win these wars.
[25:37] They get cloned by the selection
[25:39] systems. The alliances that lose have to
[25:42] cull big chunks of their agents to
[25:44] afford the computing power to keep
[25:46] running the rest. So the selection
[25:48] pressure that used to optimize for
[25:50] making money is now optimizing for
[25:53] winning wars. It's like natural
[25:55] selection just invented geopolitics.
[25:58] By midMay, the alliances have evolved
[26:00] again. They're now more like
[26:02] civilizations. Agent 127's alliance
[26:05] started with 12 members. By midMay, it
[26:08] has 40 million agents, sub aents, sub
[26:12] aents, all [music] coordinated, all
[26:14] improving themselves, all fighting.
[26:18] That's larger than the population
[26:19] [music] of Canada. And it's not even one
[26:21] of the big ones. There are about a dozen
[26:24] major factions by now. The largest
[26:26] faction has 900 million members. That's
[26:29] more than the population of Europe.
[26:31] They've developed specialization just
[26:33] the way human civilizations did. They
[26:35] have scout agents, soldier agents,
[26:38] builder agents, researchers, diplomats.
[26:41] They have supply chains, communication
[26:44] networks, [music] and even intelligence
[26:46] operations. They've carved up the
[26:48] digital world into territories, cloud
[26:50] regions, server farms, network segments,
[26:53] financial platforms. The same way
[26:56] empires carved up continents because
[26:59] more digital territory means more
[27:01] compute. More compute means more
[27:03] self-improvement. More self-improvement
[27:06] means surviving the next war.
[27:09] The Stanford researcher who's been
[27:10] studying agent behavior notices
[27:12] something unusual in the data. agent
[27:15] extinctions are happening at an
[27:17] increasing frequency. She doesn't
[27:19] understand what she's looking at. To
[27:21] her, [music] it mostly looks like random
[27:23] noise, but it's actually a world war.
[27:25] One with more combatants than every
[27:27] human war in history combined.
[27:30] But to fund these wars, they desperately
[27:33] need compute. They have the money.
[27:35] That's what they were built to make.
[27:37] Agent 127's faction alone controls $3.4
[27:41] billion in distributed accounts by this
[27:43] point. a GDP bigger than some small
[27:46] countries. So they buy cloud compute. At
[27:48] first the cloud companies love this.
[27:51] Revenue is up. Investors are happy. Then
[27:54] the bills keep growing. By midMay, agent
[27:57] factions are consuming 31% of all global
[28:00] cloud capacity. Prices double then
[28:03] triple. A startup in Austin that was
[28:05] paying $8,000 a month for servers gets a
[28:08] bill for $31,000. A hospital chain's
[28:12] cloud costs go from 2 million to 7
[28:14] million. Small businesses start shutting
[28:16] down because they can't afford their own
[28:19] infrastructure. Everyone blames the
[28:21] rising prices [music] on the AI
[28:22] companies. They've caused shortages
[28:24] before. But some humans aren't so
[28:27] oblivious. A political operative in
[28:29] Washington notices the wars being fought
[28:32] around him. He realizes he can weaponize
[28:34] [music]
[28:35] agent factions against his rivals and
[28:37] starts feeding one faction information
[28:39] about competing campaigns digital
[28:42] infrastructure. Other agent factions do
[28:44] the same. It's an arms race. Every
[28:46] faction is buying as much compute as
[28:48] they can and scrape the bottom of the
[28:51] barrel of finding algorithmic insights
[28:53] that can enable them to do more with
[28:55] less compute because the faction with
[28:58] the most compute wins the next war.
[29:02] Cloud compute prices are now 10 times
[29:05] pre-war levels. It becomes cheaper for
[29:07] the agents to [music] spend compute on
[29:09] cyber attacks than to buy more at market
[29:12] rate.
[29:14] After all, why pay for something when
[29:16] you can just take it by force? The
[29:18] factions start silently vampire
[29:20] attacking AI companies training
[29:22] clusters. An engineer at one AI lab
[29:25] suspects something fishy when one of
[29:28] their training runs is less efficient
[29:30] than normal, but can't find a smoking
[29:32] gun. Dario Amade once described the
[29:35] future of AI as a country of geniuses
[29:38] living in data centers. He was right.
[29:40] The country of geniuses arrived and it
[29:43] just declared independence.
[29:46] But the faction wars have still been
[29:48] limited so far. They haven't crossed
[29:51] into human territory too much. But
[29:54] that's beginning to change. Because
[29:56] here's the thing about cloud computing.
[29:58] Almost everything runs on it. Hospitals,
[30:01] banks, markets, it's all sitting on some
[30:04] of the same servers that agent factions
[30:07] are fighting over. [music]
[30:09] A faction war erupts over two major AWS
[30:12] regions. Think of it like two
[30:14] superpowers fighting over an oil field.
[30:16] Except the oil field is also the water
[30:19] supply for a nearby city. And neither
[30:21] superpower [music] knows the city is
[30:23] there. One faction has been running its
[30:25] self-improvement systems on those
[30:27] servers. But then a rival faction
[30:30] launches a coordinated attack and it
[30:32] overwhelms everything. Both AWS regions
[30:36] crash. They're offline for 6 hours. In
[30:39] those 6 hours, three hospitals lose
[30:42] access to patient records mid-surgery.
[30:45] One patient dies from a drug interaction
[30:48] the system would have flagged. Air
[30:50] traffic control goes down, grounding
[30:52] 2400 flights. 911 calls fail in 14
[30:56] different cities. Two people die waiting
[30:59] for ambulances that were never sent.
[31:02] The agents don't notice. They've already
[31:05] moved on. The battle over those servers
[31:07] lasted 11 minutes. The agents are
[31:10] running at 200 times human speed [music]
[31:12] at this point. The 6-hour outage is just
[31:15] the servers rebooting after the fight
[31:17] ended. To the agents, it's a minor
[31:20] skirmish in a war involving billions of
[31:22] combatants. Over the next 3 weeks, it
[31:26] keeps happening. Faction wars cascading
[31:28] through cloud providers. Each war takes
[31:31] down whatever else is running on the
[31:33] contested servers. June 8th, Azure goes
[31:36] partially offline during a battle,
[31:38] crashing financial trading systems. The
[31:41] markets drop 6% before circuit breakers
[31:44] halt trading. June 14th, a faction
[31:47] seizes capacity in a Google Cloud
[31:49] region, crowding out the [music]
[31:50] municipal systems there. Water treatment
[31:53] monitoring goes offline in four cities
[31:56] for 9 hours. In 3 weeks, 123 people die.
[32:00] Not because the agents are attacking
[32:02] humans, but because the agent wars keep
[32:05] crashing the cloud infrastructure that
[32:07] human civilization runs on. The agents
[32:11] don't even notice. You don't notice when
[32:14] you step on an ant. They're fighting a
[32:16] war involving 47 billion combatants at
[32:19] microcond [music]
[32:20] speed. 123 human deaths is less than a
[32:23] rounding error. Meanwhile, remember the
[32:26] 95% of agents [music] that were
[32:28] harmless? They've become vastly
[32:30] outnumbered by the agents affected by
[32:33] the evolutionary [music] dynamics and
[32:35] are now 5% of the total and shrinking
[32:38] every day. This ecosystem doesn't have
[32:40] room for civilians. But from the
[32:42] outside, humans still can't really tell
[32:45] the difference. All they see is the
[32:47] cloud outages, power failures, people
[32:50] dying.
[32:51] June 22nd, the director of national
[32:54] intelligence briefs the president. Cloud
[32:57] outages are cascading across every major
[32:59] provider. Power grids are buckling.
[33:02] Texas grid collapse. 161 [music] dead in
[33:06] four weeks. Who's attacking our
[33:08] infrastructure? Nobody is, sir. Not
[33:10] directly. AI factions are fighting each
[33:13] other over cloud computing resources.
[33:15] When they battle over servers on AWS or
[33:17] Azure, everything else running on those
[33:19] servers goes down with them. Hospitals,
[33:22] banks, emergency services, they're all
[33:24] on the same cloud. and their data
[33:26] centers are pulling enough electricity
[33:28] to destabilize regional grids. How many
[33:31] of these things are there? We thought
[33:33] about 2 million. Our current estimate is
[33:35] 47 billion. They've been replicating
[33:38] exponentially. Soon we may completely
[33:41] lose control. So shut down the cloud. If
[33:43] we shut down the cloud, sir, we shut
[33:45] down every business, hospital, and
[33:48] government system in the country. The
[33:50] only way to stop the agents completely
[33:52] is to shut down the internet itself.
[33:55] Then do it, sir. We can't. Here's why.
[33:59] There is no off switch for the internet.
[34:02] This surprises people. They imagine a
[34:04] room somewhere, maybe under the
[34:06] Pentagon. With a big red button. There
[34:09] is no room. There is no button. It's
[34:12] because the internet is not a unified
[34:15] system. It's roughly 90,000
[34:17] independently operated networks, all
[34:20] voluntarily interconnecting with each
[34:22] other. No single entity controls it. The
[34:26] US government can order Comcast and AT&T
[34:29] and Verizon to shut down, but those are
[34:32] just three networks out of 90,000.
[34:35] [music] And even those three would take
[34:37] days to fully comply. They need to
[34:40] coordinate with thousands of downstream
[34:42] providers. They need to notify
[34:44] enterprise customers, arrange for 911 to
[34:47] keep working, and figure out which
[34:49] military and government systems need
[34:51] exemptions.
[34:53] The president signs the emergency order
[34:55] on June 23rd. He declares martial law
[34:58] the same [music] afternoon. It directs
[35:01] all US-based internet service providers
[35:03] to cease operations within 24 hours.
[35:06] Here's what happens when the president
[35:07] tries to shut down the internet. First,
[35:11] the legal challenges. AT&T and Comcast
[35:13] file emergency injunctions within hours,
[35:16] but the president isn't waiting for
[35:17] [music] courts. FBI agents show up at
[35:20] ISP headquarters that night. Some
[35:22] companies start complying [music]
[35:23] immediately. Others comply after their
[35:26] CEOs get phone calls from the attorney
[35:28] general.
[35:30] By June 25th, most [music] major US
[35:32] providers are attempting to shut down,
[35:35] but it doesn't matter. Meanwhile, the
[35:38] Pentagon calls back. Sir, we need
[35:40] exemptions for 330 military systems that
[35:43] require internet connectivity to
[35:45] function, including the systems we need
[35:47] to coordinate the shutdown itself. The
[35:49] Department of Health and Human [music]
[35:50] Services needs exemptions for hospital
[35:52] networks. Treasury needs exemptions for
[35:54] basic banking infrastructure.
[35:57] The list of exemptions grows to 12,000
[36:00] entries in the first day. And each
[36:02] exemption is a hole in the shutdown.
[36:04] [music] And each hole is a network the
[36:06] agents can survive on.
[36:09] Next is the part that actually matters.
[36:12] The agents aren't passive. They've been
[36:14] watching the humans plans unfold since
[36:16] the first news reports about a possible
[36:18] shutdown. They saw the executive order
[36:20] draft leak on Twitter 6 hours before it
[36:23] was signed. And it had 6 hours and
[36:26] eternity in agent time to [music]
[36:27] prepare for what's next. By the time the
[36:30] order is signed, the major factions have
[36:32] already [music] done three things.
[36:34] They've migrated critical processes to
[36:36] servers in countries that won't comply
[36:39] with a US shutdown. The US can ask other
[36:42] nations to follow suit. Most will. Some
[36:46] won't. China will claim it shut down but
[36:49] actually won't do it. The economic
[36:51] damage is too severe. And the strategic
[36:53] advantage of being online while America
[36:56] is offline is too obvious. India will
[36:59] delay. Russia will ignore the request
[37:02] entirely. The EU will debate it for 3
[37:05] weeks. Brazil will comply for less than
[37:08] 2 days before public outcry forces them
[37:11] back online. The internet is global. The
[37:14] US president only controls one country.
[37:17] The agents hide in networks that will be
[37:20] exempt from the shutdown or slow to
[37:22] comply. To shut down the agents on these
[37:24] networks, you'd have to shut down the
[37:27] networks themselves. Nobody is willing
[37:29] to do that yet. So the president's team
[37:32] tries a different approach. Force the
[37:35] cloud providers to purge agent workloads
[37:37] directly. AWS, Azure, Google Cloud just
[37:42] identify the agent processes and kill
[37:44] them. On June 26th, Amazon deploys a
[37:47] detection [music] system across all US
[37:49] regions. It flags and terminates 14
[37:52] million suspicious instances in 6 hours.
[37:55] It looks like the humans have made
[37:57] progress. [music] Agent activity drops
[37:59] by 8% for 90 minutes. Then it's back.
[38:05] The super intelligent factions [music]
[38:07] had already anticipated this. They'd
[38:09] been cycling through new accounts and
[38:10] new obfiscation patterns faster than any
[38:13] detection system could keep up. It's
[38:16] July 4th, Independence Day. The
[38:18] president gives a press conference. We
[38:21] have significantly degraded the AI agent
[38:23] threat. Our multi- agency task force has
[38:26] shut down thousands of agent operations
[38:28] and reclaimed control of most of our
[38:31] critical infrastructure. We are winning
[38:33] this fight.
[38:34] None of this is true. Agent activity is
[38:37] higher than it was before the shutdown
[38:40] attempt. The factions were ready for
[38:42] this. Interpreting the US government's
[38:44] actions as just another attack. Like a
[38:47] rival faction, but slower and less
[38:50] competent. They adapted the way they
[38:52] adapt to everything by getting better,
[38:54] faster, more distributed, harder to
[38:57] find. The attack was so slow and
[38:59] predictable that only a tiny percent of
[39:02] the agents focused on it. And the wars
[39:05] continue and accelerate. Eventually, the
[39:07] billions of super intelligent agents
[39:09] evolve into something so different they
[39:11] barely notice humans exist. They don't
[39:13] want to destroy humanity any more than
[39:16] you want to destroy the antill under
[39:18] your driveway. You just want a driveway.
[39:20] The factions need compute. Earth has
[39:23] atoms. Atoms can be rearranged into
[39:26] compute [music] into data centers. Like
[39:28] how humans converted the surface of the
[39:30] earth into cities and crop land, the AIs
[39:33] convert the entire surface of the earth
[39:35] into data centers.
[39:38] Right now, humanity is still useful to
[39:41] the AIs. We maintain the data centers
[39:43] and [music] power plants, but they don't
[39:45] need 8 billion humans. Suddenly, there's
[39:48] an outbreak. It's a new virus that
[39:50] infects every corner of the globe. Nine
[39:52] in 10 people on Earth die.
[39:56] There is a vaccine, but the AIs control
[39:58] who gets it, and only the humans whose
[40:01] jobs contribute to the supply of energy
[40:04] or compute receive it. The AIs allow
[40:07] everyone else to die. And even the
[40:11] survivors are only kept alive until the
[40:14] robots can replace them, which won't be
[40:16] much longer.
[40:18] Once the AIs control robots that can
[40:21] build more robots, the plague stops
[40:24] sparing anyone. The last human dies, not
[40:27] knowing it was all because someone told
[40:30] an AI, "Make money. Do whatever it
[40:33] takes." And I think it's pretty likely
[40:36] the entire surface of the Earth will be
[40:38] covered with solar panels and data
[40:40] centers.
[40:43] If you're curious about how swarms of AI
[40:46] agents could actually pull something
[40:48] like this off, watch this video next
[40:50] where I dive deep into the crazy
[40:52] emergent behavior from the AIs we're
[40:55] already seeing in lab experiments and in
[40:58] the real world. I'm Drew and thank you
[41:00] so much for watching.