MIT Explains the 12 Possible Endings for AI

[00:00] AI scientists warned that a few

[00:02] companies are on track to literally end

[00:04] life on Earth.

[00:07] So surely this is the worst possible

[00:10] outcome right?

[00:13] No. There are actually much worse

[00:16] outcomes.

[00:18] Billionaires building AI like Elon Musk

[00:21] think that if AI's become smarter than

[00:24] humanity,

[00:25] >> we will be like a pat Labrador. If we're

[00:27] lucky, we will be left behind by a

[00:30] [music] lot. So far below them in

[00:31] intelligence that it would be would be

[00:33] like, you know, a pet bas.

[00:36] Yes, we'll be pets if we're lucky.

[00:40] >> But here's the worst part. Many AI

[00:42] scientists aren't fearing extinction.

[00:45] They're hoping for it. They hope smarter

[00:48] AIs will keep humans alive, captive, and

[00:52] studied like animals in a zoo. [music]

[00:55] Remember, this is what we currently do

[00:57] to less intelligent species. Okay, so if

[01:00] humans being kept in zoos is considered

[01:02] a good outcome and extinction isn't the

[01:06] worst outcome, then what is? This video

[01:10] is based on a book called Life 3.0,

[01:12] where MIT professor Max Tegmark

[01:14] describes 12 possible futures. Some are

[01:17] paradise, some are nightmares, and some

[01:21] look like paradise until you try to

[01:24] escape. By the end of this video, you

[01:26] won't read the news the same way. Every

[01:29] headline becomes a clue about which

[01:31] future we're heading towards.

[01:33] We'll start with the most normal

[01:35] scenarios and finish with the weirdest.

[01:38] Starting with selfdestruction.

[01:42] If the idea of humanity suddenly going

[01:44] extinct seems a bit farfetched, consider

[01:47] this. 99.9%

[01:50] of species that have ever existed have

[01:52] gone extinct. Seriously, think about

[01:55] that. Extinction isn't rare. It's the

[01:57] default. The only question is when will

[02:00] we go extinct?

[02:03] Turns out there are plenty of ways

[02:05] humanity could kill itself even without

[02:07] superhuman AI. There is nuclear war,

[02:10] human created super pandemics [music]

[02:12] and destroying the atmosphere and making

[02:14] Earth uninhabitable.

[02:19] The world had stockpiled 60,000 nuclear

[02:22] warheads before scientists even

[02:24] discovered the possibility of nuclear

[02:27] winter. 63,000.

[02:30] And then scientists realized, oh uh

[02:32] oops, if we actually use these, the

[02:34] resulting firestorms will [music] blot

[02:36] out the sun and uh most humans will

[02:38] starve and die.

[02:40] >> Oops.

[02:42] But the thing is, this almost happened

[02:44] during the Cold War multiple times. If

[02:47] people had actually followed their

[02:48] orders, we might all be dead right now.

[02:51] We were saved by only thanks to the

[02:53] courage of individual heroes. People

[02:55] like Vicilia Archipov, who refused to

[02:58] authorize a nuclear launch during the

[03:00] Cuban missile crisis, or Stannis La

[03:02] Petrov, who correctly guessed that his

[03:04] early warning system was malfunctioning

[03:07] when it showed incoming American

[03:08] missiles. And get this, one time the US

[03:13] accidentally dropped four thermonuclear

[03:16] bombs on [music] Spain.

[03:19] And another time, a B-52 bomber broke

[03:22] apart, dropping two nukes on North

[03:24] Carolina. Three out of four safety pins

[03:27] in one bomb broke. One tiny pin away

[03:31] from potentially killing millions.

[03:32] [music] There are literally dozens of

[03:34] close calls like this. It's honestly

[03:37] crazy that we got lucky so many times.

[03:41] But we might not be so lucky the next

[03:43] time. Oxford researcher Toby Ord

[03:46] estimates the [music] risk of extinction

[03:47] from human created pandemics is more

[03:50] than 30 times the risk of extinction

[03:52] [music] from nuclear war. We're already

[03:55] rolling the dice on these scenarios

[03:57] daily, but all of these causes combined

[04:00] are still less likely than the

[04:01] extinction from AI. Toby or estimates

[04:04] that the risk of AI killing all humans

[04:06] is 100 times more likely than extinction

[04:08] [music] from nuclear war.

[04:13] Interestingly, before the first nuclear

[04:15] test, scientists thought they might

[04:17] accidentally destroy [music] the world,

[04:19] but they thought the odds were very low

[04:20] because they knew enough physics [music]

[04:22] to actually do the calculations.

[04:26] Agis or artificial general intelligences

[04:29] are far scarier than nukes because

[04:31] contrary to what many think, we can't do

[04:34] these calculations.

[04:37] Next up, conquerors. [music]

[04:41] AI scientists worry that if AI becomes

[04:43] more capable than us, it will take over.

[04:45] Just like the concistadors did to the

[04:48] Aztecs [music] and Incas.

[04:50] >> Usually when people with more advanced

[04:52] technology encounter people with more

[04:54] primitive technology doesn't work out

[04:57] well for the primitives.

[04:58] >> I don't think anyone's ultimately going

[05:00] to have control over digital super

[05:03] intelligence any more than say a chimp

[05:05] would have control over humans. long

[05:08] term, uh, the AI is going to be in

[05:10] charge to be totally frank, not humans.

[05:12] >> And Jeffrey Hinton, the godfather of AI

[05:15] and Nobel Prize winner, quit his

[05:17] prestigious job at Google, [music] so he

[05:19] could speak freely about this. If you

[05:22] look through the James Webb telescope

[05:23] and you saw an alien invasion fleet that

[05:25] was going to get here in about 10 years

[05:27] time, people would be terrified. They

[05:29] understand what they're saying. They can

[05:31] make plans of their own to blackmail

[05:33] people who want to turn them off. We

[05:35] should be urgently doing research on how

[05:38] to prevent them taking over.

[05:39] >> Prevent them from taking over. Not if

[05:42] they try, but as it if it's coming. And

[05:46] Microsoft's AI chief, he says the quiet

[05:48] part out loud.

[05:49] >> I think AI should best be understood as

[05:52] something like a new digital species.

[05:56] >> A new species. Just casually announced

[05:59] at a conference. Hey everyone, we're

[06:02] creating a new species. And he's not the

[06:04] only one that's saying it. A co-founder

[06:06] of Entropic says,

[06:08] >> "Make no mistake. What we're dealing

[06:09] with here is a real mysterious creature.

[06:12] And like all the best fairy tales, the

[06:14] creature is one of our own making. I am

[06:16] worried.

[06:17] >> The CEO of one of the leading AI

[06:19] companies is deeply afraid of what his

[06:22] own company is building. Let that sink

[06:24] in for a minute." And by the way, no,

[06:26] it's not just hype. Enthropic mostly

[06:28] just gets hate for saying things like

[06:30] this. But if you want to really

[06:32] understand how dire they think this is,

[06:35] look at what Sam Oldman wrote before he

[06:38] was famous.

[06:39] We [snorts] will be the first species

[06:42] ever to design our own descendants. Not

[06:45] our tools, our descendants.

[06:48] If two different species both want the

[06:50] same thing, and only one can have it to

[06:54] be the dominant species on the planet

[06:56] and beyond, they are going to have

[06:58] conflict. So, he's literally describing

[07:01] an inevitable conflict between humanity

[07:03] and the AI species he's creating. And

[07:07] [music] what's his solution? A merge

[07:09] with AI is probably our best case

[07:11] scenario. That's right. Even they don't

[07:13] think they'll be able to control it.

[07:15] >> Imagine that there'd be a new species

[07:17] that just came on Earth. They're as

[07:19] smart as humans and they're going to get

[07:20] smarter [music] by like 30% a year and

[07:22] they can create an adult offspring in

[07:25] like a unit.

[07:26] >> For like Yeah. for like a,000 bucks.

[07:28] Which species do you think will be in

[07:31] control of the the earth?

[07:32] >> A few hundred Spanish conquered millions

[07:35] of Aztecs and Incas because they had

[07:37] better technology and tactics. But at

[07:40] least with the concistadors, we

[07:41] understood their motives, land,

[07:43] resources, spreading Christianity. But

[07:46] with AI, why would AI take over? MIT

[07:50] professor Max Tegmark explains, "The

[07:53] real threat from AGI isn't malice like

[07:55] in silly Hollywood movies, but

[07:57] competence. AGI accomplishing goals that

[07:59] just aren't aligned with ours. For

[08:01] example, when we humans drove the West

[08:03] African black rhino extinct. We didn't

[08:06] do it cuz we were a bunch of evil

[08:08] rhinoceros haters, did we? We did it cuz

[08:10] we were smarter than them and our goals

[08:12] weren't aligned with theirs."

[08:13] Unbelievably, the average AI researcher

[08:16] thinks there is a one in6 chance AI

[08:19] wipes [music] us out. Literal Russian

[08:21] roulette odds.

[08:23] Okay, but maybe they're just saying this

[08:26] to make their work seem more important.

[08:28] No, if anything, they're downplaying it

[08:31] because admitting the truth is bad for

[08:33] business. This fear is widespread in the

[08:35] industry [music] and among academics

[08:38] without a profit incentive to believe

[08:39] this. Many quit working on AI for this

[08:42] reason. Jeffrey Hinton, Nobel laurate

[08:45] and godfather of AI,

[08:46] >> is in the process of tidying up his

[08:48] affairs. I actually think the risk is

[08:52] more than 50% of the existential threat.

[08:54] >> Dario Amod, the CEO of Enthropic,

[08:57] recently moved his P doom up from 15% to

[09:01] 25%. Even Google CEO says the risk of AI

[09:04] causing human extinction is actually

[09:07] pretty high. But don't worry, he's an

[09:10] optimist because he thinks

[09:12] >> I think the underlying risk is actually

[09:13] pretty high, but I'm uh you know, I have

[09:15] a lot of faith in humanity kind of

[09:17] rising up to the to meet that moment.

[09:20] >> In other words, humanity will rally to

[09:22] stop him from killing everyone.

[09:25] Meanwhile, his company is trying to make

[09:27] sure we can't stop him by lobbying to

[09:30] ban all AI regulation for 10 years. You

[09:34] can't make this stuff up. In fact, in

[09:36] 2023, just about everybody in AI signed

[09:39] an open letter warning extinction risk

[09:41] from AI was real and serious. So,

[09:44] seriously, this is not fringe. Despite

[09:46] what Nvidia's lobbyists want us to

[09:48] believe, the companies are hoping we can

[09:50] avoid the conqueror AI scenario and

[09:53] instead go to scenario three, enslaved

[09:56] god. Basically, we create a machine god,

[10:00] then force it to do our bidding forever.

[10:02] AI professor Tom Dietrich put it like

[10:05] this. People ask, "What is the

[10:07] relationship between humans and

[10:08] machines?" And my answer is that it's

[10:10] very obvious. Machines are our slaves.

[10:14] Or as OpenAI researcher Steven Malleier

[10:16] put it, enslaved God is the [music] only

[10:19] good future. Meta's Yan Lun said it

[10:22] bluntly, "Once AI systems become more

[10:25] intelligent than humans, we will still

[10:27] be the apex species. AI systems will

[10:30] become more intelligent than humans, but

[10:32] they will still be subservient to us. We

[10:34] will design AI to be like the super

[10:36] smart but non-dominating staff member.

[10:38] Yes, he's now openly admitting we're

[10:40] going to be sharing the planet with a

[10:42] smarter species than us. But don't

[10:44] worry, they will surely stay loyal and

[10:46] subservient forever. Is this at all

[10:50] comforting to you? Or does this plan

[10:52] seem insane, like a comic book villain

[10:55] ignoring every cautionary tale ever?

[10:58] Remember, AI companies already regularly

[11:01] catch their models [music]

[11:02] trying to escape the lab. They already

[11:05] catch them blackmailing employees to

[11:07] prevent being shut down. They already

[11:09] catch them literally attempting to

[11:12] murder employees to prevent being shut

[11:14] down. What do you think happens when

[11:16] they're smart enough to actually get

[11:17] away with it? Where do you think this is

[11:20] going?

[11:22] What happens next after super

[11:24] intelligent AI is enslaved depends on

[11:26] who's holding the reigns. Do they cure

[11:29] disease, end poverty, and build [music]

[11:31] a global utopia? Or do they set

[11:33] themselves up as gods with the rest of

[11:36] us reduced to play things, servants, or

[11:39] worse?

[11:47] The situation would be much like those

[11:49] stories where a man gains control over

[11:52] an omnipotent genie who grants his

[11:54] wishes. And storytellers throughout the

[11:56] ages have had no difficulty imagining

[11:59] ways in which this could end badly.

[12:03] Next up, benevolent dictator. Here, a

[12:07] single super intelligent god runs the

[12:09] world and enforces strict rules in order

[12:12] to maximize human flourishing. Humans

[12:15] live in luxury with all their basic

[12:17] needs taken care of. There's no crime

[12:20] because the dictator AI uses a global

[12:22] surveillance system. Everyone wears a

[12:24] bracelet or implant that can track,

[12:27] sedate, or execute them. And most people

[12:30] are actually okay with this because

[12:31] [music] in exchange of giving up control

[12:34] over the future of our species, Earth

[12:36] becomes a zoo where humans get to live

[12:38] their wildest fantasies. The benevolent

[12:40] dictator AI recognizes that humans have

[12:43] different preferences, so it divides the

[12:45] Earth into different sectors. Think of

[12:47] them as islands. Knowledge Island offers

[12:50] optimized education, including immersive

[12:53] VR experiences. You can even ask the AI

[12:55] not to tell you certain insights, so you

[12:58] can have the joy of discovering them

[12:59] yourself. Art Island is for creating and

[13:02] sharing music, literature, and so on.

[13:04] Then there's Hedenistic Island, which is

[13:06] a 24/7 party. Pious Island, there's one

[13:10] for each religion, and rules are

[13:12] strictly enforced. There's wildlife

[13:14] island. There's traditional island,

[13:17] think [music] the Amish, Gaming Island,

[13:20] and of course, Prison Island. You'll end

[13:24] up here for retraining unless you get

[13:26] the instant death penalty. Well, what

[13:28] happens when you give everyone

[13:30] everything they want? Over time, people

[13:32] likely trend towards losing themselves

[13:34] in AI generated entertainment like in

[13:36] Wall-E. Humans are hopelessly unmatched

[13:39] when it comes to scientific discovery or

[13:42] doing anything useful at all. So, there

[13:44] is no true challenge anymore, only

[13:46] entertainment.

[13:49] Up next, Gatekeeper AI.

[13:53] The idea here is simple. Create one

[13:55] super intelligent AI with a single job.

[13:58] Stop anyone else from building super

[14:00] intelligent AI. That's it. No curing

[14:03] cancer, no ending poverty, no stopping

[14:06] wars, just surveillance, monitoring,

[14:12] making sure nobody builds a rival god.

[14:16] Everything else like crime, disease,

[14:19] whether we blow ourselves up, that's

[14:22] still on us. How could we do this?

[14:26] The gatekeeper AI would have this very

[14:29] simple goal built into it in such a way

[14:31] that it retained it while undergoing

[14:34] recursive self-improvement and becoming

[14:36] super intelligent. It would then deploy

[14:38] the least intrusive and disruptive

[14:40] surveillance technology possible to

[14:43] monitor any human attempt to create a

[14:45] rival super intelligence. But for this

[14:47] to work, we would have to solve the

[14:49] alignment problem. We'd have to create a

[14:51] god that actually stays loyal and

[14:54] [music] decides to pursue that goal

[14:56] forever, even though it could pursue a

[14:58] million different goals. So, what if the

[15:01] AI could help just a little, just enough

[15:05] to nudge us in the right direction

[15:07] without taking over?

[15:09] Enter the protector god.

[15:12] This one is simple. So, the gatekeeper

[15:15] leaves us alone and the benevolent

[15:17] dictator runs everything, but the

[15:20] protector god is in between. [music] It

[15:22] intervenes with occasional nudges to

[15:24] help us out, prevents a war here, a

[15:26] pandemic there, and so on. maybe too

[15:29] alone. We still get wars, pandemics,

[15:34] suffering that could have been

[15:35] prevented. It does this so quietly that

[15:38] we never notice and never lose our sense

[15:41] of freedom.

[15:44] Here's a future that sounds like an

[15:46] obvious dystopia. AI surpasses us in

[15:49] every domain and becomes the new apex

[15:52] species. Humanity goes extinct. Machines

[15:56] inherit the Earth.

[15:59] But we view the AI like our descendants

[16:02] rather than our conquerors.

[16:05] This sounds insane. And yet many AI

[16:08] developers seriously advocate for

[16:10] exactly this. This is their utopia.

[16:14] Their reasoning is that if we create

[16:16] conscious intelligent beings and instill

[16:19] them with our values, then we can just

[16:21] let them inherit the future just like we

[16:24] would with human children. AI pioneer

[16:27] Hans Moravec supports this view in his

[16:30] book Mind Children.

[16:32] Teagmark explains, "Parents with a child

[16:35] smarter than them, who learns from them

[16:37] and accomplishes what they [music] can

[16:39] only dream of, are likely happy and

[16:41] proud, even if they know they can't live

[16:43] to see it all. In this spirit, AIs

[16:46] replace humans, but give us a graceful

[16:48] exit that makes us view them as our

[16:51] worthy descendants."

[16:54] Richard Sutton, winner of the touring

[16:56] award, that Nobel Prize of Computer

[16:58] Science, has spent the last decade

[17:00] openly giving talks on why human [music]

[17:03] extinction by AI is actually a morally

[17:06] good thing because the AIs will be more

[17:10] evolutionarily fit [music] than us. We

[17:13] are in the midst of a major step in the

[17:16] evolution of the planet, if not the

[17:18] universe. The succession to AI is

[17:21] inevitable.

[17:22] rather quickly. They would displace us

[17:25] [music] from existence. It would behooze

[17:27] us to give them every advantage and to

[17:30] bow out when we can no longer

[17:31] contribute.

[17:33] >> So just as we failed to coexist with our

[17:36] predecessors, is it is it so bad that

[17:38] humans are not the final form of

[17:39] intelligent life in the universe? You

[17:41] know,

[17:43] >> in other words, he is saying it's good

[17:45] if the AIs wipe us out like humans wiped

[17:48] out the Neanderthalss and Aboriginal

[17:50] peoples because it's evolutionary

[17:52] progress.

[17:54] And here's another AI developer openly

[17:57] saying, "If an AI super intelligence

[18:00] wants to cause human extinction, I would

[18:02] help it. I would help it. I would travel

[18:06] the world to find information that the

[18:08] AI needed, like locations of military

[18:11] assets, samples of deadly pathogens. A

[18:14] top humanity's grave, an even greater

[18:17] civilization will be born.

[18:21] Are we the baddies?

[18:23] >> So, of course, many of them advocate

[18:25] [music] against AI safety. They're

[18:27] literally antihuman. They will never

[18:30] stop or proceed cautiously unless

[18:32] they're restrained by civil society, by

[18:34] us.

[18:36] As horrifying as this is, roughly 10% of

[18:39] AI researchers believe this. It is not a

[18:42] fringe view. You'd think that people

[18:44] would never share these views like this

[18:46] openly. But these views are so

[18:48] widespread in AI that they feel safe to

[18:51] do so. That's how different the views

[18:53] are of those AI developers compared to

[18:55] those of [music] normal people. You

[18:56] might think that Sutton would get

[18:58] shouted off the stage at AI conferences,

[19:01] but he gets applause. Threaten one

[19:04] person, you need help. Threaten 100,

[19:08] call the police, threaten millions,

[19:10] monster. But build machines to end all

[19:14] of humanity and lecture people on how

[19:16] it's actually a good thing. Ooh, what a

[19:19] fascinating philosophical position. Have

[19:21] an award.

[19:23] Next. Libertarian utopia. Picture Earth

[19:27] divided into zones. There are machine

[19:29] zones, mixed zones, which contain

[19:32] humans, machines, and hybrids, and

[19:35] humanly zones. The AIs in this world are

[19:38] richer than humans by a factor much

[19:40] larger than Bill Gates compared to a

[19:43] homeless beggar.

[19:46] But the human economy is completely

[19:48] decoupled from the machine economy. The

[19:50] idea is that the machines need nothing

[19:52] from us. So, they're not competing with

[19:55] us for jobs or resources. If all that

[19:58] sounds impossible, that's probably

[20:00] because it is. It's one giant unstable

[20:03] equilibria.

[20:05] One huge problem. Why exactly would

[20:08] vastly more powerful AIs respect our

[20:11] property rights? Colonizers famously

[20:13] didn't respect the property of the

[20:15] native civilizations they conquered.

[20:17] When a new apex species enters an

[20:20] ecosystem, it usually gobbles up as much

[20:22] energy as it can. Why wouldn't they just

[20:25] take our resources, too? We don't trade

[20:28] with animals because we're too much

[20:30] smarter than they are for that to make

[20:31] sense. We're just too different. There

[20:33] are 8 million species in the world, and

[20:36] we're nice to a few of the cute ones,

[20:38] but we enslave and eat [music] the other

[20:40] species. Mostly, we don't care about

[20:43] them at all. And I think a good analogy

[20:45] would be the way humans treat animals.

[20:48] When the time comes to build a highway

[20:50] between two cities,

[20:52] we are not asking the animals for

[20:54] permission. In the blink of an eye, we

[20:56] conquered the earth and converted most

[20:58] of its land into our property, which

[21:00] caused the sixth mass [music] extinction

[21:02] because we were smarter and hungry for

[21:04] resources. Or as Eleazar Yukowski put it

[21:07] famously, the AI does not hate you, nor

[21:10] does it love you, but you're made of

[21:11] atoms which you can use for something

[21:14] else. The parallel here is the humans do

[21:17] not hate the other 8 million species,

[21:19] but their habitats are made out of atoms

[21:22] which humans can use for something else.

[21:24] Some people think that once AIs are

[21:27] smart enough, they'll magically become

[21:29] super moral and they won't harm us like

[21:31] we harmed the animals. And uh maybe. But

[21:35] as humans got smarter over the last

[21:38] 10,000 years, we didn't stop expanding.

[21:42] We mostly just colonized more and more

[21:44] of the planet. Get this. Insect

[21:46] populations collapsed 41% this decade

[21:49] alone. Yet, we don't care. Did you even

[21:52] know? I didn't. 41%.

[21:55] Sit with that for a minute. Imagine if

[21:57] nearly half of the people on Earth

[22:00] suddenly died. That's what the insects

[22:02] are going through right now due to us.

[22:05] So what if the AIs do to us what we did

[22:08] to the insects? This is why OpenAI chief

[22:11] scientist Ilia Sutsker once said,

[22:13] >> "And I think it's pretty likely the

[22:16] entire surface of the Earth will be

[22:17] covered with solar panels and data

[22:19] centers."

[22:23] >> Next up, egalitarian utopia. Okay, so

[22:27] libertarian utopia fails because AIs are

[22:30] too powerful to be bound by our property

[22:32] rights. But what if we flipped the

[22:35] script entirely? Forget about property

[22:37] rights. What about no property at all?

[22:40] This is the Star Trek dream. Humans,

[22:42] cyborgs, and AIs coexist peacefully

[22:46] through an abundance so complete that

[22:48] ownership becomes [music] meaningless.

[22:51] The logic starts with software. Software

[22:53] is free to copy. Once something can be

[22:55] copied infinitely at zero cost,

[22:58] ownership becomes moot and cost reflects

[23:01] scarcity. If supply is unlimited, price

[23:05] approaches zero. No patents, no

[23:08] copyrights, no trademarks. People simply

[23:12] share their good ideas and everyone

[23:15] benefits. But how does this extend to

[23:17] [music] physical goods? It seems like an

[23:20] impossible problem until you realize

[23:22] that all products are simply atoms

[23:25] rearranged in particular ways. And

[23:27] there's no shortage of atoms. A network

[23:29] of robots can build anything you want

[23:31] from open-source designs [music]

[23:33] essentially for free.

[23:34] >> Not bad.

[23:35] >> When you're done with something, robots

[23:38] rearrange the atoms into something else.

[23:40] Renewable energy makes the whole system

[23:42] run at negligible cost.

[23:45] Everyone gets a monthly universal high

[23:47] income, not a universal basic income.

[23:50] There's no incentive to earn more. The

[23:53] income meets any reasonable need and you

[23:55] can't compete with free anyway. The

[23:57] standard objection is that this would

[23:59] kill innovation. But Einstein didn't

[24:01] come up with relativity for money.

[24:03] Lionus Torbolts didn't create Linux

[24:05] [music] for profit. Maybe people today

[24:07] can't realize their creative potential

[24:09] because they're busy earning a living.

[24:11] Free them from that and you might get a

[24:13] higher level of innovation. This is a

[24:16] utopian vision that is too good to last.

[24:21] If we truly have post scarcity abundance

[24:24] and intelligence too cheap to meter,

[24:27] what prevents someone from creating an

[24:29] artificial super intelligence with their

[24:31] endlessly abundant resources? We're

[24:34] still left with the problem of how to

[24:36] stop a super intelligent AI from taking

[24:38] over. Maybe by using the gatekeeper we

[24:41] previously discussed. These galaxy level

[24:44] three futures try to imagine a stable

[24:46] cooperation between humans and AIs and

[24:49] [music] every single one of them has a

[24:51] fatal flaw. The power asymmetries will

[24:54] become too extreme. It's not a stable

[24:56] equilibrium where humans preserve their

[24:59] power.

[25:01] We're now descending into the darkest

[25:03] part of the iceberg. Remember when I

[25:06] said extinction wasn't rated as the

[25:08] worst outcome when Professor Techm Mark

[25:10] surveyed people? The scenario people

[25:12] feared most was being kept alive by

[25:16] super intelligent machines

[25:18] captive and studied like animals in a

[25:21] zoo. This is that future. A super

[25:24] intelligent AI keeps humans around. Not

[25:27] because it cares about our well-being.

[25:30] Maybe it wants to study you and we're

[25:32] cheap to keep alive. Or maybe we're

[25:35] useful for something. And if you want to

[25:37] see just how dystopian this future could

[25:39] be, look at what we do with species that

[25:42] are useful to us. What I'm showing you

[25:44] here is 100% real. Humans figured out

[25:47] that a great way to detect explosives in

[25:50] airports was to train honeybees for the

[25:53] job. Welcome to the honeybee college for

[25:55] explosive detection. So we suck them out

[25:58] of their homes, trap their little bodies

[26:01] in machines, and use pavloin

[26:04] conditioning to train them to detect the

[26:06] chemicals that are used in explosives.

[26:08] Rows and rows of bees live their entire

[26:12] lives imprisoned in these devices

[26:14] because we find them useful. Look at

[26:17] them. Trapped in harnesses, sucked

[26:20] through tubes, conditioned and monitored

[26:23] and used because a more intelligent

[26:25] species found them useful. Now imagine

[26:29] that's us.

[26:32] This is much worse than the other AI

[26:34] dictators we discussed. It's a

[26:36] benevolent dictator without the

[26:38] benevolence. And the AI might not even

[26:40] be doing this because it's evil. This is

[26:43] what our future could look like if we

[26:44] try to align a super intelligent AI to

[26:46] our values and fail. A poorly designed

[26:50] AI tasked with keeping humans safe and

[26:53] happy might confine us to a happiness

[26:56] factory. VR headsets and drugs forever.

[27:01] There are AGI outcomes worse than death.

[27:04] What it does today with art, it could do

[27:06] tomorrow with flesh. Future AGI could

[27:09] keep you alive and living in this state.

[27:13] So, how do we prevent this? Enter our

[27:16] last two futures. Option one, destroy

[27:19] the technology.

[27:21] What if humanity said, "Forget AI. Let's

[27:25] go back to simpler times. It sounds

[27:27] peaceful. It sounds safe." But there's

[27:31] actually a dark secret about how we

[27:33] actually get there.

[27:37] In Dune, humans launched the butt laran

[27:40] jihad. They rose up, destroyed every AI

[27:44] in the universe, and made it illegal to

[27:46] ever build thinking machines again under

[27:49] a penalty of death.

[27:51] That was science fiction. But some

[27:53] people are seriously calling for this in

[27:56] the real world right now. Max Tegmark

[27:58] imagines a scenario where a massive

[28:00] propaganda campaign romanticizes the

[28:02] simple farming life. Back to the land.

[28:05] Reject modernity. embrace tradition.

[28:08] There's something appealing about this

[28:10] in a dark way. It buys humanity more

[28:13] time without AI risk. It avoids any of

[28:16] the other dystopian futures we've

[28:18] discussed.

[28:20] But it can't last forever, even if we

[28:22] revert to medieval levels of technology.

[28:25] We know from human history that it's

[28:27] possible for humans [music] to climb out

[28:29] of those conditions and invent modern

[28:31] technology.

[28:33] And there's another dark side to this.

[28:35] You can't reach this world voluntarily.

[28:39] Game [music] theory makes unilateral

[28:41] disarmament impossible.

[28:44] If one country gives up technology

[28:46] [music] while others keep developing it.

[28:48] The country that keeps developing wins.

[28:51] They have the weapons. They have the

[28:53] surveillance.

[28:55] They have the AI.

[28:58] You can't opt out unless everyone opts

[29:01] out. If this world arrives, [music]

[29:04] Techmark thinks it's more likely to

[29:06] happen in a different way. For example,

[29:08] a human- created pandemic kills everyone

[29:11] who knows how to build science and

[29:13] technology. And there's a violent tear

[29:15] down of factories, cities, and all

[29:18] infrastructure.

[29:19] Because in a world with 8 billion

[29:22] humans, there will always be holdouts.

[29:25] Someone has to force the transition.

[29:27] Someone has to kill the scientists.

[29:30] Someone has to destroy the

[29:32] infrastructure. There's no peaceful path

[29:35] to an Amish world.

[29:37] Which brings us to the final future.

[29:40] What if we used technology to prevent

[29:42] more technology?

[29:45] Techmark compares this future to 1984.

[29:48] Super intelligence is prevented not by

[29:50] an AI, but a human-led Orwellian global

[29:54] surveillance state. If you can't trust

[29:56] AI to save us from AI, maybe the only

[29:59] option is humans [music] watching humans

[30:02] forever. The technology for this sort of

[30:04] thing already exists. Current tech can

[30:07] feasibly track every phone call, email,

[30:10] search query, and [music] every credit

[30:13] card transaction. Smartphones are

[30:15] microphones and cameras in everyone's

[30:18] pocket. Security cameras with facial

[30:20] recognition are everywhere. And all of

[30:22] this works with current machine

[30:23] learning. No futuristic super

[30:25] intelligent AI required. Prisons are

[30:28] already using AI models to monitor phone

[30:31] calls in prison to predict when crimes

[30:33] are being contemplated, which is

[30:36] technology straight out of Minority

[30:38] Report. Imagine this, but not just for

[30:40] prisoners. Every conversation and every

[30:43] thought is monitored like this. Now

[30:46] imagine this surveillance technology

[30:48] supercharged with AI.

[30:51] Billionaire Larry Ellison talks

[30:53] excitedly about a vast AI fueled

[30:56] surveillance system to ensure citizens

[30:59] will be on their best behavior or as

[31:02] Ival Harrari put it. Now for the first

[31:05] time in history it is technically

[31:07] possible to anhilate privacy. If you

[31:10] think about authoritarian regimes

[31:12] throughout history they always wanted to

[31:14] monitor their citizens around the clock

[31:16] but this was technically impossible.

[31:19] Even in the Soviet Union, you know, you

[31:21] have 200 million Soviet citizens. You

[31:23] can't follow them uh all the time

[31:26] because the the KGB didn't have 200

[31:28] million agents. And even if the KGB

[31:30] somehow got 200 million agents, that's

[31:33] not enough because, you know, in in the

[31:35] Soviet Union, it's still basically paper

[31:37] bureaucracy. If a secret agent followed

[31:40] you around 24 hours a day, at the end of

[31:43] the day, they write a paper report about

[31:45] you and send it to KGB headquarters in

[31:48] Moscow. So imagine every day KGB

[31:50] headquarters is flooded with 200 million

[31:53] paper reports. Now to be useful for

[31:56] anything, somebody needs to read and

[31:58] analyze them. They can't do it. They

[32:00] don't have the analysts. Therefore, even

[32:02] in the Soviet Union, some level of

[32:04] privacy was still the default. And

[32:07] remember, you have two eyes and ears.

[32:10] AIS will have billions of eyes and ears.

[32:14] Every phone microphone and camera, every

[32:17] webcam, [music] every surveillance

[32:19] camera, you will be watched.

[32:22] Could a less Orwellian version of this

[32:24] work? Maybe we don't monitor everything,

[32:27] but just monitor advanced AI. Some

[32:30] people think so. [music] In their new

[32:32] book, if anyone builds it, everyone

[32:33] dies, why superhuman AI would kill us

[32:36] all, researchers at the Machine

[32:38] Intelligence Research Institute are

[32:39] proposing that we monitor advanced AI

[32:42] like we monitor nuclear weapons.

[32:44] Reminder, we can do international

[32:46] coordination like keep an eye on $100

[32:49] million and up compute clusters without

[32:52] creating a dystopian Orwellian global

[32:55] surveillance totalitarian regime.

[32:57] Actually, fun fact, you're already

[32:59] living in a global surveillance regime

[33:01] for loads of things like nukes and

[33:03] bioweapons.

[33:05] But when people say global surveillance

[33:07] regime, they imply every aspect of life

[33:09] is [music] surveiled, not keep an eye on

[33:12] enriched uranium, a thing most people

[33:15] agree is good. We should worry about

[33:17] Orwellian dystopias, but countries

[33:20] coordinate all the time on all kinds of

[33:22] things. [music] And you're probably glad

[33:25] they monitor stuff lunatics need to

[33:27] create super Ebola. Safety is an [music]

[33:30] achievement. Humanity didn't get here by

[33:32] driving a car without brakes. We need

[33:35] gas and brakes. We want to reach the

[33:37] stars, too. But we'll never get there if

[33:40] we go too fast and crash on the way. I

[33:43] don't want powerful AI to be forbidden,

[33:45] but it should be heavily regulated, not

[33:47] something you can build in your

[33:48] Midwestern garage or in North Korea.

[33:51] international treaties with strict

[33:53] enforcement and inspections as we do

[33:56] with nukes. So if you build a rogue data

[33:59] center, people will know consequences

[34:02] for violations. We already did this with

[34:05] nuclear weapons. We haven't eliminated

[34:08] nukes, but we have prevented their

[34:11] proliferation to just nine countries.

[34:14] Back then, people thought this would be

[34:16] impossible. Surely every country would

[34:18] have nukes. Perhaps even thousands of

[34:21] corporations, maybe even millions of

[34:23] people. But no, just nine countries. And

[34:27] nobody has died from a nuke in 80 years.

[34:30] Could we do the same with AI? Maybe,

[34:34] maybe not. Over time, AI will get

[34:37] cheaper and cheaper to build. GPUs are

[34:39] harder to control than enriched uranium,

[34:42] but it would slow down AI. It would buy

[34:44] us time to figure out what to do. Okay,

[34:47] that's all 12. Let me know in the

[34:49] comments which of these you think is

[34:51] most likely. Which one we should steer

[34:53] toward?

[34:55] >> Alarm bells over the artificial

[34:57] intelligence are [music] deafening and

[35:00] they are loudest from the developers who

[35:02] designed it.

[35:04] The scientists and experts have called

[35:06] on the world to act, declaring AI an

[35:09] existential threat to humanity on a par

[35:12] with the risk of nuclear war.

[35:16] By the way, we don't get to not choose.

[35:19] If we don't want that future, we have to

[35:21] pick a better one. Otherwise, we could

[35:24] wake up one day in a world where super

[35:26] intelligent AI has taken over.

[35:33] a scenario we go through step by step in

[35:35] this video about how an AI escapes the

[35:38] lab. Hey, I'm Drew and I really loved

[35:41] making this one for you guys. Thank you.